Inspired by the awesome Lenny, I adapted the idea of a gift guide for Core Matter. And it turns out my past life as an edtech founder was helpful as well. So many of the gifts I came across for toddlers and kids are from my firsthand experience teaching kids to code many years ago.

And thanks to the community on X, who contributed so many awesome products for the bench builders category. I’ve kept these free of affiliate links.

Thanks to everyone who replied to the polls, especially @alexembodied, @Wbackdown, @benbeerengineer, @dcsan, @liminalsunset_, @TouficJrab, @boxcardavid, @_baldtires, @MikeAHFerguson, @Joyingeth, @RohitSatyaV, @reticente07, @LiamTal and @lasermatts.

In this gift guide:

1. For friends with kids

2. For the software engineer building their first robot

3. For the robot founder deploying robots, or visiting Shenzhen for the 1st or 10th time

4. For the bench builder in the garage

5. For the Core Matter reader catching up over the holidays

For friends with kids

Bee-Bot · $99.75 · ages 3 to 8
This trusty little Bee-Bot was the first robot I used in the age-3 curriculum teaching coding, back in the day. I’d still get it for my toddler, now that he’s turning 3 soon. Plug and play, no screens, and space friendly.

Cubetto Playset · about $225 · ages 3 to 6
A pricier choice than Bee-Bot that lasts a little longer, with more functionality. Aesthetically, it fits well into a Montessori play area and a midcentury San Francisco home.

Button Box · DIY and open source, kits in beta
I’ve been lurking on their WhatsApp channel for a few months now. This is a DIY cardbox box with a button that you program to allow toddlers to leave voice messages for loved ones on WhatsApp. My toddler loves leaving voice notes to his grandparents via Whatsapp, so it’s definitely on my holiday project list.

Nex Playground · $299 · ages 5+ (3 to 4 with a grown-up)
Turn your living room into an arcade, dojo and more. So many friends with kids have been raving about it. It helps burning off some excess energy especially during cold winter days. I am getting it for my niece and nephew (6 and 9 years old).

BBC micro:bit V2 Go kit · $19.95 · ages 8+
This is for older kids, but I think it’s the best way to get started building something, even for adults.

For the software engineer building their first robot

MicroDuck · $399 pre-order | Reachy Mini · $399 Lite, $499 Wireless
Both MicroDuck and Reachy Mini have a lead time of a few weeks so they may not arrive before the holidays. But who can resist that cuteness? Reachy Mini also recently got a hoverboard upgrade (€135), so it can run around with its younger sibling soon. When asked which one to get, a friend of mine used to say, “always get both.”

SO-ARM101 · $199 to $299
This is the arm that went viral when @cdngdev used it to have GPT-6 Astra paint the Golden Gate Bridge. If you’re looking to try Astra on robots, this is a good starting point. I have a long list of projects I want to do with these arms.

“A good robot arm kit under a thousand would be handy” (@sheikhakash69)

Bambu Lab A1 mini · $199
A must-have for anyone getting into hardware/robots, and a great way to learn how to make 3D drawings.

3Dconnexion SpaceMouse Compact · $179
This is suggested by quite a few folks on X. I’m so curious how it works, though I’m still quite attached to my Mac mouse.

Suggested by @alexembodied

Le Creuset 14-oz mugs, Oyster, set of 4 · $79.95
I’ve been in the market for nicer coffee mugs and can’t take my eyes off these. They come in so many colors. Definitely needed for all those nights wrangling robot arms and code!

Asimov 1 DIY kit · $20,000
This is for the friend who has everything. Menlo Research, maker of Asimov, ships these in monthly batches (already shipping, and from the US). It has a big community on Discord, in case you have questions. Menlo estimates about 20 hours to build it. Might be a fun Christmas holiday project at home?

For the robot founder deploying robots, or visiting Shenzhen for the 1st or 10th time

These could also be good gifts for your supplier in Shenzhen.

Anker Laptop Power Bank, 25K, 165W · $119.99

Suggested by @boxcardavid

Sipeed NanoKVM-Go · $79 pre-order

“plugs into tailscale and any computer to allow you to access it remotely with low latency. perfect for lab use” (@liminalsunset_, on the NanoKVM Pro)

O’Keeffe’s Working Hands hand cream · about $8
I don’t work in a factory (yet), but my hands are always dry. This is the ultimate workhorse.

JISULIFE portable neck fan · $29.99
Anyone who’s been on a factory floor knows how hot it gets. Keep cool with these neck fans.

Zojirushi stainless steel travel mug · $28.96
Like most Asians, I have so many Zojirushi products growing up and still do at home now, from rice cooker to kettle to thermos. Their products are top notch and last so long. This smoky blue color is so beautiful. Stay hydrated on the floor (pair with the neck fan).

Refurbished Lenovo ThinkCentre Tiny mini PC · $199.99

“Quieter and sturdier than a Pi once the workload gets real.” (@TouficJrab)

Starlink Mini · from $149, plus a plan

“Starlink, absolute necessity for our deployments” (@RohitSatyaV)

Also loved: Toyo toolbox, fast SSD, Subtle earbuds, Field Notes

For the bench builder in the garage

This was the most popular section in my X polls. Lots of pros on my feed.

Mitutoyo 500-196-30 caliper · about $175 to $185

“set of quality calipers from Mitutoyo” (@alexembodied). Also @reticente07 and @LiamTal.

Hakko FX-888DX soldering station · $121.47

“Good quality soldering iron. The cheap ones are garbage and you’ll hate yourself for saving money.” (@Wbackdown)

Fluke 117 multimeter · about $245 to $305

“Fluke multimeter + probes kit. Having certainty in your measurements is the best debugging tool.” (@benbeerengineer)

Wera ball-end hex key set · $43 to $80

“a good set of hex keys is under $250 and saves so much frustration” (@Joyingeth)

Garrett Wade headband magnifier · $49.95

Suggested by @barriosdrew

FNIRSI DSO-152 pocket oscilloscope · $39.30

Suggested by @alexembodied

Leg massager and foam roller set · $4.88
A home spa after a long day at the bench. It does magic and saves you a trip to the spa. I got mine from Taobao and is probably one of the most used products at home.

Also loved: 12V bench power supply (@dcsan), thermal camera, heat-set inserts, Klein digital angle finder ($64.98, @_baldtires), electric screwdriver with hex bits (@MikeAHFerguson)

For the Core Matter reader catching up over the holidays

(Hope not! Enjoy the festivities.)

Keyboard fidget toy, 5-pack · $15.89

Mini singing bowl · $13.99
I found both on Taobao (the stress reducer was Rmb 7, about $1, and the gong was Rmb 3, about $0.40). There’s an Amazon premium here, but I quite love both.

If you try any of these, let me know if you love them. Or share your favorite finds this year with me :)

Read more

China races for the brain, while the US fixes the unit economics. Unitree, like other Chinese makers, is racing to catch up on the brain while fending off dozens (if not hundreds) of competitors at home. Agility, like other US makers, has to make the unit economics work and build a humanoid that sells above cost. It spends about 4.5x what Unitree spends on R&D, on about 1/134th of the revenue.

Both now have to grow into their valuations. For Agility, that means delivering on its orders and proving its robots work reliably in customer deployments. For Unitree, it means holding gross margin and growing beyond academic buyers, which also depends on better AI.

These are financial stories as much as operating ones. Agility is billed as the first US-listed pure-play humanoid company, and Unitree is the most valuable listed humanoid maker in our comparison. Unitree’s debut made China’s regulators stricter, and they now look at whether revenue comes from independent customers.³³ Agility’s trading will be the first US test.

Disclaimer: this is not financial or investment advice.

In this piece

1. Humanoid Robot Financials Compared: Agility, Unitree, UBTech, AgiBot and Peers

2. Agility Robotics’ S-4: $1.8mn of Revenue, a $300m Order and a Policy Tailwind

3. Unitree Q3 2026: Revenue Up 70%, Margins Down and a Race to Catch Up on AI

4. China Races for the Brain, the US for Unit Economics

5. “If you don’t surge, how will the other robot makers go public?”

6. Footnotes

1. Humanoid Robot Financials Compared: Agility, Unitree, UBTech, AgiBot and Peers

In 2025, Agility spent more on R&D ($91.6m) than Unitree, Leju and Youibot combined (about $41m).

2. Agility Robotics’ S-4: $1.8mn of Revenue, a $300m Order and a Policy Tailwind

Agility (AGLT) has 2 things going for it: reshoring and AI.

Reshoring. Agility is betting on reshoring and reindustrialization, and that there will be a handful of domestic winners that can do heavy-duty industrial work.

Washington has moved its way this year. In July, the FCC added foreign-produced advanced robotic devices to its Covered List, which bars new models from July 28, 2026 onward from receiving the authorization they need to be imported or sold in the US.⁷ Unitree’s 6 existing models keep their authorization.⁸ On Oct 29, the White House issued a Section 232 proclamation on imported robotics and industrial machinery threatening national security. It did not talk about tariffs yet, but opened 180 days of negotiations with trading partners, till April 26, 2027.¹⁷

Agility has been part of the push. Peggy Johnson, its CEO, endorsed the GUARD Act at its introduction in June and made the case for it in a July op-ed as a House subcommittee took it up. The bill would put adversary-made humanoids and quadrupeds through a national security review and onto the FCC’s Covered List.¹⁸¹⁹ About 75% of Digit’s parts are sourced in the US.² The tailwind has a catch. Section 232 also covers parts and components. Johnson has said US developers still need to source some components internationally.²⁰

AI. At IROS in late September, Digit ran an end-to-end, whole-body mobile manipulation demo for about 12 hours over 3 days, with several stretches running for hours without a reset.²¹ Chris Paxton, an engineer who works on AI at Agility, said the demo ran at “1x speed, autonomous, in a previously-unseen location (not in training data).”²² An attendee who spoke with the team reported a success rate of roughly 80% to 90%, from a policy trained on about 500 teleoperated demonstrations. Paxton described the work as research that is not yet running in production.²³

The weaknesses show up in the numbers.

Revenue. The revenue base is thin. The 5 robots sold to one investor brought in about $1.05mn of 2025’s $1.8mn.² First-half 2026 sales were $174k, less than the price of one of those robots.

Cost. Agility is the only publicly listed humanoid company in our comparison selling below cost, at low volume. Its bill of materials per humanoid has actually gone up: about $150,000 for Digit 5 at launch, against about $125,000 for Digit 4 today.²

Timing. Digit 5 has slipped. The first S-4 said late 2026, and the amendment says early access in the first half of 2027.² Amazon, an investor that ran several Digit pilots, was dropped from the customer list in the amendment without explanation.²⁴

Backlog. Agility has $300mn of multi-year orders for Digit 5. I did not find a total order backlog in the filings of the Chinese companies in our comparison. The orders come from a single contract: 1,000 Digit 5 robots on a 3-year robots-as-a-service term, tied to milestones the company has to meet. The filing says the figure is not a measure of current period revenue.² The customer is a related party, as reported from the amended filing, and receives warrants that vest as robots are deployed, about 453 per robot.¹⁰

3. Unitree Q3 2026: Revenue Up 70%, Margins Down and a Race to Catch Up on AI

Unitree’s strengths are scale and profit. Revenue grew 70% in Q3 to Rmb665m ($94mn), and 56% in the first 9 months to Rmb1.82bn ($256mn).³ It sold more than 5,500 humanoids in 2025.⁴ It is also the only profitable company among those in our comparison that disclose profit, with $49mn of net income excluding one-off items in the first 9 months.³

The weaknesses are margin and AI.

Gross margin is trending down. For Chinese makers, including Unitree, domestic competition is their main worry. AgiBot surpassed Unitree’s shipments in the first half,²⁵ with its company-reported first-half revenue was Rmb2bn ($300mn) , against Rmb1.15bn ($172mn) for Unitree.⁶²⁶

Unitree cut humanoid prices deliberately in 2025 “to consolidate industry position,” and its cheaper G1 model became the volume seller.⁴ Humanoids are still its best-margin product, which is why selling more of them raised margin in 2025. That premium is shrinking. Humanoid gross margin led the rest of Unitree’s business by about 43 points in 2023 (on 5 units), 17 points in 2024 and 6 points in 2025.⁴

The 2026 drop at Unitree, from 59.8% in the first 9 months of 2025 to 54.0% in the first 9 months of 2026, isn’t explained in the Q3 report. Quarterly reports in China give no breakdown by product, so the cause is unconfirmed until the 2026 annual report. By product, Unitree’s humanoid margin is trending down while its share of revenue goes up. My guess is this trend will continue. Unitree’s 54% gross margin as of 9M 2026 is still much higher than its Chinese peers; UBTech’s 44.7% and LeJu’s 40.8%.

Catching up on AI is driving up R&D. Unitree founder Wang Xingxing has repeatedly said he underestimated AI: “The biggest mistake I’ve made over the past decade-plus is not believing in AI enough.”²⁷ He titled his October keynote “Believe, believe in AI as much as possible.”²⁸ He defines the “ChatGPT moment for embodied intelligence” as: “In 80% of unfamiliar scenarios, a robot can complete about 80% of tasks via voice or text commands.”²⁹ He believes this could happen as early as next year, shortened from the 2 to 3 years he gave in August.²⁷²⁹ At Unitree’s partner ecosystem conference in October, he said: “Physical AI has already shown signs of being broken through. At the fastest, 1 to 2 years. At the slowest, 2 to 3 years.”³¹

The R&D spending has grown significantly, and we believe will likely continue to climb. R&D rose from 8.5% of revenue in 2025 to 19% in Q3, and Q3 net profit fell 17% as R&D jumped 250%.³ Even so, Unitree’s R&D at about 15% of revenue for the first 9 months is still lower than UBTech, Youibot and Leju (24-25% of revenue), let alone Agility (5000%) (refer to Exhibit 1).

4. China Races for the Brain, the US for Unit Economics

The challenges facing Unitree and Agility reflect wider shifts in physical AI in China and the US.

China: the race moves to the brain. Reuters reports that Chinese robot makers have mastered the moves and are now putting their money into the AI that lets robots handle many kinds of work, as hardware prices fall.¹¹ Unitree has company:

• Spirit AI raised $280m in February to scale its general-purpose embodied models.¹²

• X Square has published research on robots learning a handling skill from a single human video, and its Series C valued it above Rmb20bn ($2.9bn).¹¹¹³

• Robbyant, Ant Group’s robotics unit, open-sourced its LingBot-VLA model in January and released 6 more models in July.¹⁴

• UBTech offered a chief scientist up to Rmb124m ($18m) a year to lead its AI model research back in April 2026.¹⁵

Embodied AI companies in China raised Rmb93.5bn ($13bn) across 322 deals in the first half of 2026.¹¹

US: the race is to make the unit economics work. Agility’s gross margin was −151% in 2025. Each robot has to cost less to build than it brings in. Agility has to find customers who will pay a price that leaves a positive gross margin. Its own target is a bill of materials of about $75,000 at 1,000 robots a year.²

Serve Robotics, the Nasdaq-listed maker of sidewalk delivery robots, shows the same problem outside humanoids. It booked a gross loss of $8.8m on about $3.2m of revenue in Q2 2026 as it grew its robot fleet. It cut its 2026 revenue guidance from about $26m to $9m to $10m.¹⁶

5. “If you don’t surge, how will the other robot makers go public?”

The stakes reach beyond both companies. In China, regulators have already hit the brakes. In September, Reuters reported that they were slowing a rush of humanoid listings, with one person saying humanoid IPOs had effectively been frozen for now. The main trigger was Unitree’s volatile start: its shares rose more than fivefold on debut, then fell 55% from the peak. Regulators are focused on whether revenue from local-government-backed projects, such as robot data-collection centers, reflects demand from independent customers.³³

At least two dozen humanoid-related companies have filed to list in Hong Kong alone.³⁴

With JPMorgan’s two-notch upgrade to Overweight on October 25,³⁰ Unitree’s stock has rebounded more than 70% from its October 15 low.³² After the rebound, a highly upvoted comment on a Chinese news app put it bluntly: “If you don’t surge, how will the other robot makers go public?” (你不大涨，其它机器人怎么IPO。)³⁵

My view is that both are in for a tumultuous few years. As first movers in their markets, they have to break through the industrial adoption cycle under public scrutiny, and there will be a lot of turbulence before the dust settles.

Tesla’s Optimus ramp over the next year will show what volume does to the cost of a US-built humanoid, the question Agility’s margin depends on.⁹

It was the best of multiples, it was the worst of margins. For now, investors are pricing the multiples. The next 12 months of filings will show whether the margins catch up.

Footnotes

At a precision gear factory in Songgang, Shenzhen, the owner told me about a customer’s design for an intricate part. It couldn’t be made as drawn. The factory caught the problem early and helped the customer change the design.

On the same visit, I saw experienced workers inspecting parts and checking tolerances by hand. They had spent years learning this work. Their judgment helped determine whether a part was ready to ship.

A US actuator startup founder told me about the challenge from the other side. As he scaled production, the challenge was not in sourcing machines, but in finding experienced people to run them.

I see an opportunity in helping manufacturers shorten this learning curve. How might AI capture what experienced people know, help others use it and test whether it works? Making that knowledge easier to pass on could help new factories get into production sooner.

I’ll explore this over the next few months as a series. Today, we’ll start with parts manufacturing because the products we want to build here depend on a supply of reliable parts. Following a custom-machined part from drawing to delivery gives us a concrete way to see where manufacturing gets st. By the end, you’ll have a framework for understanding where AI could help and how to judge whether it works. Components, subassemblies and finished products will follow in later pieces.

Thank you to Yeok (@yeokch), Arthur (@darthur) and Adi (@adiprasad27) for sharing their experience and feedback, and to Ted Xiao (@xiao_ted) for the presentation at the Humanoid Hub Conference that helped inform this piece.

In this piece

1. The manufacturing stack: where should the US start?

2. Where AI can improve manufacturing: from drawing to delivery

3. Shenzhen’s manufacturing ecosystem: could the US build its own?

4. Verification: what counts as proof in manufacturing AI?

5. 3 ways AI could close the US manufacturing knowledge gap

6. Footnotes

1. The manufacturing stack: where should the US start?

Building domestic manufacturing capability starts with deciding what we want to make. In Apple in China, one of my favorite books I read this year, Patrick McGee describes how Apple helped build manufacturing expertise in China through engineers, equipment and work on supplier factory floors.1 That capability developed over decades of making products and iterating the processes.

I believe materials and parts are foundational to rebuilding US manufacturing because they feed the layers above. Read the stack below from the bottom up, starting with materials and standard items, then parts, components, subassemblies and finished products. A factory assembling a robot relies on suppliers for housings, brackets and other parts. More domestic suppliers could give US teams more options and shorter turnaround times when iterating on designs.

Which product falls in which layer depends on who the buyer is. A gearbox can be a finished product for a gear maker and a component for a robot manufacturer.

This piece focuses on the parts layer. Making subassemblies and finished products brings different questions about how components fit together, how to automate assembly and how to test the finished system.

2. Where AI can improve manufacturing: from drawing to delivery

Imagine a regular day in a factory: A customer sends in their drawing and asks for a quote. Before the factory can commit to a price and lead time, someone has to work out how to make the part with the equipment, people and time available.

Process planning is a classic optimization problem. The same part can be made with different tools, orders of operations, people on the team, and different combinations of those, depending on how long it takes and which machine or human is available. What separates a great factory from a good one is how optimized and flexibly adaptable the shop’s equipment, the machinists’ experience with the material and tolerances, and the customer’s deadline are.

The 5 kinds of work below follow a custom part from a request for a quote to delivery. They overlap and repeat.

These tasks are interconnected. A different material can change the machining plan. A late delivery can change which route is practical. A quality problem can send a job back to setup or design.

3. Shenzhen’s manufacturing ecosystem: could the US build its own?

When I talk to people who have built hardware in Shenzhen, they are skeptical of the ambition of the US “replicating Shenzhen.” I can see why.

The magic of Shenzhen extends beyond any individual factory. It comes from the density, interconnected tissue and speed at which everything operates. Nearby specialists who have worked together over many jobs. The trust built between suppliers and customers. And because many factory employees move from their hometown to work close to the factory, work is a big piece of their life and they often work across the weekend.

In my interview with Eric Pan of Seeed Studio, he described working alongside Pollen Robotics to adapt Reachy Mini for production. His account points to 3 parts of that ecosystem.

Density gives factories specialists within reach. Tooling, machining, testing and repair are easier to coordinate when the relevant people are nearby. For example, a shop quoting a part that needs heat treatment could work with a nearby specialist to price that step.

Relationships save time. Trust and repeat work help suppliers understand requirements, handle changes and solve problems together.

Demand sustains the network. Orders from multiple customers help specialists keep equipment running and skilled people employed.

On top of that, Shenzhen truly never sleeps. In a few factory visits, I learned that employees live next to the factory in dormitories provided by the factories. They work 10-12 hour shifts to bank overtime pay. They choose to work 6-day weeks, or 26-day months, to bank holidays to go back to their hometown at the end of each working month.

How might the US build that capability, density and speed across more geographically dispersed hubs? Could live quoting and resource sharing help factories coordinate across distances? What would still need to happen close together?

4. Verification: what counts as proof in manufacturing AI?

“Progress follows verification.” Ted Xiao, founding engineer at Prometheus, shared in his talk at the Humanoid Hub Conference. His argument was that even after AGI, verified knowledge of the physical world would remain scarce.2

When thinking about physical AI such as manufacturing AI, verification is the measuring stick to evaluate which problem is ready to be solved and where the bottlenecks remain. An easy way to visualize what this means is by seeing the contrast between robotics data and physical data. In robotics, teams have had to collect demonstrations and experience from robot runs. Across manufacturing, decades of designs, test logs and production experience already exist. But much of that knowledge is scattered across companies, held in people’s heads or guarded as a trade secret.

When I interview machinists and engineers who work on the factory floor, the word “muscle memory” comes up a lot. It might be the angle used to hold a tool, the timing of an operation or an adjustment for a particular machine. This is the tacit knowledge that physical AI needs to collect. And it won’t be collected through data farms like robotics data.

Some software tasks have quick checks, such as running a test to see whether code produces the expected output. In manufacturing, running a check involves material, machine time and a production run. As tools wear and materials change, a trial run that’s tested to be successful can degrade over time.

For a company building manufacturing AI, think from the vantage point of verification and access to data:

1. How will it access the know-how and records inside factories?

2. How will it generate feedback it can trust?

3. How quickly and cheaply can it find out that a proposed method is wrong?

Verification and access to tacit and hidden data lead to progress. That gives us a way to screen opportunities: where can we get the data, measure success, afford the experiments and reuse what we learn?

5. 3 ways AI could close the US manufacturing knowledge gap

Below we explore 3 approaches to close the manufacturing knowledge gap in the US.

Read more

Each robot had a human operator controlling it remotely from the company’s office in the Mission. A staff member dropped them off and picked them up. For most of the hour, there was no company staff member in the house.

This experience reinforced my belief that home robots should be available to rent by the task or hour. For the user, paying $60 to try a cleaning service is a much smaller commitment than buying a robot or taking on a monthly subscription (Weave lists Isaac 1 at $449 per month or $7,999 up front [1], and 1X lists NEO at $499 per month or $20,000 to own [2]). For the robotics company, sending the same robots into different homes could increase utilization and give it more opportunities to collect data and improve both hardware and software before the next deployment.

From observing this visit, I also saw how much work surrounds that hour of cleaning. Below I share a) my experience as a first time user of a humanoid cleaning service, b) the economics for the user and for the company, c) what a future with humanoids and advanced robots doing productive work around us would look like.

If you’ve hired a humanoid to clean your home, be it in SF, or Shenzhen, or elsewhere, I’d love to hear about your experience.

1. The task is the product

I paid $30 per robot-hour, but the booking ended with work unfinished. Do I care whether it’s teleoperated or fully autonomous? No. What matters most is to have the whole home cleaned.

Around my neighborhood in San Francisco, a full-home clean costs roughly $200–220 for 2 human cleaners working about 2.5 hours. That comes to $40-44 per worker-hour. The robot service has a lower hourly price, but my $60 booking left the job unfinished. In later correspondence, one of the founders, Cornelia Weinzierl, shared that they could potentially do a longer visit by swapping batteries after an hour.

What can the cleaning robots do?

Vacuuming, mopping and wiping surfaces seem easy for them. They were doing it almost as well as a human.

The robots seemed slower on delicate tasks, such as moving objects to clean a surface and picking up small or soft objects like toys and socks. Chelsea dropped a toy, struggled with a pillow and had trouble balancing on the playmat. For the tasks they completed, grippers were enough.

They didn’t attempt unloading my dishwasher, as it was running at the time. After the service, I learned from another user that it was described as an “advanced task” by the company. The robots also didn’t fold our laundry. Weinzierl later told me that loading and unloading the dishwasher are possible but slow, and that some dishes may be placed at unusual angles. She also said the robots can “fold T-shirts, pants, shorts, skirts, diverse types of towels, roll up socks. Hang shirts, blouses, dresses, light jackets and similar. They struggle with folding bedsheets or tight sports long sleeve shirts if they are inside out.” Frankly, I find that pretty impressive with their grippers. That suggests five fingers may not be necessary for these household tasks.

I was told the batteries needed swapping after about an hour. Extending my booking would have required swapping the battery.

2. The economics extend beyond the house

Tau Robotics charged $30 per billed robot-hour. Beyond Sergey and Chelsea’s labor, the visit required several kinds of human support:

1. A staff member handled drop-off and pickup. The robots also needed help taking their shoes off and putting them back on, which the staff member handled at drop-off and pickup.

2. One operator was dedicated to teleoperating each robot. They were Tau staff working from the Mission, according to the person who dropped off the robots.

3. A driver transported them in a van.

4. Someone responded to my text request.

5. Between visits, batteries need swapping and equipment, clothes, socks and grippers need cleaning.

What would improve a robot operator’s economics?

The case for optimizing utilization

At this stage of autonomy, I believe a shared fleet is a promising way to collect data and improve the service. Sending robots into more homes could expose them to different objects, layouts and failures. Collectively, this can bring a better experience for the user from each visit.

The value of collecting data in customers’ homes depends on how quickly it leads to better performance. For a comparable clean, that means handling difficult objects with less help and using fewer operator minutes. Weinzierl told me that for future visits, “It would probably start vacuuming parts autonomously. Still mostly teleoperated.” Lessons from those visits could also guide hardware changes that make battery swaps and delivery faster.

A low-hanging fruit to dial up the throughput would be apartment buildings or student housing. If enough residents book, a pair of robots could move between units throughout the day, with a place in the building to swap batteries and reset the equipment. This can increase the cleaning sessions per day, lower driver costs and could provide repeated practice in similar layouts.

3. Hospitality affects the product

Sergey waved when he came in. I would have liked both robots to introduce themselves. I didn’t initially realize they could respond by voice. When I texted a request to wipe the mirror, they did it.

Those interactions made a difference. So did the noise. They stomped around the house, which took a little getting used to. Another user told me the robots had scratched their hardwood floors when the robots weren’t wearing socks. For our visit, both wore gripper socks.

I accepted the fact that the data would be collected to train the robots. Other users may feel differently. What gets recorded, who can access it and whether a customer can opt out should be clear when booking.

4. What the robot economy looks like

The visit also got me thinking about what would need to exist around a larger fleet of working robots. The driver, battery swap and text message were small parts of my booking. Across a fleet, each becomes an operating responsibility.

Local depots and transport. A neighborhood depot could provide charging, battery swaps, equipment cleaning and repairs between jobs. Could robots eventually travel to the next booking in an Uber or Waymo? Where would robot charging stations be? Will they be in gas stations like Tesla charging ports? Someone would still need to solve loading, securing the robot, unloading and entry into the building. The van trip is part of a longer handoff.

Identity, safety and insurance. A household should be able to verify which company sent the robot and who is responsible for the visit. I would also want clear instructions for stopping it and a service contact if it malfunctions. What safety checks should a home-service robot pass? Who covers damage to the robot, the home or someone inside it?

Ratings and security. Booking history and customer feedback for each robot could help households choose a service and help operators find recurring problems. Connected robots also need controls over who can operate them and access household recordings. A customer should be able to manage access to their data and revoke that access.

Some of this could be handled by the robot company. Some could become services shared across fleets. There will be many new business opportunities from supporting robots doing productive work around humans.

Would I hire Sergey and Chelsea again? Yes, if they can finish the whole house. I would like to book about 2 hours. And I’d be curious to see them fold laundry and unload the dishwasher.

Sources

[1] Weave Robotics, Isaac 1. Advertised pricing, checked October 19, 2026.

[2] 1X, NEO order page. Standard subscription and Early Access ownership pricing, checked October 19, 2026.

Visit observations are from my October 18 home cleaning. Additional capability statements are attributed to Tau Robotics cofounder Cornelia Weinzierl in follow-up correspondence.

Tyler takes us from an early Blue River lettuce-thinning machine with eighteen desktop computers to his work at John Deere and the specialty-crop applications Bonsai is pursuing. When we recorded this episode in Aug 2026, he said Bonsai had deployed about 70 large machines with equipment partners since the prior year. He sees room to build around the jobs and crops those partners already understand.

My favorite part of the conversation was Tyler’s sharing on where the physical world interrupts the model: birds circle a working machine, simulated farm footage acquires a T. rex, and modular robots travel to Australia. Tyler also compares 35 gallons of diesel a day with about 3 for the hybrid electric machine. He also shared practical advice to robotics founders: how do you learn from early field use while protecting customer trust?

In this episode we cover:

• Blue River’s eighteen-computer lettuce-thinning machine

• Camera perception in dusty Australian orchards

• John Deere, specialty crops and machines designed around autonomy

• World models, synthetic training data and cattle egrets

• Air freight, modular hardware and hybrid power

• Utilization, fuel consumption and cost per acre

• Field learning, product readiness and customer trust

Tyler describes the demo as “just 1% of the business.” His advice puts reliability, product-market fit and early paid field experience in the same discussion.

Tyler Niday: https://www.linkedin.com/in/tyler-niday-16a67650

Bonsai Robotics: https://www.bonsairobotics.ai/

Watch on YouTube:

I sit down with founders, researchers and operators building physical AI worldwide, tracing the stack from components and supply chains to deployment. See the published episodes on The Core Matter Show page.

Recorded at Reservoir Farms, Salinas, August 19, 2026.

Chapters

00:00 The demo, the farm and the fuel bill

00:20 Blue River and 18 desktop computers

01:36 From computer vision to learned models

03:00 Dust and the Australian orchard

04:27 Cameras, lidar and working distance

05:57 Why a software company builds machines

08:06 John Deere and the specialty-crop opening

09:09 A mixed fleet of tractors and Amigas

10:12 World models and simulated farm footage

12:06 The birds that stopped the machine

14:20 Why Australia needs farm autonomy

15:36 Shipping robots to Australia

16:50 Hybrid power and the charging constraint

18:34 Utilization and cost per acre

20:20 Orchard units and the Amiga fleet

21:40 Amiga Flex, Amiga Max and the job

22:50 Field learning and customer trust

24:27 The 2031 vision for individual plants

Transcript

[00:00:00.000] Tyler Niday: The demo is just 1% of the business, right? So we can simulate the whole farm before we ever get there. It had 18 desktop computers on the back of a tractor doing lettuce thinning. Instead of burning 35 gallons of diesel a day, it burns about three.

[00:00:20.367] Michelle Sun: So take me back to Blue River.

[00:00:22.720] Tyler Niday: And then I came across Blue River, and it was the most magical ag machine I’d ever seen. It had 18 desktop computers on the back of a tractor doing lettuce thinning. Lettuce thinning is a process that happens right outside here in Salinas that historically, it was about 20 to 40 people with a hoe going through the field and thinning lettuce. The lettuce is planted at like a two-inch spacing because the germination rate of the crop is poor. And then you thin it out to grow.

[00:00:47.160] Tyler Niday: So this machine had 18 desktop computers. It was pre-AI on the back of a machine doing thinning, running at four miles an hour. It was the most beautiful machine I’d ever seen in the agriculture space. It was very, very special. Fast forward, we had a service business. AI came out. I think we were one of the first companies to use the NVIDIA TX1 chip using YOLO-based deep learning. And we built an actual green-on-green detector so we could differentiate weeds from crops. And then the acquisition happened, which was pretty special and awesome.

[00:01:15.820] Tyler Niday: And got to work with one of the best companies in agriculture, John Deere. And then got to go in and build out the autonomy team with the now CEO. We’re focusing on big, large-scale corn, cotton, and soy tractors. So really cool process. A bit crazy to put 18 desktop computers on the back of a tractor and run a service business. But it was a pretty special machine to watch work.

[00:01:36.780] Michelle Sun: You’ve really seen the pre-AI architect to today. So what did you learn in that period that you’re using building Bonsai Robotics today?

[00:01:46.980] Tyler Niday: It’s been crazy to go from early-day computer vision, right, and OpenCV algorithms to detect green versus brown to where we’re at right now with foundation models and world models. And what I learned is that you really have to look to biology to inspire how to build the system. Historically, robotics have been built in a very deterministic way. You write hundreds of thousands of lines of code. If this happens, then this happens. Now with current day and trends, you can do these much more complex tasks that are trained on data from expert operators to enable much bigger platforms.

[00:02:19.860] Tyler Niday: So the big differentiator that we’re doing now is that we’re using these foundation and world models to have one platform that can scale to all the specialty crops and beyond. So I don’t need to write a different code stack for almonds versus the lettuce outside versus the strawberry crops. We have this big platform with one model with no code changes that can go into all these diverse crops. And it works in dust. It works with occlusion. It’s not a deterministic stack. For example, if you had a time of flight sensor and you’re looking at an obstacle and there’s a limb covering it, you’re not going to be able to see through that.

[00:02:51.620] Tyler Niday: Whereas with a biology-inspired learned model, you can infer that something’s there from a trained model, so it’s a pretty exciting time.

[00:03:00.520] Michelle Sun: So we have machines like these on the field and tell me about conditions where you didn’t expect that the machine break. Is it the dust or is it random environmental factors?

[00:03:14.000] Tyler Niday: So we are a full stack robotics company for the rugged world and we started in probably some of the most adverse conditions to make a big machine learning end-to-end model work. The almond orchards of Australia. It’s removed. There’s limited cell service. You don’t have GPS connection. It’s in sandhills. And the amount of dust that’s generated is crazy. So we had to solve this problem so that we could scale to these other crops. I think some big examples are dust, right? Traditional sensors and approaches with lidar-based or stereo vision-based stacks don’t work that great in dust, right?

[00:03:49.460] Tyler Niday: You’re blinded by it. So we tried to model the system like a brain works, right? Where we can infer things and see through things ultimately. That was one big challenge. Yeah, I think our stack is really unique in that we’re a full stack. So we don’t only build our own platforms that go beyond labor. Like this machine removes labor from the equation of a grower, but additionally reduces the cost of the machine and also the operation expense. This is hybrid electric. Instead of burning 35 gallons of diesel a day, it burns about three.

[00:04:18.280] Tyler Niday: So there’s a lot of intricacies in that platform and making a full stack robotic system work that you have to have a lot of great experience to learn through ultimately.

[00:04:27.240] Michelle Sun: You mentioned about some conditions where, you know, the lidar doesn’t work. And so I noticed that you also picked using cameras. So talk to me more about that design choice.

[00:04:37.020] Tyler Niday: I think modeling robot behavior as a human brain works is really good. So we’ve collected a crazy amount of camera data, right? And we’re able to do a lot of things with that. We see the world in color and can associate a 3D environment around us from a learned model. In a lot of our applications, we don’t have to see things super far out very accurately. We’re worried about the next 30 or 50 feet, right? So lidar is a super great tool at seeing things super far out very accurately and creating a really good map.

[00:05:08.880] Tyler Niday: It’s a bit expensive. Prices up came down. But historic methods are much more deterministic. You now can do full end-to-end lidar models, which is great too. But we really wanted to build a reliable, robust solution modeled after a human that works well in these conditions, right? That can infer through dust, right? And visualize the whole world around you accurately like a human does.

[00:05:30.160] Michelle Sun: So how many cameras are there on a robot?

[00:05:33.220] Tyler Niday: Yeah, it depends on the application. This machine has four cameras on it behind us. Some of our platforms have quite a bit more depending on the size of the array. And with our bird’s-eye view model and end-to-end models, we’re able to input these cameras to create the 3D world for us to navigate through prior to feeding it into the world action model to actually go deeper and figure out how to interact with that environment.

[00:05:57.400] Michelle Sun: You mentioned that the industry is moving from iron to intelligence. And you then also operate machines as well. So how do you think about owning a hardware versus going to the model side?

[00:06:11.000] Tyler Niday: Yeah, we’re a full stack company. We build robots. We build world models and foundation models and the complete autonomy layer. Our company started just as a software platform. So we started just building the autonomy stack. We have a very intuitive app that not only enables you to control the autonomous machines, but manage your farm. You have all your operational data. This data comes not only from the autonomous machines, but we have a compute box that you could outfit the rest of your manually driven fleet with as well.

[00:06:37.580] Tyler Niday: So getting that data is critical and informing the user to be able to use, navigate and make informed decisions is a big starting point. Autonomy has to be very reliable and very robust. And we decided to start the business by scaling with our channel partners. We have Orchard Machinery Corporation and Flory that makes really good hardware so we can stay laser focused with the capital we had on the software, right? So that the app, right, the whole perception stack, the AI models and collect tons of data on these platforms.

[00:07:08.000] Tyler Niday: We’ve now deployed about 70 of those big OEM units since last year with those partners. And then what we realized was there was a huge opportunity to design platforms around autonomy versus around a human. Machines got bigger and bigger and bigger to be optimized around a human’s performance. That was the ROI you would get is a bigger machine. And they got really expensive. And if you design the machine around the autonomy system, you’re not only replacing the labor aspect to the grower or end user, but you’re also reducing the cost of the machine.

[00:07:40.900] Tyler Niday: You’re making it easier to work on. You’re reducing the amount of fuel you have to use. So there’s a much bigger savings if we built the full stack on these focus specific tasks that could get to 4,000 hours of utilization a year than just being a software player. You definitely need big tractors and horsepowers for specific things. But there’s a niche, mostly in specialty, where we believe autonomy first form factors can really scale.

[00:08:06.340] Michelle Sun: You know John Deere from the inside, right? So what does a company your size have advantage over a company like John Deere?

[00:08:13.740] Tyler Niday: You know, John Deere is a spectacular company. And I learned a lot there. And they have a bunch of very, very good people. I think one of our advantages is that we are able to focus on industries that they might not put all the attention into up front, like the specialty aspect of vineyard, strawberry [unclear], right, which you saw earlier. That’s an application that I don’t think John Deere has a ton of interest in, but it’s a big, big market for us right now.

[00:08:38.360] Tyler Niday: And there’s a huge amount of labor and replacements that we can go. Additionally, we can move very fast, right? We’re a four-year-old company. We can make decisions quickly and adapt new technology and put it in the field super, super fast. We have 60 people instead of 70,000 people. So making a decision might be a little faster. We might not always make the right decision, but we can get it out there very quickly. And I think that’s a big advantage as well.

[00:09:03.040] Michelle Sun: So more agile and focusing on specialty problems to really get on the field faster. And so we talked about how the company’s full-stack robotics, and so with both the world model going into other machines, but also you have Amiga here, your own machines on the field. Where do you see the split of the company going in, let’s say, the next three to five years?

[00:09:27.820] Tyler Niday: Yeah, so right now the split of existing OEM machines, manufacturing machines, and Amigas is about 50-50. And I see that continuing for the next three years. After three years, I’m not really sure. I think maybe we’ll go more heavy into Amigas as autonomy and implements get more advanced to be able to do more jobs on the farm. Tractors are the ultimate Swiss Army knife, right? You can do everything with it, and you’re going to need tractors to do specific things. So I think it’ll stay about 50-50.

[00:09:56.060] Tyler Niday: We need both. We need these platforms that could drastically improve the ROI of very repetitive tasks. But you also need tractors to do those specific applications. So I think it’ll be this combination of a rainbow fleet to really make the farm succeed and really reduce the bottom-line cost.

[00:10:12.000] Michelle Sun: Tell me more about the biology-inspired world model. And what kind of capability does it have right now? Like, where does it work on the farm? And where is the limitation?

[00:10:21.840] Tyler Niday: Yeah, so we use it in two aspects, right? Our world model is feeding our generative sim data. So we can actually create a field and run a platform. But we also use it to augment and train data. We have a full automated 3D labeling pipeline for our foundation model, right? And that, like, we can generate copious amounts of data for new environments or new implements very quickly. I think our world models for generation of a scene normally starts with an HD map in self-driving cars.

[00:10:49.860] Tyler Niday: Farming usually starts with a 2D map. So we have this world model platform where you can take a farm anywhere in the field from, like, a Google Maps view, And then you generate the full map of the farm, so a [unclear] boundary. You can then plan a path, and then it’ll translate that to an almost picture-perfect scene, right? So we can, like, actually see what our cameras will see, and people can’t even tell the difference. We made a joke of one of our cameras, and we had an elephant walk in front of the machine, right?

[00:11:17.920] Tyler Niday: And I thought it was obvious to people that it was fake, but so many people are like, where are you guys running, you know? And so then we made a T. rex one to just make sure that people actually got it. So we can use this to simulate edge cases, to train new data, but really generate things we’ve never seen before. So we can have the full stack run in sim, thinking that it’s actually out in the real world, Generating dust data, generating holes like negative obstacles and [unclear], and actually seeing what the machine will do.

[00:11:48.340] Tyler Niday: So we can simulate the whole farm before we ever get there. But additionally, let’s say we were going into a fully new environment, full new crop, we can generate that and actually train our models prior to even getting out there and running. So there’s two big aspects we can use it for there. It’s pretty fun.

[00:12:06.800] Michelle Sun: Tell me a story where a machine surprising you. What happened? Like, what did you change afterwards?

[00:12:13.420] Tyler Niday: Yeah, I think the bird one would be a great one. And so this has actually happened to me twice in my career in autonomous ag so far. The first day we were engaging the implement, the job being done. We were chisel plowing a field and we were digging up the ground. And when you dig up the ground, if the soil is fertile, there’s a lot of worms in the ground, right? And what likes worms? Birds. So we had this beautiful system that worked great.

[00:12:38.040] Tyler Niday: We had this amazing segmentation model. And then all the birds were up and we were moving, right? So the birds would just circle the machine. And I don’t think the machine moved more than 10 feet in a 10-hour period, right? Because we’d have all these obstacle stops. Similarly, we were recently deploying with a large grower in Australia and we’re doing slashing. Slashing is the Australian word for mowing in ag. It’s a better word. I think Australians got some good words. I like it.

[00:13:02.780] Tyler Niday: But when you slash, there’s also worms and bugs that are brought up, right? So we had this bird called the cattle egret. And these birds circled the machine and our system was detecting it. It hadn’t seen this before and was stopping. We didn’t have a lot of data historically to train on these birds, right? To make our system at 99.x9s of reliability to drive through these birds. So we used the world model to create copious amounts of data of these cattle egret birds.

[00:13:29.540] Tyler Niday: And it was all auto labeled through our 3D pipeline and then could be deployed in a very short period. We’re able to use this world model to solve these edge case problems that you wouldn’t get until you put a new implement on the machine, until you’re in a new environment and discover it. And I think that will also carry on where the world model can actually get to these applications. Right now, I think a nice parallel is humans have to have dreams to survive.

[00:13:54.280] Tyler Niday: It’s kind of weird, but it’s been proven that if you don’t dream, it’s like this off-channel learning cycle and you’ll go crazy, right? And your dreams are normally crazy edge cases, right? They’re not normally going to happen, but you wake up some days and it’s crazy. I think right now that’s where world models are at, right? You can fine-tune what you’re dreaming about. So you can get these edge cases and you need it as a function for this biology-inspired model to thrive and get past just discovering it in the field.

[00:14:20.220] Michelle Sun: You mentioned Australia quite a bit, and how did you decide to go all the way there?

[00:14:26.840] Tyler Niday: Yeah, well, A, I just like Australia and I think the people are awesome. But apart from that, I think it’s ripe for autonomy. I’ve been going to Australia for 15 years, and so we’re in Salinas right now. Salinas is pretty close to the San Francisco Bay Area. But when you drive to the Central Valley, it’s a little more remote. But generally, to get to a farm, it’s about two hours, right? When you’re in Australia, the closest farms that we work on are sometimes 10 or 12 hours from a major city, and there’s nothing in between.

[00:14:55.360] Tyler Niday: These farms are generally planted along a river, and they’re huge, right? So labor is even more scarce than in the United States. It’s really hard to find someone to operate these machines, but they’re massive operations. Some almond orchards we’re on are 5,000 or 10,000 acres. Like, we’ll start a machine up at the shop, and it’ll have to do its autonomous route to the field. It’s going to spray, and it’s over two miles of autonomous navigation. So there’s a huge need there for autonomy to keep agriculture thriving.

[00:15:23.020] Tyler Niday: And historically, Australia has been about five years ahead on the ag tech adoption curve of new technology. So people need it even more so than the United States right now. So I think it’s a prime hotspot where you can scale quickly.

[00:15:36.720] Michelle Sun: Getting these machines all the way from Salinas or in the area all the way to Australia. So how do they get there? It’s a pretty sizable parcel.

[00:15:46.720] Tyler Niday: So historically, you pack it up in a shipping container and ship it. And we learned this from our OEM partners. And it can take six weeks, right? And then you have to go through customs. However, our Amigas, we actually have air freighted there too. So the little ones we’ll put on a plane, and it’ll be there in two days, which is a bit wild. These ones are quite a bit bigger, and the cost is substantial to ship it. But we have air freighted these Amiga Maxes to Australia as well.

[00:16:09.900] Tyler Niday: You couldn’t do that with a full-size tractor. I think we’re on the upper cusp. The base platform weighs about 1,300 pounds, but pretty cool you can do this. And I think another aspect is these are modular. Like, these are powered, steerable wheels. So it’s really easy to put together. Tractors normally have hydrostatic hydraulic lines. So like, literally, there’s fluid being pumped across the tractor, right? It’s inefficient, but also a very complex system to manage, and you couldn’t assemble things there. This is a robot where it’s all electrical connections, and you can bolt it together and plug it in.

[00:16:41.620] Tyler Niday: It’s like big Lego. So it makes some of these applications doable, where you could flat pack it and get it somewhere in a quicker time than you would with a normal big machine.

[00:16:50.860] Michelle Sun: And you mentioned about the uptime as well, right? So that’s also intentional design.

[00:16:55.260] Tyler Niday: Yeah, most definitely. So my big thesis is electrification of machines has a big impact on the efficiency. Like, I had a Tesla for 100,000 miles, right? And I only changed the windshield wipers, the tires, and the air filters, right? It was very minimal, the amount of work, versus now I have a big F-250 truck, right? And I have to take it in for an oil change. There’s things that go wrong all the time. So I think electrification is a huge win on the efficiency and reliability of these big machines.

[00:17:25.720] Tyler Niday: Tractors are either massive mechanical assemblies with a ton of parts that can fail, or hydraulically driven machines that also have a ton of parts. So electrification simplifies it. However, I think you need farming never stops. Some of our machines run 4,000, 5,000 hours per year. If you needed to stop and charge, recharge a machine during the critical time, it could kill an operation. For a big John Deere 8R series tractor, it would take 13 Tesla Model 3 batteries to run for a few hours, right?

[00:17:55.180] Tyler Niday: It’s not, horsepower needs power. So our big thing is you need to be able to get a machine back up and running in under 10 minutes. And we believe there’s a threshold for how big a battery can be on the platform. It needs to be swappable. So our little machines have swappable batteries. Or these bigger applications are hybrid electric. So this has a small diesel generator on the front that’s trickle charging the battery. So as the battery declines, we just add a bit of power and keep it charged all the time from a diesel power source.

[00:18:22.700] Tyler Niday: And instead of using 30 gallons of diesel or 35 gallons of diesel a day, you’ll use three. So there’s a big, big win, especially with the rising costs of diesel and different fuel sources.

[00:18:34.220] Michelle Sun: So you mentioned about 4,000 hours a year. How does the math work out for a farmer, a farm owner that is buying a machine?

[00:18:40.820] Tyler Niday: Yeah, it all comes down to utilization, right? And if you’re just replacing an operator on an existing tractor, you have to have high utilization to make it pencil out per whatever your autonomy costs are or your subscription. This is where we think these Amigas are very unique in that we’re redesigning the form factor around AI and autonomy instead of designing around a human and a cab, right? Like the cab of a tractor is a lot of money, right? Sometimes there’s like refrigerators in it.

[00:19:08.820] Tyler Niday: There’s amazing sound systems, right? There’s a full suspension seat. It can be a big cost of the tractor. Removing that’s a big aspect. So these platforms have a much quicker ROI than just a straight replacement of a human operating the machine because we’re not only adding the labor aspect with higher reliability, high performance autonomy, but we’re able to make the machine cheaper on a per acre per pass basis by reducing the cost of the machine with electrification, but also drastically decreasing the OPEX around repairs and fuel consumption.

[00:19:41.440] Tyler Niday: So it’s a win-win where we’re kind of stacking all aspects on the per hour per acre operational cost ultimately of the machine.

[00:19:49.720] Michelle Sun: And is that the number that the growers think about as well?

[00:19:52.740] Tyler Niday: Yeah, they are running like a full profit and loss statement based on metrics to make sure it pencils out per their machine matrix, right? Some big operations have 5,000 tractors, right? And it’s not like one tractor, right? There’s like of that 5,000 tractors, there’s probably like 50 to 60 different form factors, right? And getting that matrix right is a big aspect of it. So they’re definitely running very detailed ROI calculations based on real metrics, based on the machine performance to get to an answer.

[00:20:20.920] Michelle Sun: 400 units right now.

[00:20:22.440] Tyler Niday: Yes.

[00:20:23.060] Michelle Sun: Tell me about the split in terms of different farms, different countries. Where are you seeing the most demand coming from?

[00:20:28.780] Tyler Niday: Yeah, most definitely. So this is a new product. We’re very excited about this. This is an Amiga Max and we’re seeing a large demand from this. But historically, we started deploying Bonsai-enabled orchard units last year. Since then, we’ve sold about 70 of those units. And these are big tractors, right? Like about 4 to 5x the size of the one behind us. We see a lot of those going to Australia and California. So about 70 of those units and the rest has been Amigas.

[00:20:57.780] Tyler Niday: Amigas originally were as much smaller footprint, great for specific applications, did a lot of work. But we’ve now expanded and put the full Bonsai autonomy stack on these Amigas and also increased the size of the Amigas. So this is a full new market we’re breaking into as well, which we’re really excited. We were in nine countries globally right now, but we’re seeing predominant focus for the next few years in Australia and the U.S., mostly California, Washington, and Florida.

[00:21:26.580] Michelle Sun: So no Europe yet?

[00:21:27.860] Tyler Niday: There’s a couple machines in Europe and we’re planning to expand towards the end of next year to Europe. We think it’ll be a big use case, but we want to make sure we nail these markets first as we have the full business side of the equation set up in these industries.

[00:21:40.860] Michelle Sun: So you mentioned that there was a small Amiga and then this one Amiga Max. What are the unlocks in terms of capabilities that the Max version is able to use?

[00:21:50.960] Tyler Niday: The small one. The Amiga Flex is a small unit and it’s really about low cost autonomy. It’s a fixed platform. It’s very simple and it can go into a bunch of different applications. We’re deploying, I think there’s one at the Space Force right now to different ag fields across the road. That machine also can do specific applications. Like we have a spot spraying application. So it has a tank of material and instead of spraying the whole ground, it just sprays the plants.

[00:22:16.160] Tyler Niday: The big unlock and the bigger form factor is we can do more and compete much more in these bigger applications. This one’s actually set up to spray apples, right? And a little platform can’t carry 300 gallons worth of material, right? And also can’t power a big, you know, fan to blow the spray out in these applications. So that’s one of the aspects. But additionally, this machine has full articulation so it can go over much more adverse terrain, bigger tires and four wheel steering.

[00:22:42.780] Tyler Niday: So it can get into different hills, conditions and edge cases much quicker than some of the little platforms or even the competitive tractor.

[00:22:50.820] Michelle Sun: There’s been a lot of development in the ag tech industry and lots of physical AI startups coming in. And for someone like yourself that’s been at it for quite a while, like what are some things that you would advise like new startup founders to think about?

[00:23:04.640] Tyler Niday: Yeah, you know, it’s really exciting that robotics in these ruggedized environments is mainstream because it’s such a huge opportunity to solve for. The biggest thing I would advise is that the demo is just 1% of the business, right? Getting past the demo and actually building repeatable, scalable, reliable products is such a big challenge. And you want to be centric around that and not ship machines early, right? Like make sure your product is steady, great, reliable and has product market fit before you jump the gun and continue out past that.

[00:23:38.100] Tyler Niday: So I think that’s a big aspect. And it’s always a battle, right? Like you’re driving for revenue, but your machine has to work. And finding that sweet spot in between the two is a critical moment that has to happen to be successful.

[00:23:50.980] Michelle Sun: Getting the machine on the field is also important to get it better. So there’s this tradeoff to get on the field and still retain the trust to the customers.

[00:24:01.540] Tyler Niday: I think that’s another huge aspect that I’m always pushing my team very hard on. It’s like we’re more or less like getting to the field early because there’s so much new things you learn in the field versus in the shop. You can do every test in the shop you could imagine, but until you see that machine run and get a little bit of paid customer traction, you don’t actually know if it’s a product that’s going to work in scale or not. So get to the field early is a big one for sure.

[00:24:26.160] Tyler Niday: I’m glad you brought that up.

[00:24:27.680] Michelle Sun: By 2031, let’s say it all works out the way that you envision for Bonsai. How does it look like? Who owns the machine? What’s running on the fields?

[00:24:36.900] Tyler Niday: So by 2031, I’m hoping that not only ag, but beyond is the whole operation system of work is reimagined holistically around AI instead of around humans operating machines. The vision of Bonsai was started in ag, right? And we believe autonomy is the enabler to instead of managing fields and having these machines run through every field, We can manage plants and give every plant the master bonsai treatment. There’s competition corn growers in the Midwest that produce 3x the amount of corn versus, like, the top growers, like 600 bushels of corn instead of 200.

[00:25:10.640] Tyler Niday: They’re using the same energy source, the sun, right? But they’re treating the plants differently. And I think autonomy going through the field, you have eyes on every plant sometimes 50 to 100 times a year. You can start managing these farms on a plant level basis and really improve the efficiency on a per acre basis. I’m very excited for that. Farming is one example, but this same thing can be applied to all these outdoor rugged industries, right? Depending on the job you’re doing, whether you’re in mining or construction, you can really optimize the workflows and make things a lot better.

[00:25:42.840] Tyler Niday: And I think autonomy and these big foundation models are the enabler to do that.

[00:25:48.940] Michelle Sun: Thank you.

[00:25:50.620] Tyler Niday: Thank you so much.

[00:25:51.460] Michelle Sun: That was a great conversation.

Robotics has a “100,000-year data gap,” according to Berkeley professor and Ambi Robotics cofounder Ken Goldberg. His 2025 comparison put the human-equivalent reading and viewing time behind large vision-language models against roughly one year of teleoperation data. [1]

Suppliers are moving quickly. Mecka announced a $60mn Series B on October 7. XDOF emerged from stealth in June with $70mn in funding. On October 9, microagi said it had reached $100 million in ARR within six months of launching Shift. [2], [3], [4]

In this piece, we look into how collecting data builds the operating muscle and customer relationships to sell evaluation and deployment. That can turn a dataset order into recurring revenue. The question is how much becomes recurring profit after engineering, hardware and support costs.

Sections

1. Seven ways to close robotics’ data gap

2. Beyond datasets: selling evaluation and deployment

3. The financial comparison: Scale AI and its robotics challengers

4. Turning recurring revenue into recurring profit

5. LLM vs. robotics data: what drives delivery costs

6. What if the Scale AI of robotics becomes bigger than Scale AI itself?

Seven ways to close robotics’ data gap

What do robotics models train on? The main sources fall into seven overlapping categories:

1. Internet video: existing footage of human activity, usually without robot action labels.

2. Egocentric video: first-person recordings of people performing tasks, sometimes enriched with additional sensors.

3. Handheld capture: demonstrations recorded with portable tools such as UMI’s camera-equipped grippers, without a robot at the collection site.

4. Robot teleoperation: a person controls a robot while its actions and sensor readings are recorded.

5. Physics simulation: virtual robots interact with simulated objects and environments.

6. World-model experience: learned models generate predicted scenes or outcomes, sometimes conditioned on actions.

7. Deployment experience: recordings of working robots, including successes, failures and interventions.

Three tradeoffs in robot data

Three key qualities determine how useful robot data is. Each collection method has different tradeoffs:

• Scalability: How data scales with labor, equipment and cost. Teleoperation needs robot capacity and operator time. Egocentric and handheld capture remove the robot from collection. Simulation generates repeated trials without an operator for each run, but still requires compute, setup and validation.

• Physical fidelity: How accurately the data represents embodiment, motion, contact and outcomes for the target robot.

• Diversity: Coverage of tasks, objects, environments and conditions, including failures and recoveries.

What drives demand for robot data

Labs need broader coverage as they tackle new tasks. And data strategy is already a competitive concern in the LLM market. This translates into the robotics market.

• Competitive pressure. “The labs don’t want the other labs to figure out what data they’re using to make their models better,” Handshake CEO Garrett Lord told TIME about LLM training data after Meta’s Scale investment. [5]

• Limited coverage. Beyond Goldberg’s 100,000-year comparison, hardware, labor and logistical coordination limit collection across environments.

• New tasks and failure modes. When model labs look to train for new tasks, robot configurations and difficult cases, they need additional demonstrations.

Utilization and quality determine the economics of data plays

Labor cost is only one lever for a data collection operation. Data collection economics depend on staff and equipment utilization, and on the share of recorded data that passes quality checks.

Cost per accepted trajectory-hour = total collection and processing cost ÷ accepted trajectory-hours.

Consider a hypothetical station with one operator costing $30 per staffed hour. If recording occupies 60% of staffed time and 80% of recorded duration passes quality checks, labor costs $62.50 per accepted trajectory-hour. At 30% recording time and 60% acceptance, that rises to $166.67, a 2.67X difference. Both cases exclude hardware, premises, separate processing staff and overhead.

Tooling, infrastructure and software that improve utilization and data quality can reduce the cost of usable output. One example is developing specialized models to annotate video data more accurately. [16]

Beyond datasets: selling evaluation and deployment

A robotics data supplier operates hardware, trains operators, sets up tasks and checks outputs. Those capabilities can support policy evaluation and, with additional integration work, deployment. These enable a data supplier to sell to two types of customers:

1) The model lab needs to test and improve a policy. It needs to compare checkpoints, reproduce failures, collect corrections and measure progress. New releases and regression tests can generate repeat work on the same stations and procedures.

XDOF lets customers select tasks, test checkpoints and receive results and trial recordings. It also offers corrective collection. Its evaluation service currently specifies bimanual YAM arms from I2RT, so infrastructure reuse depends on the robot and task. [6]

2) The factory operator needs a reliable production task. Parcel sorting, for example, requires the model, arm, sensors and workflow to function together. Repeat work can come from additional cells, new sites, changing products and ongoing support. Integration knowledge may transfer, while site conditions require fresh engineering.

microagi’s investor Redalpine describes recreating a customer’s workstation, purchasing hardware and training the system before renting the working installation to the customer. That model combines continuing revenue with equipment financing and operating responsibility. Other deployments may use customer-owned hardware, as microagi’s current site describes. [7], [8]

Mecka states: “We bring the hardware, data capture, post-training, integration, and ongoing operations.” microagi’s stated ambition is to deploy one million robots. [2], [9]

The financial comparison: Scale AI and its robotics challengers

Revenue, gross margins, funding and valuations show the scale these companies are pursuing. The reporting periods and definitions differ.

[10]

Read more

My sensors were working just as hard. My eyes scanned for the gate sign and gaps in the crowd. My body adjusted to the carry-on and backpack. My fingers kept hold of the lukewarm coffee I’d bought in Pittsburgh hours earlier. My ears listened for the final call.

If I were a humanoid, making that tight connection while keeping hold of my coffee would require all these sensor feedback working together. That’s the topic of today’s post - the senses of a robot. In autonomous vehicles, Waymo combines cameras, lidar and radar, while Tesla has built its driving system around camera-based vision. Sensor choices change what the hardware measures and what software has to infer.[1]

Recently, I asked a Formic deployment engineer what broke most often. His answer was sensors. Jake Panikulam of Main Street Autonomy, a startup that focuses on sensors calibration, shared with me that dirt on sensors as a source of failures.[2]

Let’s go through the six senses, their suppliers, and where deployment breaks. Then we’ll ask what technical progress would make robots more useful, and where specialists can compete with established suppliers.

In this piece, we cover:

• What sensors do humanoid robots use

  • Sight: what’s around me?

  • Depth: how far away is it?

  • Balance and body position: where are my limbs, and how fast am I going?

  • Force: how much load am I carrying?

  • Touch: is the cup slipping?

  • Hearing: is the boarding gate closing?

• Where deployment breaks

• What technical progress would make robots more useful?

• Will automotive suppliers lead, or can specialists win?

What sensors do humanoid robots use?

Sight: what’s around me?

How it works: A video camera captures images many times a second. 30 fps (frames per second) means 30 images each second. During a vision-guided task, the camera supplies a live stream of frames. Software then uses those images to recognize the gate sign, the cup or the person walking into the robot’s path.

What it measures: Light at each pixel, which software interprets as objects and scenes.

Where it’s located: Head cameras give a view of the surroundings. Wrist cameras can give a closer view of a grasp, although the hand or object can still block them.

Suppliers: The supply chain spans image-sensor chips, camera modules and integrated perception systems.^[1]

Depth: how far away is it?

What it measures: Distance to surfaces across the scene.

How it works: Three common depth-camera approaches are:^[2]

1. Stereo compares two views taken a known distance apart. Nearby objects shift more between the images than distant ones.

2. Structured light projects a known pattern and uses where it appears in the camera image to calculate distance.

3. Time of flight sends out light and measures its return timing, directly or through a change in phase, to estimate distance.

Many lidars use time of flight, the third approach above: they send out laser pulses and time the reflections. A spinning lidar sweeps its view around the robot, whereas a MEMS-scanned lidar steers light with a tiny moving mirror inside the unit. Scanning decides where it looks. Ranging determines how far away something is.^[3]

Where it’s located: Depth cameras can sit on the head, torso or wrist, depending on the view needed. Unitree’s G1 combines a depth camera with 3D lidar.^[4]

Suppliers: Hesai and RoboSense supply lidar. Orbbec and RealSense supply depth cameras, with some suppliers combining several sensing functions in one module.

Balance and body position: where are my limbs? and how fast am I going?

Proprioception is body awareness. If you know whether your elbow is bent with your eyes closed, you have proprioception.

How it works: Encoders read an optical or magnetic pattern as a joint rotates. Then, software takes joint angles, combined with the robot’s limb lengths, to calculate where its hands and feet sit relative to its body.^[5]

An inertial measurement unit, or IMU, adds acceleration and rotation-rate measurements, with accelerometer, gyroscope.^[6]

What it measures: Encoders report joint position, while IMUs report acceleration and rotation rate. Software combines IMU, joint and contact measurements to estimate motion and orientation, then the controller adjusts the motors to keep the robot upright.

Where it’s located: A body IMU can be mounted in the torso or pelvis, while encoders sit at the joints. The exact mounting depends on the robot.^[5-1]

Suppliers: Bosch and ST supply IMU chips. VectorNav supplies inertial modules. Renishaw/RLS and Kübler supply encoders.

Force: how much load am I carrying?

How it works: Imagine a tiny metal spring inside the wrist. When the robot lifts or pushes something, the metal bends a little. Small electrical sensors attached to it, called strain gauges, stretch with it. Stretching changes how easily electricity flows through them. After testing with known loads, the robot can translate that change into how hard it is being pushed, pulled or twisted.^[7]

What it measures: Push and pull are forces. Torque is their turning effect around an axis. A six-axis sensor measures pushes and pulls along three directions, plus twisting around those three directions. Force is measured in newtons. Torque is measured in newton-meters.

Where it’s located: At the wrist, it senses the combined load passing between the hand and arm. Ankle sensors can measure loads from the ground. Joint torque sensors are built into the joint or actuator, for example at the elbow or knee, to measure the twisting load transmitted through it.

For the coffee cup, wrist force sensing helps track the overall load, while fingertip touch tells me whether my cup is slipping.

Subscribe now

Suppliers: ATI, Bota Systems and SRI.

Touch: is the cup slipping?

How it works: Tactile sensors turn contact at a surface into a signal. An optical design uses an internal camera to watch a soft material deform. A magnetic design detects how embedded magnets move as that material is pressed or sheared. Other designs use changes in resistance or capacitance.^[8]

What it measures: A taxel is a sensing site, analogous to a pixel. An array can map contact across a fingertip or palm. What it reports depends on the design. Some measure local force, others capture deformation that software interprets. Temperature requires a temperature-capable sensor.

Where it’s located: Besides fingertips and the palm, touch can extend to feet, arms and the torso. That creates coverage beyond the camera’s view, but also more surfaces, wires and connections to maintain.

Suppliers: GelSight, XELA, PaXini, Daimon and Tekscan use different sensing approaches.

Why do tactile sensors need replacement?

Soft material deforming is part of how these sensors work. Wear happens when abrasion, tears or changes in material response alter the signal. Replacing the skin or coating may require calibration or model checks. Magnetic interference can also disturb readings, but is a separate issue from physical wear. Service life depends on the material, load and task.^[9]

Hearing: Is the boarding gate closing?

How it works: Microphones capture sound. Software interprets voices or alarms. A robot can use a wake word like an Amazon Echo, a button, or continuous command recognition.^[10]

What it measures: An audio signal.

Where it’s located: The microphone is usually on the head or torso, with motor noise and vibration influencing the design.^[10-1]

Suppliers: Infineon and ST supply MEMS microphones.

Where deployment breaks

In deployment, measurements must line up in space and time, reach the controller quickly enough, and remain usable after repeated use and servicing. A missed task can come from the sensor, its calibration, the processing pipeline or the robot’s response.

The airport trip makes three failure modes concrete:

Occlusion and timing: Someone steps into its path. The robot needs a usable view to see the travelers around it, an accurate distance estimate and enough time to change course. A blocked camera or delayed measurement could leave it reacting too late. It needs to detect and avoid the person at its walking speed. A high camera frame rate helps refresh the scene, but processing and motor response also take time.

Grip feedback: The cup starts slipping. As the robot turns around a suitcase, the cup shifts in its hand. Its wrist sensor measures the overall load, while fingertip sensors help detect the slipping contact. The controller needs to adjust the grip before the cup falls, without squeezing hard enough to crush it. The robot needs to complete the trip without dropping, spilling or falling over.

Contamination and wear: Performance degrades during use. A smudged lens could make the gate sign harder to read, much like dirty glasses. Repeated use can also wear tactile surfaces or change calibration. Waymo’s built-in sensor cleaning offers a relevant lesson: keeping sensors usable is part of deployment.^[11]

What technical progress would make robots more useful?

When it comes to sensors, it’s about getting “good enough” so it’s no longer a blocker for capabilities. Cameras, IMUs and force sensors are established product families from other industries, but their specifications still need to meet the robot’s operating conditions. ^[12] Below is a map of the gaps in the sensors offering today for robotics capabilities.

Read more

Shan is one of the professors whose name was mentioned frequently when I speak with robotics and tactile researchers. He has chaired ViTac workshpos for X consenstuive years. So he’s seen the tactile industry evolve close-up.

The question that stuck with me was about replacement skin. A tactile sensor’s soft contact layer wears with use. Replacing it can change the material properties and alignment that a force-prediction model depends on. Shan’s GenForce research helps transfer knowledge learned with one sensor to another, reducing the need to collect force-calibration data again.

We also discuss his work with Unilever on assessing hair and fabrics, the force-control problems he describes with Ocado, and how the task should shape the finger, sensor and control policy together.

There is an Easter egg near the end: I ask Shan what he would build if he started a company. His answer includes a surprise announcement. I’ll let him tell you.

Topics:

• The contact information behind material assessment and manipulation

• GenForce and learning across different tactile sensors

• ViTacGen, object pushing and learning-time touch

• Replacement skin and the challenge of preserving calibration

• Designing the sensor, finger and control policy together

• Industrial research with Unilever and Ocado

• Open datasets, shared benchmarks and a surprise announcement

GenForce research · ViTacGen research

Shan Luo · King’s College London profile

Watch the conversation:

Explore more conversations on The Core Matter Show.

Recorded remotely on August 21, 2026. Published October 2, 2026.

Chapters

00:00 Cold open: what happens when robot skin changes?

00:29 Why Shan Luo chose tactile sensing

02:04 Scaling tactile data and simulation

03:40 The interface between different sensors

05:04 Material assessment, cleaning and insertion

09:18 Vision, contact and the last mile

11:54 Manipulating transparent objects

14:29 GenForce: transferring force-prediction knowledge

17:59 Hardware, shared datasets and standards

21:43 ViTacGen: learning from touch, deploying with vision

25:42 How tactile sensor skin wears

28:43 Replacement skin and model calibration

30:49 Manufacturing cost and consistency

32:38 TacDiff: designing the finger and policy together

36:47 Industrial research with Unilever and Ocado

40:01 ViTac workshop and open tactile data

45:11 A startup question and a surprise announcement

49:43 What warehouse robots could feel in 2031

Transcript

00:00 · Shan Luo

But the real question is, can we make sure the model can still work if we change the skin? The group has to keep collecting essentially the same kinds of physical interaction again and again. That’s why at deployment for some tasks, the knowledge can be carried by vision without requiring the tactile hardware to be present. Thank you for the invitation. I’m very, very happy to join the podcast.

00:29 · Michelle Sun

So you did your PhD at King’s, and then you went to MIT, Leeds, Harvard, Liverpool, and you’re now back at King’s holding the chair. When did touch become the thing that you want to spend the next 20 years on?

00:47 · Shan Luo

my interest when I started my PhD. when touch was a field that didn’t receive very much interest compared to other fields like computer vision or natural language processing. But to that time, when I started my PhD, we were considering enabling the surgeon to have a sense of touch for minimally invasive surgeries, and we were considering different sensors, and we didn’t find any suitable sensors in the market that can be deployed in minimally invasive surgeries. That’s how I studied my interest in the sense of touch. And back to that time, I discovered that it was very challenging to get the sense of touch in the very limited space, like minimally invasive surgeries.

01:46 · Shan Luo

And then I started working on designing tactile sensors and also these processing for tactile data. So that’s how I started my journey into the sense of touch and also tactile sensing.

02:04 · Michelle Sun

Amazing. And what were some of the biggest breakthrough in the field of tactile sensing in the recent five years, let’s say, and what are still unsolved in a field right now?

02:17 · Shan Luo

I would say in the past five years, the biggest breakthrough would be how we can scale the tactile data, especially in simulation and also in the real world data collection. We created this simulation for tactile sensors for the first time, and it was followed by many other researchers, including MIT and many others. So this enables the data collection in the simulation, and we can train our robot agents equipped with tactile sensing in the simulation so that we can transfer such robot agents to the real world.

03:01 · Shan Luo

And this has attracted a lot of interest from different fields. And we can see how this can be integrated with vision and other modalities and enable robot dexterity. And on the other hand, the development of models for processing tactile data is another breakthrough. So we have seen various models that can enable the understanding of tactile data and how to embed tactile data into the bigger ecosystem of the intelligent systems. I believe these are the two major breakthroughs for tactile sensing in the past five years.

03:40 · Michelle Sun

And what are still unsolved right now? What are the frontiers that you’re looking at that you hope that more people can work on it and it will lead to the next breakthrough?

03:53 · Shan Luo

Yeah, I take the interface as the missing piece of the whole building blocks. So we have seen different tactile sensors that we have seen these models. But we haven’t seen an interface to integrate different tactile sensors into this ecosystem. So different tactile sensors, they have different advantages and disadvantages. For example, like vision-based tactile sensing, they have high resolution for these different areas of the body parts. So vision-based tactile sensing can be challenging to be equipped to these body parts. For other tactile sensors like magnetic-based tactile sensors, they are very fast and they can be made very small.

04:46 · Shan Luo

But the resolution can be a challenge. So that’s why we need different types of tactile sensors in the ecosystem. And we need the interface so that the different tactile sensors can talk to each other. And these can be integrated into the whole intelligent ecosystem.

05:04 · Michelle Sun

For sure. One of the things that when I first learned about tactile data is that the sensors are also different and they produce different types of data. And so we’ll definitely talk more about this later in our conversation about your work to help integrating various data as well. And so let’s take a step back and talk about what touch is for when it comes to robotics and manipulation tasks. So let’s talk about the manipulation task that requires tactile sensing today. What are some tasks that touch is the requirements?

05:46 · Michelle Sun

Without a touch, it cannot happen. And then what are some tasks that touch is an improvement to the task?

05:55 · Shan Luo

Yeah. I would say for this task that require physical interactions, the touch will be a requirement, not improvement. So I will give three examples of such tasks. So one can be material assessments. For example, like hair assessment. This is a daily task for humans. So for vision, we can see the hair. But we can’t estimate how we can feel the hair. And really, when we get the contact with the hair, we can feel the hair. And we know the smoothness, stiffness, and other friction properties of the hair.

06:39 · Shan Luo

That’s the time when we use touch to evaluate the materials. So similarly, for other materials like garments or other materials in our daily life, we really need to use touch to evaluate their material properties and how they feel. So that is one example. And we have been working on this with Unilever. As Unilever, they have these products like shampoos or conditioners. And in the normal practice, they recruit people to assess hair or fabrics using their own hands so that they can evaluate how much friction when the hand is interacting with the hair.

07:25 · Shan Luo

But however, this can be biased and this can be very costly for their product development. So they have been working with us to develop these tactile sensing solutions so that we can use artificial touch to evaluate such materials. And this cannot be done by vision. And vision can only give you the appearance information. So that is one example. Another example is wiping a table. So how we clean a table using a wipe. So in that case, vision is definitely occluded by the hand. And we need to use a sense of touch to detect how much force we have applied to the table.

08:10 · Shan Luo

And if we detect any dirt on the table and how clean we have cleaned up the table. So this is another example. We need to rely on the sense of touch in doing the task, not the vision, as the vision is occluded by the hand itself. Another example is how we can insert a USB or a connector into our socket. So in that case, vision cannot feel how we have inserted the USB into the socket. So we need to use a sense of touch as the fingers to feel how much reaction force we can get from the socket.

08:53 · Shan Luo

So for the interactions between the socket and the USB. So only using the sense of touch, we can insert the USB into the socket. But that’s not with vision. So these are the examples. I can say we rely on tactile sensing instead of vision. So in that case, tactile sensing is not an improvement, but it’s a requirement.

09:18 · Michelle Sun

I love that. And I love that you’ve given examples of both from a commercial standpoint with Unilever and then from a home setting, where now a lot of people look at the home robots. And they’re hopeful to have these robots helping to do their chores. And then also on just a daily life standpoint like inserting a cable into a USB plug. And in terms of when you think about vision only manipulation, there’s been quite a lot of improvement over the years. How has that shaped your view of tactile sensing and where tactile sensing fits in the manipulation stack?

10:01 · Shan Luo

Yeah. I have seen many developments in vision-based manipulation. And this has improved the manipulation task a lot. But I would say in many of these tasks, vision can get the robots to do the task, but not complete the task. And this turns out to be clearer for robotics researchers that tactile sensing is important in completing the task. So usually when we have these objects on the table, we use vision to know where the objects are on that table. But when we pick up these objects by the robots, so when the contact is established, tactile sensing is much more important than vision.

10:55 · Shan Luo

As this tactile sensing can tell you how the contact dynamics happens between the fingers and the cup, as this may change the state of the objects. And the state of the objects cannot be perceived by vision only, as vision can be occluded by the object or by the hand. So in that case, tactile sensing is important to complete the task. And I would call it like last mile problem solver for manipulation task.

11:26 · Michelle Sun

For sure. That’s definitely a great way to frame it, right? So vision is great to start and probably do 80%, 90% of the task available. And then the last mile really to complete the task in all sorts of conditions where vision cannot even see or is occluded, then tactile really comes into play. And I think as humans, we take it for granted, right? Like the sense of touch gives us so much information. And sometimes we don’t think about it. And we think about like our eyes or our vision more.

12:01 · Michelle Sun

But actually, the touch has been informing us a lot, like feeling what’s in our pockets, finding out things, and without having to pick up the stuff in our pockets and show it to our eyes. And another thing that you worked on was about transparent objects. How do you see that? Do you think that’s one of the best proofs about touch being in a requirement? I know you shared like three other really great examples as well. Is there like a strongest argument for touch is the requirement for manipulation?

12:40 · Shan Luo

Yeah, I would like to put it in this way. Touch is a requirement for completing the task, especially in this example of transparent objects manipulation. So in that work, we have done some experiments to compare vision-only manipulation and also vision-guided touch manipulation. So in these two scenarios, if we look at the depth maps from vision of these transparent objects, so it’s quite sparse. We cannot detect these objects on the table. And it’s very challenging to use vision-only information for the manipulation. However, in our model pipeline, so we use vision to guide touch to pick up the objects.

13:34 · Shan Luo

In that case, vision can give us some sparse cues of the objects. So for example, the edges of the cup, especially the top edges of the cup, even though the body of the transparent object is very challenging to detect by vision. But it can give us very sparse information so that we can use touch to verify if the edges are there and we can estimate how the object is standing on the table. And then touch can complete the task of picking up the objects.

14:14 · Shan Luo

So with this example, I would say vision can give us some global view, can give us some initial estimates of the object state. But touch will be the one that will complete the task in the whole loop.

14:29 · Michelle Sun

That’s a good segue into GenForce. You publish GenForce, which is when you take a step back. So GenForce is the research you’ve done where you took a force model trained on one sensor and make it work on another model and on another sensor. So as we mentioned earlier, the tactile sensors are such a different challenge, different from cameras or other types of sensors, where the data that it generates are very different, sensor to sensor types. For someone outside of the field that are not familiar with tactile sensing, what does GenForce make possible?

15:16 · Shan Luo

I would put it in this way. So with GenForce, we significantly reduce effort from a sensor-specific calibration to a transfer problem. So traditionally, we have these sensor manufacturers, and we need to get these calibration models for each sensor. And even for the sensors from the same manufacturer or the same sensor, if we use it for a certain period of time, so the force prediction will be very different from the force predictions out of the factory. So it means we need a model, and we need a lot of data to calibrate the sensor so that it can meet the requirements for a certain task.

16:07 · Shan Luo

But with GenForce, we can have the force prediction model trained on one sensor to be adapted to another sensor. So for example, if we have a model for force prediction trained for GelSight, the vision-based tactile sensor, so this model can be adapted for another type of sensor. It can be a vision-based tactile sensor like TacTip, or a GelSight sensor with a different elastomer, or a tactile sensor of a totally different sensing transduction sensors, for example, like a uSkin from XELA with magnetic-based tactile sensors.

16:50 · Shan Luo

So in that case, we don’t need to collect a large amount of data to return the force prediction models for the new sensor. Instead, we just use the prior knowledge trained for another sensor, and then we can adapt it for another sensor. So in that case, we can save a lot of efforts for the force prediction for different sensors. And also, in the same time, as we use the sensor, we can dynamically adapt our force prediction models for the same sensor so that it can keep the accurate predictions of the force for the same sensor.

17:29 · Michelle Sun

This kind of reminds me of vision, where vision data compound, because different data from different cameras generate the same type of data. Different cameras generate the same type of data, and that can be used by the same vision model. With GenForce, you’re kind of unlocking that capability, where tactile data set become the assets, and the sensor becomes the inputs. In this scenario, with GenForce becoming more powerful, does the sensor hardware commoditize? And if it does, who ends up holding the cross-sensor data set?

18:11 · Michelle Sun

Is that a company or a consortium?

18:13 · Shan Luo

So there are two questions here. So first, does the sensor hardware become commoditized? So to some extent, yes. But I don’t think the hardware becomes unimportant. So what GenForce start to show is that the data and knowledge learned from that data don’t necessarily have to remain locked to one particular sensor. So if I collect a very valuable tactile data set using one sensor, and I can transfer what I have learned to another sensor, so then the values are to shift from the individual piece of hardware towards the data models and also representations build on top of it.

19:01 · Shan Luo

But hardware still matters a lot. So different sensors have different spatial resolution, sensitivity, bandwidth, durability, and also form factors. As I just mentioned, so for different body parts, so we may need different types of tactile sensors. So that’s why the tactile sensor hardware is still very important. So you can’t recover information that the hardware never measured in the first place. So I see this less as hardware disappearing and more as hardware becoming part of a larger interoperable tactile ecosystem. For the second question, so who should host the cross-sensor data set?

19:47 · Shan Luo

My preference would be for something closer to a consortium or standardization organization rather than a single company owning everything. So for tactile sensing to reach something like the scale that computer vision achieves, we need many groups to contribute data from different sensors, robots, objects, and manipulation tasks. But for those data sets to become generally useful together, we also need standards. So apart from the sensor specification, resolution, durability, and many other metrics, we also need these standards to standardize the data collection process to benchmark different sensors.

20:33 · Shan Luo

For example, what object stimuli we should use for benchmarking such sensors, like how much force, which trajectories we need to use to contact the tactile sensors. So these are the protocols we use for data collection. So only with that, we can standardize our data collection, our datasets, and we can benchmark these models for tactile sensing. So in essence, I believe the best practice would be to have the consortium or standardization organization to own these datasets or these protocols.

21:14 · Michelle Sun

For sure. And you touch on a really important point where different parts of the robot body would require different types of tactile sensing. And so that’s why where the tactile sensing hardware actually plays a really important part to contribute to this cross-sensor data sets. And you worked on GenForce at your lab, which makes tactile data portable across hardware. And you also worked on ViTacGen, which is generating tactile signal from a camera, and it is up to 86% success rate using vision alone.

21:57 · Michelle Sun

So on one hand, GenForce is saying that data is the assets and it’s portable across different hardware. And for ViTacGen is saying that the signal is able to transfer without the hardware, the tactile side of the hardware. Which of the two changes more for a company that sells tactile sensors today?

22:21 · Shan Luo

Yeah, I actually think both are right, and they are complementary rather than conflicting with each other. So GenForce asks a question, if I need physical touch, can the knowledge I learn from one tactile sensor transfer to another? And the answer is yes. So that makes tactile hardware easier to deploy because new tactile sensors don’t require us to start again from zero. But on the other hand, ViTacGen asks a different question. So once I have learned from this physical touch hardware, do I always need the tactile sensor?

23:03 · Shan Luo

at deployment time. And in some tasks, the answer may be no. And in the ViTacGen work, the task we were addressing is object pushing. So in that task, we just use an end effector of the robot to push one object from position A to position B. So in that case, the real tactile information we need is the geometry information of the objects. So in that sense, we can transfer the geometrical information learned from tactile sensing to vision from its views. And so in that sense, we don’t need the tactile sensing in the deployment once we had trained the models using tactile sensing and vision together.

23:50 · Shan Luo

So with ViTacGen, we use tactile data during the learning process to teach the system about physical interaction within the relationship between what the camera sees and what the tactile sensor would have perceived. So that’s why at deployment for some tasks, the knowledge can be carried by vision without requiring the tactile hardware to be present. But I don’t interpret ViTacGen as saying that the tactile sensors are not needed. So in fact, we need tactile sensing to create the physical experience that the model learns from in the first place.

24:30 · Shan Luo

So we can learn from humans as well. So we physically touch objects and materials and build associations between how things look and how they feel. So later, we can look at something and predict that it will be soft, rough, or slippery without touching it. But that prediction comes from experience with the physical world. So for tactile sensor company, I think these results point towards a broader business than simply selling a piece of hardware. So the sensor generates the physical experience. So the data captures that experience, and the model makes that experience transferable.

25:15 · Shan Luo

And then, so I think about the whole stack, so we need the hardware to gather the initial data. But also, the tactile data can tell us the experience, how we can interpret the physical world, and also the models that can turn that data into reusable physical intelligence. So that’s my standpoint for these two questions.

25:42 · Michelle Sun

Yeah, that’s really cool. And I see that the tactile sensor data actually is also the input for ViTacGen, in addition to human interaction data, to make that prediction better. So it’s actually important in the whole stack and important in improving the whole loop. And your lab currently has all these different sensors, so GelSight, TacTip, uSkin hardware, et cetera, side by side. What do you replace the most often? And how often do you have to replace them?

26:19 · Shan Luo

That’s actually quite a different question to answer. So it really depends on what experiments we have. So we have various tactile sensors, like the GelSight Mini, we have TacTip, and we have uSkin. Most of them were donated by this company to our research lab. And we also make our own sensors to fit our task requirements. So from our experience, we may replace tactile sensor after a certain period of time, as the extensive use of the tactile sensor will make the elastomer of the tactile sensor [unclear].

27:18 · Shan Luo

Especially for various tactile sensors, we have this painted area, we have the painted layer. So this may get thinner as we have extensive use. So usually, we may say after a few months of use, so we may replace these elastomers. But for some other sensors like uSkin, as we have a rubber material, it can be more robust. And we may replace such material for a longer time. But this is similar to our human skin. So we get new skin after a certain time. So we are not having the same skin as the one we were born.

28:04 · Shan Luo

So that’s why it’s similar for artificial tactile sensors, as this skin is the part that gets most physical interactions with the environment. So that’s why we may need to replace such elastomers for a certain time. And also, it depends on what task we have. So if we have gentle contact with objects, we may use it for a longer time, like one year. But if we have extensive shear force, so this can make this elastomer easier to get damaged. So we may replace it in a shorter time.

28:43 · Michelle Sun

Yeah, so around anywhere between a couple of months to a year. And you touch upon the elastomer, so I wanted to follow up on that. That’s an interesting parallel that we’ve seen with the human skin. Our skin gets shedded throughout the year. And essentially, it’s always changing. It’s almost like a replaceable layer for our body to shield against the environments. Or let’s say this robot starts to also have consumable skin. Do you think the replacement cycle would be sometime around a year, a couple of months?

29:23 · Michelle Sun

And how much would it cost per robot per year to have a skin like that?

29:32 · Shan Luo

I would say it really depends on what tactile sensors you may have. As the skin we use for vision-based tactile sensor could be very different from the ones we have for magnetic-based tactile sensors. But normally, we take such skins as consumables. In practice, this may cost a few dollars for the skin, a small patch for the finger pads. So in that sense, if we replace the skin after like three to six months, so the cost is not high. But the real question is, can we make sure the model can still work if we change the skin?

30:20 · Shan Luo

So that’s a real tricky question for now. If we change the elastomer, the material property may be slightly different. And also, the alignment of the elastomer with the sensor can be slightly different. How can we make sure the models we have developed for the previous skin can be reused for the new skin? And that’s how we can use GenForce to make it happen.

30:50 · Michelle Sun

And your outlook paper also said something about how manual the process right now is when it comes to making these sensors. So optical sensors are made by mixing, painting, coating, all done by hand. And e-skin needs many manual steps in the clean room. So for touch to... and you mentioned that from different skins, sometimes the model doesn’t work, and maybe in simulation, it needs to be tweaked as well. So for touch to ship on a production robot, what matters more? Is that taking the price down further or getting every unit to read the same so that the model works across different hardware?

31:39 · Shan Luo

I want to say both. So on one hand, if we push the tactile sensors into the market, so we need to get the price down so that it can scale up. It can be used by the robots. That it will be deployed in factories, in homes. So that’s why we need to minimize the cost of tactile sensors so that we can minimize the total cost of the whole robot. So that is one hand, and definitely it’s the trend to minimize the cost. On the other hand, so we really need to make sure how we can scale up a different tactile sensor.

32:16 · Shan Luo

As we just mentioned, we may need different types of tactile sensors for different parts of the hands or different body parts. So in that case, how we can have different tactile sensors talk to each other, how we can have this interface between different tactile sensors, that’s OK. How we can get tactile sensing into the bigger intelligence system.

32:38 · Michelle Sun

For sure. So let’s shift gears and talk about TacDiff a little bit. Your view is that robot parts get built first, and algorithm shows up afterwards. And with TacDiff, you’ve been spending five years designing the sensor shape and the control policy together. Where is the biggest gain when it comes to getting these capabilities up? So is it a better sensor or designing the better policy and the finger at the same time?

33:12 · Shan Luo

Yeah, I think the bigger gain comes from designing them together, because there isn’t really a universally best tactile sensor or robot finger. And this is a better design for a particular task. And this is based on our experience in designing the tactile intelligence systems. So for example, using a parallel gripper, using a cube-like tactile sensors. So we can pick up objects like the phones or cups. But if we use such a parallel gripper to pick up a piece of fabric, a piece of paper, it’s very challenging. So the main reason is that the form factor cannot enable the picking up.

34:00 · Shan Luo

If it is a cube-like form factor, it’s very challenging to adapt such a sensor to pick up a soft object of very thin objects. So that’s how we design these RoTipBot, with a finger like these cylinder shapes, like our human finger, so that we can roll over the objects. And we can use the shear force to pick up the object, like a piece of paper or a piece of fabric. That is not possible by a cube-like tactile sensor for a parallel gripper. So this gives us the idea how we can design form factor and also other parameters, like control policies and also these material properties of the elastomer altogether to optimize its designs for a given task.

34:55 · Shan Luo

So all of these will be dependent on the task given by the user. So that’s why we propose TacDiff. We want to use simulation to optimize these different parameters, so that we can optimize these parameters for a given task. So with TacDiff, we want to reverse the traditional thinking. So we go to one design, so traditionally, we don’t consider what task we want to have. We don’t want to guess the best design. It results in the considerations of specific tasks. So here, we start with the task and ask, what morphology, what sensing and what control policy should the robot have to perform this task well?

35:46 · Shan Luo

So take manipulating a very delicate object as an example. So you may want a soft fingertip with high force sensitivity so that the robot can detect very small contact forces. But if the task is inserting a component during assembly, the optimal finger might be stiffer with sensing focused in particular areas where contact tell you about alignment. So this example can give us more motivations to use these simulations that can optimize these design parameters. So ultimately, the question changes from what is the best tactile sensor to what is the best tactile robot for this task.

36:34 · Shan Luo

So here, we use TacDiff, this differentiable simulation so that we can optimize these parameters based on the given task. And then we can get better performance for that task.

36:47 · Michelle Sun

For sure. And in the beginning of the podcast, you mentioned about Unilever, the work that you’ve been doing on the shampoo and testing the texture of the hair. And you also work on the deformable and packaged goods as well. With an industrial partner like this, what have you learned from working with a partner like Unilever? And what are some things that required of your work that your research have not been producing? And you have to work with them to solve some of the problems.

37:23 · Shan Luo

Yeah. So there are some problems that we would like to solve together with our industrial partners. Taking Unilever as an example, so in their practice, they recruit people to assess hair so that they can have subjective assessment of hair, and then they can evaluate their hair properties. Also, they have some objective measurements practice. They use this tribology technology to collect the friction data, and then they have this simple linear regression model to predict these hair properties. But the information is very limited, collected from the friction data.

38:15 · Shan Luo

So compared to their practice, using these tactile sensors we have developed, these vision-based tactile sensors, we can capture much richer information. For example, the fiber distributions, the texture of the hair, that cannot be disclosed by the friction data collected from the load cells. So this can give richer information that can be comparable to subjective assessments. So that’s what we have solved, and we can map this richer tactile data into the space that can be aligned with subjective assessments. That was not possible before our solutions. Another example is with Ocado technology.

39:09 · Shan Luo

So Ocado is the largest online grocery in the UK, and they have been operating its automated warehouse. But in that case, they still rely on human resources to handle these delicate objects, like fruits, these delicate objects. And they cannot detect these forces between the interactions of the objects with these robots. So with tactile sensing, we can detect these forces with delicate objects, and we can moderate the forces applied to these objects. So these are the problems that we are solving in the industry, and they are not possible without tactile sensing.

40:01 · Michelle Sun

For sure. And you’ve founded the ViTac workshop, and this year is the seventh year at ICRA. The theme was learning to see and feel. Every year, you see where the whole field of tactile sensing has been developing, and what are they submitting to the workshop. What has stood out to you for this year, and what are you seeing that are emerging as trends that you’re excited about?

40:32 · Shan Luo

Yeah. The biggest change I have seen over the past two years is transition from hardware-focused research to intelligence-focused research. So when I started the ViTac workshop series, so many of the papers we received were on the hardware, how we designed the tactile sensors of different sensing principles, different form factors that we can equip such tactile sensors to different types of robots. But recently, we have seen many papers working on foundational models, simulations for tactile sensing, how to scale up data collection for tactile sensing. So we are seeing more intelligence elements in the submitted papers.

41:24 · Shan Luo

Definitely this is a trend in the community. So from what I can say, it’s also a reflection of what we have developed over the years. So we can see intelligence can be built on top of its hardware sensors. So we have seen a convergence of the sensor developments. So we have got good sensors for these robots. But now we want to answer the question, how we can use such sensors in robotic systems? How we can make it benefit these tasks? So the questions have changed from, can the robot feel?

42:06 · Shan Luo

To what can the robot learn from feeling? So that’s the transition I have seen in the past years.

42:16 · Michelle Sun

Yeah, so from hardware-driven design to intelligent focus research. And that actually ties everything together really well, with the more and more tactile data being collected in the real world in deployment, not just in the labs. And how do we make use of it and to build better model for the robots? And you hold a Turing open source AI fellowship. In tactile, which layer benefits most from being open? Is that the data sets, the trained model, or the sensor designs?

42:56 · Shan Luo

Ideally, all three, so the data sets, the models, and the sensor designs. So they benefit from being open. So the community can benefit from open sourcing these designs. So this can accelerate the designs of tactile robots. But if I had to prioritize, I would put data sets and models first. So the reason is that their value compounds, so tactile data, are still relatively scarce compared to with vision or language. So every research group tends to have their own sensor. So its own experimental setup and its own data sets.

43:42 · Shan Luo

So if those data sets remain siloed, every group has to keep collecting essentially the same kinds of physical interaction again and again. So that’s why I would say the tactile data collection is the first priority I would like to highlight to be open-sourced so that we keep the same or agreed protocols for data collection. And we can share such data sets across research groups. So then we can interoperate these models across these data sets. So the models can then be deployed across different data sets. And so these models can be benchmarked.

44:33 · Shan Luo

So we know how the model can progress. And this will be similar to what we have already got for ImageNet. So we have this data set that is benchmarked, well documented. And also the models can be benchmarked. So similarly for tactile sensing, we need the data set first. And then we need to benchmark such models. And we can reproduce these results across different sensors and across different tasks. So that’s my view for open-source studies.

45:11 · Michelle Sun

Yeah, it’s interesting that it actually is a sequential thing, right? So the data set actually unlocks better models. More data unlocks better models. And then it informs better designs for the sensors. So I know that, Shan, you’ve been in academia. And I’m sure a lot of people ask you this. But if you were starting a company today instead of running a lab, would you build a sensor, a data set, or a simulator, or a different answer?

45:43 · Shan Luo

Actually, I have already started the company. It’s called Hand Intelligence. So we are building this venture. And to answer your question, I wouldn’t choose just one of these three. So what we are building is really a full-stack physical AI company. Because I think the sensor, the data, and the models are deeply connected to each other. So you need a sensor to interact with the physical world and collect the high-quality tactile information that generates the data, which I think will become an increasingly important asset, as we discussed earlier.

46:24 · Shan Luo

And then you need the models and simulation to learn from the data, scale it, and transfer that physical intelligence across different sensors, models, and tasks. But if you only have the sensor or the data sets, you will be limited to being a hardware supplier, or you don’t have a way to generalize to different tasks. So that’s why we also developed these AI models, including these foundation models and world models that is embedded with vision and tactile sensing, so that we can complete these physical tasks.

47:06 · Shan Luo

So I think the long-term value is not simply in giving a robot a better fingertip, it’s in giving robots a better understanding of physical interaction. So as a very first step, what we are developing now is funded by ARIA. So ARIA is a funding body in the UK, and they have a Robot Dexterity programme through this program. So Hand Intelligence is developing this interface so that different tactile sensors can talk to each other. And we are generating these unified representations that can have different tactile sensor manufacturers can adopt these unified representations, and we can get these benchmarks or representations for different tactile sensors.

47:55 · Shan Luo

And they can be used for downstream tasks of robots. So if you ask me, sensor, dataset or simulator, my answer will be the interface between all three is where I would build the company. And that’s actually what we are doing at Hand Intelligence.

48:16 · Michelle Sun

How do you spell it? Like just for our notes, like Hand Intelligence, is it like hand?

48:22 · Shan Luo

Yeah, Hand Intelligence.

48:25 · Michelle Sun

I see, I see. Is that live right now? Like just for our editing, do you want this to be mentioned in the podcast or like when are you guys going live?

48:37 · Shan Luo

Yeah, we are still at a very early stage of the venture. And so we want this company to be mentioned as this can be heard by like investors or any interested companies. So this can be mentioned to the audience, the podcast. Also, at the same time, we are still as the stealth mode. So we are not disclosing very much detail of the company. But we want people to know we are doing this already.

49:16 · Michelle Sun

Awesome, awesome. And if you have a link or any website that I can point to, then we can include that into the show notes and the editing as well.

49:28 · Shan Luo

Yeah, at this moment, we don’t have a dedicated website yet. Yeah, so if any of the audience is interested, so they can reach me. And I’m happy to talk more about our company.

49:43 · Michelle Sun

OK, sounds good. Let’s wrap up the last question. And then I’ll stop the recording. And I’d love to learn more as well. So this is exciting. And I’m sure with the company, you guys thought about it a lot when it comes to the next five years of where the industry is going. So take me to 2031, a robot works eight hour shift in a warehouse. How much of that shift involves hand intelligence, touching something, and what is on its fingers?

50:14 · Shan Luo

Yeah, in five years, I would imagine that these robots have been deployed in warehouses. And each of the robots has tactile sensors equipped to their hands. And they can detect the contacts, the force interacting with the objects. And they can decide how much force can be applied to these objects, and it can better complete this task. So we have seen some demos. They are just looking at these boxes. And they are not really feeling the contact with these objects for now. But in the five years, I would say when these tactile sensors are in their hands, they can really tell which objects they are touching, what material they are touching, and how much force they need to apply to that object.

51:08 · Shan Luo

And this will close the loop and help them to complete the task.

51:14 · Michelle Sun

Amazing. Well, thank you so much, Shan, for your time. It’s great having you today.

51:18 · Shan Luo

Thank you, Michelle, for the invitation. And it’s great to share what we have got so far with the audience.

From pre-conference coffee on the CMU campus to lightning talks, lunchtime plenaries, walking meetings around the exhibition floor and hosting a final panel at 9 p.m., the days were full. Over two days, I had more than 50 conversations across 45 companies and labs. Below are some observations:

1. Humanoids are losing their legs

When I visited Sharpa’s Mountain View office in July, the team proudly showed me its North humanoid at the poker table. At IROS, it was mounted on a stationary base. No wheels, either.

The representative on the floor told me the change was intended to allow the robot to ship into the US following the FCC’s July restrictions on foreign-produced advanced robotic devices. The definition covers qualifying mobile robots, including humanoids and quadrupeds, and excludes fixed, stationary industrial robots.

‘So customers can add back their wheels if they want’, an exhibitor next to me commented. I thought it was a funny, very tangible example of how a regulatory change can show up in product design.

2. Selling hands, collecting data

Hands and grippers were everywhere, with quoted prices ranging from about $700 for a gripper to $50,000 for a highly dexterous hand. Many had new models that looked leaner and more human. From the demos alone, it wasn’t easy to tell how they differentiated themselves.

Wuji told me it ships 800-900 hands a month globally, with US prices of roughly $13,000–$19,000 including tax and customs. Sharpa quoted $50,000 for its Wave 01 hand (the W02 price is not yet disclosed) and named Google and Nvidia among its research customers.

I expect the market to become more clearly segmented by buyer: humanoid makers buying production grade, economical hands in volume, research labs paying for advanced capabilities, and factories buying grippers that last.

Several hand companies were also showing tools for collecting data. Sharpa has switched out Manus gloves for its own exoskeleton, quoted at $10,000 a pair. Wuji, Chestnut and PaXini also showed UMI gloves designed for data collection.

That makes sense to me. A lab might buy a handful of robot hands, then need a much larger stream of demonstrations to train them. Selling the collection hardware is one opportunity. Providing the data itself could create recurring revenue, if the hand makers can show that it improves training. PaXini, for example, operates its own data collection facilities.

3. Actuator know-how lives on the factory floor

There were plenty of actuator companies on the floor, too. MyActuator told me it serves US customers from Fremont and has two factories in China producing 500K units annually. It focuses on custom engineering and aims to bring down turnaround to weeks.

And from my conversation at maxon, I learned that many of its patents are in process engineering. The know-how inside a plant also lives with the people operating the machines. It’s hard to transfer, even when the same company opens a new factory in a different country.

I increasingly believe that building a US actuator supply chain requires much more than opening factories. The challenge is scaling production while consistently making precise, reliable products that last. Some of the hardest knowledge to acquire comes from years of working with the machines and refining the process.

4. Sensors: can you source them, and will they last?

A professor doing medical robotics research commented that none of the tactile sensors in the market today lasts, and the biggest bottleneck is in materials science, not models. A few hours later, a PaXini sales representative told me its Hall-effect sensors last for 10 million cycles and described testing the sensors after boiling them in a hotpot and dropping them in ice water. An engineer at Daimon said BMW and Siemens use its sensors in deployments. Sharpa demo’d its e-skin, made in-house, that detected pressure on parts of the robot’s body, including its shoulders and elbows.

My guess is that the professor and these suppliers are right, but in different scenarios. How do we measure the durability: under what loads, doing which tasks, and with how much maintenance?

An entrepreneur who helps robotics companies deploy their systems told me that one of his biggest frustrations is finding competitive domestic lidar sensors. He described a US lidar option priced at $60,000 versus a $500 imported option that performed better for his application. I’d like to dig into the exact models and specifications, but that price gap was shocking to me.

5. Getting a robot to work is a big business

Commercial installations still require integration, programming and ongoing support. The Standard Bots team described deployment engineers traveling to customer sites in various states, often starting with a single arm. Applications included programmed welding and CNC machine tending, with newer AI applications also in development.

Forward-deployed robotics engineer (FDRE?) feels like an increasingly common job description.

Independent evaluation is another need I’ve encountered while writing about robot foundation models. How do you compare capabilities objectively beyond the demo or press release? Robocurve works with universities on benchmarks for different use cases and plans to offer independent evaluations to companies and governments.

I like the picks and shovels plays that enable deployment. Integration, calibration and evaluation are all common pain points across robotics companies. If someone can standardize, automate and programmatically solve these, there can be a durable business.

Share

A couple of other conversations stayed with me. The Robotis team told me its actuators are used in Disney’s Spider-Man and Olaf robots. Entertainment may be one of the fastest-traction deployment categories; robot fights are another example beyond theme park robots. Professor Yu Sun at the University of Toronto gave an inspiring talk on robotics at micro-nanometer scales, on IVF and fighting brain cancer (Learn more here: https://sun.mie.utoronto.ca/).

I’m writing this piece on the 7AM flight out of Pittsburgh, sleepy and smiling. IROS brought the endurance athlete back out of me, and gave me plenty of reasons to keep showing up.

The FCC’s restrictions on foreign-produced advanced robotic devices put sourcing on the agenda for companies making covered robots, with conditional approvals available. [1] If US robotics companies need more domestic actuator supply, what must be built or expanded, and which suppliers can turn that demand into a durable business?

The filings from Chinese manufacturers, alongside established suppliers’ production experience, give us a way to examine that question. They show customer concessions, changing product offerings and the cost of expansion. What I want to understand is which of those lessons matter for the US businesses being built now.

If you work in actuators or related fields, I’d love to chat. For paid readers, I’ll be sharing a PDF version of the report that includes Parts 1-4 and additional data and analysis next week.

In this piece:

1. What America actually needs to build

2. What makes actuator manufacturing defensible

3. Can an actuator manufacturer deliver venture returns?

4. Does selling the whole joint improve the business?

5. What remains imported after the factory opens?

6. What else needs to be built around actuator factories?

1. What America actually needs to build

For a robotics founder in the US today, what is the hardest component to buy: the motor, the reducer or the complete actuator? The motor contains the stator and rotor. The reducer is the gearbox, which can be planetary, harmonic or cycloidal. A complete actuator, sometimes called a joint module, combines more of the motor, transmission, sensing and control functions, depending on the design.

Domestic production already exists. Harmonic Drive LLC says most products it sells are made in Massachusetts. [2] The question is whether available supply meets a customer’s requirements for price, performance and volume.

Suppliers can offer off-the-shelf products that cover common needs or work on customer-specific designs. For example, Atlas Motion Systems, an actuator startup in the US, says it designs systems around customer requirements and allocates manufacturing resources to its partners as customers scale. Its website describes prototyping in Southern California and volume production in the Philippines. [3] The same supplier can do both.

For a US robotics founder, the question is which part of the joint they can already buy domestically at the price, performance and volume they need, and which part still needs another supplier.

2. What makes actuator manufacturing defensible

Making an actuator involves manufacturing processes, cost optimization and innovation that improves performance or lowers cost. Just because someone can tear down an actuator and see the design doesn’t mean they can reproduce the process behind making it. Process innovation can improve cost, performance or delivery while producing consistent quality.

On August 13, Schaeffler said it had validated its formed strain-wave gearbox for humanoids, claiming manufacturing-cost reductions of more than 25% and material-use reductions of more than 75%. The company reports supplying more than two million formed automotive gearboxes over ten years. That is accumulated process experience. It plans humanoid mass production for 2027. [4]

For stator winding, conductor geometry interacts with bending, twisting, welding and insulation limits. Research on automotive hairpin windings shows how closely the design and manufacturing process interact. It helps explain the work involved in reproducing a winding process, while the economics of a smaller robot motor still need to be established. [5]

The filings make this concrete. Harmonic-reducer maker Laifual measures capacity from flexspline-line throughput. Planetary-reducer maker Newstart measures capacity at ring-gear machining. Precision-transmission supplier Seenpin describes long delivery and commissioning cycles for expensive equipment. [6][7][8]

A new supplier could develop these capabilities in-house, or buy qualified output from specialists. The question is then: which processes does it need to control to meet customers’ cost and performance requirements? Bringing a process in-house adds equipment and ramp-up costs. That investment needs to earn its keep through better economics, dependable capacity or performance that customers value. [14]

When evaluating how defensible an actuator supplier’s business is, I would ask:

• What would take a capable competitor the longest to reproduce?

• What yield, cycle time, tool life and sustained performance has the supplier demonstrated?

3. Can an actuator manufacturer deliver venture returns?

Growing humanoid and advanced-robotics production could expand demand for domestic actuators. But the volume a supplier wins depends on its customers, market share and how much robot makers produce in-house. The filings also show that average selling prices can fall as businesses grow. [7][9]

US$ equivalents below use a fixed Rmb7.19 per US$1 for scale; percentage changes use the original Rmb figures. [15]

An actuator supplier can grow total gross profit even while prices fall. Look at Newstart, a planetary reducer maker. Its 50–240mm planetary reducers illustrate this: from 2024 to 2025, average selling price fell 6%, from Rmb844 (about US$117) to Rmb791 (about US$110), while units sold rose about 70%. The company attributes part of its broader margin pressure to price concessions for leading embodied-robot customers. [7]

For this product category, total gross profit rose from Rmb119.1 mn (about US$16.6 mn) to Rmb166.4 mn (about US$23.1 mn). Newstart earned less gross profit per average unit but sold enough additional units to increase the total. This was a recovery from 2024; category gross profit remained below its 2023 level. [7]

Margin erosion is not observed across the board. Look at another example here: Direct Drive Tech, or DDT, a Chinese manufacturer of direct-drive actuator modules and robots. From 2023 to 2025, its direct-drive module average selling price fell from Rmb77 (about US$10.7) to Rmb31 (about US$4.3) while the category’s gross margin rose from 8.3% to 20.0%. Management attributes margin improvements to design changes, better yield, purchasing and spreading factory overhead across more output. Consumer-robot modules accounted for 88% of company revenue in 2025. The changing application mix also affects this category average. [9]

Subscribe now

Newstart shows how additional volume can offset lower gross profit per unit. DDT showed improving margins as their consumer-product business scaled. These cases show how profits can continue to grow while prices fall. Applying DDT’s consumer-robot economics to humanoid joints would require separate analysis.

For a US actuator startup, growing demand creates room to build a larger business even as prices fall. But the company still has to win that volume and earn enough from it to cover R&D, sales and the cost of running the business. It also needs to have a sound capital plan to fund equipment, facilities and inventory before those orders turn into cash.

For an investor, I’d consider how much cash the next production line can generate relative to what it costs to build and ramp. If every increase in sales requires another large equipment purchase and equity round, growth can absorb much of the value it creates. Reusing designs and equipment, filling capacity and collecting customer payments sooner can reduce the amount of new equity needed to expand. [14]

Debt can reduce the need for new equity, but adds interest and repayment obligations. Hadrian provides an example from US advanced manufacturing: it announced a US$360 mn revolving credit facility in August to support infrastructure, machinery and related hardware. [10] An actuator startup would still need to establish how much it can borrow, on what terms, and whether its cash generation can support the repayments. The financing plan is part of the venture-return question. [14]

4. Does selling the whole joint improve the business?

How do actuator suppliers differentiate from one another, other than price?

What I see from studying suppliers such as Laifual and Zhongda is that their offerings are extending from individual transmission components into integrated products. [6][11]

Laifual, a Chinese manufacturer of harmonic reducers and joint modules, reported joint-module average selling prices rising from Rmb2,010 (about US$280) in 2023 to Rmb7,074 (about US$984) in 2025, about 3.5X the earlier level. The company attributes the increase to added content: low-voltage modules introduced in 2024 included drivers and braking systems, while some 2025 modules added customer-specified housings and water-resistant cables. The invoice grew along with what was supplied. [6]

Zhongda, a Chinese maker of geared motors, precision reducers and integrated actuator units, reported 2025 gross margins of 21.44% for precision reducers and 25.08% for intelligent actuator units. Its integrated products combine reducers, motors and drives. The integrated category earned a higher gross margin. Differences in products and customers leave the contribution of integration unresolved. [11]

Selling more of the joint can increase the average selling price. Whether the supplier keeps enough of that revenue is another question. Does the customer save engineering and assembly work by buying the full module? Can the supplier reuse its design, tooling and testing across customers? If each order needs extensive new engineering, a higher selling price may still leave a difficult business to scale.

For a complete-joint supplier, I would test five things:

• Engineering: How much integration and assembly work does the customer save, and how much of the supplier’s design can be reused?

• Performance: Does the module meet the required duty cycle, size, mass and thermal limits, and does that performance carry across customer programs?

• Manufacturing: Can the supplier deliver qualified modules at a consistent yield and cost as volume grows?

• Service: Does the customer get dependable replacements and repairs, and can the supplier provide that support profitably?

• Sourcing: Can the supplier document origin and maintain qualified supply as components or vendors change?

5. What remains imported after the factory opens?

An actuator involves many components, and the components need raw materials, machines and labor to produce. Local assembly can leave upstream dependencies elsewhere.

Some examples of these upstream dependencies:

• China accounted for 94% of global sintered permanent-magnet production in 2024, according to the IEA. [12] For a motor supplier, the questions include where magnets and other inputs come from, and which production and testing steps it controls.

• For reducers, the flexspline and ring-gear processes mentioned above are specific capabilities to investigate. [6][7]

• For linear actuators using precision screws, Seenpin’s equipment and commissioning disclosures raise a specific question: can the supplier order, install and qualify additional precision grinders and inspection equipment in time for the promised ramp, and how will it fund that investment before the new capacity earns revenue? [8]

Domestic players are moving to solve these dependencies. For example, USA Rare Earth is partnering with Pasqal and Riven Systems to develop more efficient rare-earth separation chemistry. The September announcement describes a research program upstream of finished-magnet production. Turning that research into commercial supply remains a further step. [13]

Labor is the other input to plan for. Actuator companies told us that people with the skills to operate their production machinery are hard to find, so they hire recent graduates and train them in-house. [16] For a supplier promising a fast ramp, the time it takes to train operators belongs in the schedule alongside equipment lead times.

So, a US robotics founder needs to ask their supplier:

• Which components and production steps depend on imports, and where do they come from?

• What would be most constrained if production rose 10X or 100X in the next 12 months?

• What equipment, supplier qualification or financing would be needed to remove that constraint?

• How many trained operators would that ramp need, and how long does it take to train one?

6. What else needs to be built around actuator factories?

As domestic actuator supply ramps up, I see opportunities beyond finished actuators. The dependencies above suggest five areas to investigate:

Read more

I hosted the Bill of Materials huddle, essentially a panel where the whole room gets to participate. It was right up my alley, and harder to moderate than I expected. I had prepared plenty of questions. We ran out of time because people had so much to share and ask for help with.

Below are some of my notes from the half day conference, hosted at the beautiful Golden Gate Club at the Presidio.

1. Who babysits the robot?

Dyna’s Lindon Gao framed deployment around throughput, quality and autonomy. What stuck with me is their experience deploying robots for napkin folding. What trips up the customer was that someone still has to move the inputs and outputs, ie feeding the fresh pile of napkins and moving the napkin pile out of the robot’s way.

In his framing, the industry is chasing the 99%, but actually 99% is just the starting line. Customers then need the right quality, speed and level of human intervention. Getting robots to work productively means reducing as much as possible the human work required for the robot to succeed at its assigned task.

2. The BOM conversation kept returning to manufacturing

It turns out lots of insights surface when you put component suppliers, PhDs from CMU, ex-Tesla engineers, robotics founders and deep tech investors in a room.

Gear manufacturing drew a lot of discussion. Participants talked about the difficulty of combining high torque with low inertia, then got into grinding, hobbing and heat treatment. We also talked about sourcing the equipment behind those processes, including furnaces, coating and surface treatment.

Buying the machines leaves a hiring problem. Participants described difficulty finding experienced operators and having to train new graduates from scratch. That really stood out to me as a lot of conversations have been around component manufacturing, but reshoring manufacturing also needs a plan for developing the people who will run it. If you know anyone working on this area, I’d love to chat.

The advice on specifications was practical. A useful tip is, instead of bringing a dream (often unrealistic) list of specs to an actuator supplier, start with the problem they’re trying to solve and use an off-the-shelf component before building something custom. That echoes with my conversation with Eric Pan at Seeed studio. When starting a buyer-supplier conversation, explain what the machine needs to do, then work backward to the component.

Supplier responsiveness was part of the sourcing discussion. One contrast that stood out to me is quick exchanges with Chinese suppliers over WeChat, including on a Saturday, vs. the difficulty getting responses from US suppliers (“90% don’t reply”). The China anecdote is true from my personal experience of sourcing parts from China. In general people are willing to respond to work messages over the weekends. Responsiveness matters to a small team trying to keep development moving. Policy changes also came up as a source of uncertainty over where to invest.

Wire harnesses and power deserve more attention. Both came up when we moved beyond actuators. The audience had specific requests for wire-harness recommendations and linear-actuator leads. If you have a supplier to recommend, I’d love to hear which applications and requirements they can support.

Side note, I counted at least 4 domestic actuator companies in the room, including Conifer and Bimotal. That is an encouraging sign for US robotics teams looking for domestic options.

3. A good pilot needs a path to a purchase order

Rhoda founder Jagdeep Singh’s advice was to be selective about proofs of concept and establish a route to commercial deployment. He said there’s no shortage of demand when it comes to PoCs. Corporate curiosity about physical AI can fill a startup’s calendar. He recommended robotics startups be mindful of who they partner with, ask early who can approve the purchase, what success means and who owns the rollout.

4. What sets a robotics startup apart

The deployment huddle discussed a crucial questions. As models become easier to access and distill, where does a robotics company build durable differentiation? Product experience, distribution, cost efficiency and speed were all part of the discussion. The most difficult part in building a sustaining robotics company may not be the hardware, but the interaction between the machine and human workflow.

Are actuators the bottleneck of humanoids or robotics? Which part exactly limits supply? Who are the main suppliers for each component that goes into the actuator? In today’s Part 3, we cover the global supplier map of actuators, for each component and where the facilities sit. We also look at where actuator supply could become constrained.

I spent the last month reading through dozens of corporate filings from the Chinese robotics supply clusters, and they sharpened the question for me. Scaling actuator supply means securing qualified parts for a particular robot design, at the required volume.

The supplier map shows where to investigate: precision manufacturing, motor performance, and materials and bought-in components. Those exposures move differently. Reading Laifual’s filings made me separate two questions: what a reducer costs, and how much qualified output a supplier can deliver. Today, we map the global supplier reality layer by layer.

Subscribe to get the next ones in your inbox. If you work in this space, please reach out. I would love to compare notes.

This piece covers six sections.

1. Bill of Materials: The Processes and Players Behind the Joint

2. The Global Actuator Footprint: East vs. West

3. The Friction Points: Feedstock vs. Process Control

4. Falling prices can coexist with tight capacity

5. Magnets: The Downstream Chemical Chokepoint

6. What could limit actuator supply at scale?

1. Bill of Materials: The Processes and Players Behind the Joint

We walked through each component from an engineering perspective, covering the metrics and what they do in Post 1, section 1. To dig into the supply chain dependencies, I find it helpful to look at these components from a manufacturing lens: the material, manufacturing process, and supplier set.

Quick refresher: the actuator is the muscle of the robot. It moves the joint. A shoulder actuator needs to be strong, since it is lifting a heavy arm and whatever the hand is carrying. A finger actuator’s speed requirement depends on the task.

Stator and motor. One option is a frameless brushless motor. In a slotted design, the stator is a stack of steel laminations with copper windings. The rotor carries permanent magnets. Frameless versions ship as rings that the robot maker fits into its own housing. Unitree’s filing states its motors are self-developed, and that it outsources some production steps, winding among them. Merchant suppliers include maxon, Kollmorgen (a Regal Rexnord brand), Novanta, MOONS’ and ThinGap.

Rotor magnets. One option is sintered neodymium-iron-boron (NdFeB), made by powder metallurgy: mill the alloy, press it in a magnetic field, sinter it, machine it, coat it, magnetize it. These are the magnets whose field interacts with the energized stator windings to produce torque: they are what makes the actuator spin. Processing is concentrated in China, which accounted for 94% of global sintered NdFeB magnet production in 2024, according to an IEA report published in April 2026. [1] Suppliers include JL MAG, Ningbo Yunsheng, Earth-Panda, Vacuumschmelze, Proterial, Shin-Etsu Chemical and TDK.

Reducer, low-ratio planetary. This is a quasi-direct-drive route, covered in Post 2. Precision planetary gearboxes can use ground gear teeth. Planetary-reducer suppliers include Zhongda Leader and Newstart.

Reducer, strain-wave. This is a precision route for robot joints. A strain-wave reducer uses a thin steel flexspline that deforms during operation. Suppliers include Harmonic Drive Systems, Nidec-Shimpo, Leaderdrive, Laifual and Schaeffler. Harmonic Drive LLC also has a Massachusetts production facility. This layer is expanded in sections 3 and 4.

Reducer, RV and cycloidal. This is the heavy-load precision route. The named suppliers are Nabtesco of Japan, Shuanghuan via its Huandong unit, and Zhongda Leader.

Planetary roller screw. This is one linear actuation route: a rotary motor turns a roller screw to create straight-line motion for high-force joints. It has a threaded screw, a set of grooved planetary rollers, and a nut. It is difficult to make at the high end because several rollers must share the load at once. Small errors in thread shape, pitch or alignment create uneven contact. See more in Section 3. Suppliers include Rollvis and GSA of Switzerland (one group since 2016), Ewellix (a Schaeffler company, with roller-screw production at Armada, Michigan), and Bosch Rexroth. Chinese entrants include Leaderdrive, Seenpin, and Hengli, which reported sampling and initial mass production in 2025.

Encoder. Encoder options include magnetic rings and optical discs read by sensors. Suppliers include Renishaw, Heidenhain and Broadcom.

Output bearing. Cross-roller bearings are one option for supporting joint loads. Suppliers include IKO, THK and Schaeffler. Laifual reports making crossed-roller bearings for its own products; Leaderdrive’s SKF joint venture is focused on R&D and industrialization of high-precision robot-joint bearings.

Driver board. The board carries power-switching and control electronics on a PCB. Servo-drive suppliers include Leadshine, Elmo and Kollmorgen; CubeMars lists a standalone driver board.

Housing and brake. The housing and brake depend on the joint design. Spring-applied brakes are one option, such as Mayr’s ROBA servostop.

2. The Global Actuator Footprint: East vs. West

The supplier examples span Japan, Europe, China, South Korea and the US. US production includes Harmonic Drive LLC’s Beverly, Massachusetts facility for precision reducers, Ewellix’s Armada, Michigan roller-screw plant, ThinGap’s Camarillo, California motor production, and eVAC’s Sumter, South Carolina magnet plant. VAC said in July 2026 that eVAC began commercial magnet production in 2025. Sanhua said in its 2025 annual report that it is expanding overseas production of its actuators, and updated in Aug 2026 that it’s processing through batch delivery and production-line ramp.

These are representative suppliers of actuator components. Exhibit 1 below records disclosed manufacturing locations where they could be matched to the product; other entries are marked manufacturing location unverified.

China’s actuator market is split between precision transmission specialists, integrated module suppliers, and automotive Tier 1 manufacturers pivoting into robotics.

3. The Friction Points: Feedstock vs. Process

Now this is the exciting part. Now that we know at a high level how each component is made and who makes it, it’s time to zoom into the requirements and potential exposures for each component. Broadly speaking, there are two places to investigate. (1) Feedstock and bought-in parts: the raw material, plus the merchant components that arrive finished, like sintered magnets and encoder chips. (2) Process and machining: what turns the raw material into the part, and the machines that do it.

Strain wave reducer. The process requirement is manufacturing control and accumulated know-how around the flexspline. The manufacturing challenge is making a thin flexspline that repeatedly deforms while maintaining gear accuracy and the required fatigue resistance. Material selection, heat treatment and machining must be controlled together.

Planetary reducer. The process requirement is manufacturing precision gears. Grinding the gear teeth is one process used in precision planetary gearboxes. Newstart, a Zibo, China-based maker that filed for a ChiNext IPO, reported ASPs for its mainstream 50-240 mm planetary reducers of RMB 780 in 2023 and RMB 791 in 2025. Its whole-company gross margin fell from 55.49% to 48.05% over the same period. [4]

Quasi-direct drive. The performance requirement sits in the motor itself. With less gear reduction, the motor must supply more torque for the same joint requirement. Take Direct Drive Tech as an example. This company sells direct-drive actuator modules for consumer, industrial and commercial applications. Its draft PHIP shows that, from 2023 to 2025, enameled copper wire rose from 6% to 27% of group material costs while the magnet share fell from 23% to 18%. These are group material-cost shares across its product mix. [5]

Planetary roller screw. The potential process exposure is the machining step. Leaderdrive’s H1 2026 report describes Chinese industry dependence on imported high-precision thread grinders, with high equipment prices and long delivery times. [6] Seenpin’s June 2026 application similarly describes reliance on imported high-end precision grinders. [7]

Rotor magnets. The potential exposure sits upstream in rare-earth separation and downstream in NdFeB sintering capacity, with China at 94% of global sintered NdFeB production in 2024, according to an IEA report published in April 2026. [1]

Encoders and driver boards. The potential exposure is bought-in availability: merchant chips and read-heads for the encoders, and the power semiconductor chain for the driver boards.

4. Falling prices can coexist with tight capacity

Laifual’s disclosures put price and capacity on the same page. They also show why a lower average selling price does not answer whether a supplier can meet the qualified weekly output for a particular design.

From 2023 to 2025, Laifual’s blended harmonic-reducer ASP fell 28%, from RMB 795 to RMB 571. Its small-size category fell 37%, from RMB 634 to RMB 397. These are category averages, not same-SKU price changes. Laifual attributed the 2023–2024 decrease to a higher share of small-size products and strategic price adjustments, and the 2024–2025 decrease primarily to the higher small-size share. Over the full 2023–2025 period, its harmonic-reducer sales rose from 115,315 to 291,515 units. [8]

H1 2026 tells a separate capacity story. Laifual reported designed harmonic-reducer capacity of 270,000 units and actual production of 256,900 units, a 95.1% utilization rate. The company also said that, against a backdrop of earlier relatively constrained production capacity, it had prioritized industry-leading customers. By July 2026, designed capacity had reached 64,800 units per month, so the H1 utilization figure should not be read as a September measure of spare capacity. [9]

A September 2 seller offer adds a point-in-time procurement reference, not proof of a fulfilled ramp. Laifual offered size 17, 20 and 25 component sets, each at a 100:1 ratio and a quantity of 100, for US$145, US$160 and US$173 per set EXW, respectively, with a stated lead time of three to four weeks. [10]

The buyer question is concrete: for the specified part and qualification standard, how many units can the supplier deliver each week while maintaining quality?

5. Magnets: The Downstream Chokepoint

Magnets represent a highly concentrated layer of the actuator supply chain. Recall that permanent-magnet motors carry magnets on or within the rotor. Electric current flows through the copper windings in the stator, creating an electromagnetic field. The rotor magnets follow that field to spin the motor. Torque per unit of mass is a critical metric for a humanoid robot that must carry its own weight.

These are made from a sintered alloy of neodymium, iron, and boron (NdFeB), commonly called “neo magnets”.

It turns out that “rare earths” are actually not that rare. Economically viable concentrations are uncommon, and their chemical similarities make them hard to separate. [11]

To map the potential exposure for robotics, I group the value chain into 3 phases using International Energy Agency (IEA) data. All three are concentrated in China:

1. Mining (Upstream). Hard-rock ore is extracted, crushed and concentrated. China supplied 60% of the four magnet rare earths (neodymium, praseodymium, dysprosium and terbium) in 2024. Other producers include Myanmar, Australia and the US. [1]

The potential exposure here is access to economically viable deposits. Extraction routes differ: hard-rock ores require crushing, while ion-adsorption clay deposits, including those in Myanmar and southern China, can be leached.

2. Separation and refining (Midstream). This is a highly capital-intensive and chemically complex stage, and China accounted for 91% of refined output of the four magnet rare earths in 2024. A common process is Liquid-Liquid Solvent Extraction, routing dissolved rare earths through repeated extraction, scrubbing and stripping stages. [1]

The potential exposure here is environmental policy and process know-how. Waste treatment depends on the ore and processing route. Where the ore contains thorium or uranium, processing can concentrate radioactive material in residues.

(3) Sintered permanent-magnet production (Downstream). Metals are alloyed, milled into powder several microns in size, pressed in a magnetic field and sintered. [12] China accounted for 94% of global sintered permanent-magnet production in 2024. [1]

Two processing challenges in this stage are machining yield and grain boundary diffusion.

• Machining and Yield Loss: Shin-Etsu describes grinding sintered magnets with diamond abrasives to reach the required dimensions. Removing material during machining creates a yield-management problem. [13]

• Grain Boundary Diffusion (GBD): A heavy-rare-earth diffusion source is brought into contact with the magnet, followed by diffusion annealing. This can improve coercivity, the magnet’s resistance to demagnetization. [14]

Getting magnet supply set up is so much more than finding and opening a mine. The potential constraints include navigating midstream separation chemistry, waste management costs, and downstream precision processing.

§6: What could limit actuator supply at scale?

Congratulations on making it to the end of this post! This piece has got longer than I originally anticipated. But this also shows how deep and complex supply chains are. As a final reward of finishing the post, let’s apply what we’ve learned on the supply chain to tackle the million dollar question: if annual humanoid production rises from an illustrative 20,000 robots to one million, what breaks first?

Read more

Scott is a mechanical and aerospace engineer, a 2-time robotics founder, and one of the sharpest public readers of humanoid-hand architecture. He has worked through Tesla’s hand iterations, Figure’s decision to abandon its first tendon-driven design, 1X’s routing choices, and the direct-drive approaches behind Wuji and Sharpa Wave.

This was a particularly fun and wide-ranging conversation. We used three extreme designs to escape the usual tendon-versus-motor argument: Allonic’s braided structure, Daxo’s maximalist system with up to 120 tendons, and Tacta’s hydraulic hand, glove, sensing, and data stack. Tacta is especially interesting because it is designed as a complete manipulation system for cobots and industrial arms, without trying to fit inside a humanoid.

The side paths were just as useful. Scott proposed a decathlon for humanoids, complete with pole vault, hurdles, baton exchange, and the same sand pit humans use. We also asked whether robots should copy human movement at all, and whether a general-purpose body still becomes specialized through its profession, as humans do.

In this episode we cover:

• Why Tesla and Figure reached different conclusions about tendons

• What Bowden tubes solve, and what they add at the wrist

• Why joint count and independent control are different metrics

• What the Humanoid Games reveal about robot-specific movement

• Scott’s proposal for a humanoid decathlon

• What Allonic, Daxo, and Tacta teach at the design extremes

• Why Tacta is building a manipulation system outside the humanoid

• How generalized bodies can still become specialized workers

Scott Walter: https://www.linkedin.com/in/scott-walter-ph-d-b2a78ab/

RoboStrategy: https://robostrategy.co/

Watch on YouTube: https://youtu.be/VbwM0t1oJQI

Every episode, I sit down with the founders, researchers, and operators building physical AI worldwide, tracing the full stack from components and supply chains to what actually deploys. See all published episodes here.

Chapters

00:00 Cold open: the labor prize and the 5-year forecast

00:24 The Fermi paradox of robotic hands

01:18 Tesla, Figure and the tendon debate

05:10 Bowden tubes and routing tendons through a wrist

11:15 Joints, actuators and real degrees of control

14:38 Why the pinky matters more than it looks

17:23 Humanoid Games as real-world robotics experiments

21:57 Scott’s humanoid decathlon challenge

26:28 Tesla’s hand iterations and abandoned designs

27:55 Wuji and Sharpa Wave put motors in the fingers

38:29 What 3 extreme hand designs can teach us

38:53 Allonic’s braided hand and portfolio disclosure

51:15 Daxo’s maximalist hand with up to 120 tendons

01:02:36 Tacta measures how workers use their fingers

01:05:18 A robotic-hand system designed outside the humanoid

01:13:12 Data-center cables as a tactile manipulation task

01:16:22 General-purpose bodies and specialized work

01:19:18 Why this 5-year robotics forecast may be different

Full Transcript

Below is the full transcript. There may be errors as it’s AI-transcribed.

00:00:00 — Cold open — excerpt montage

Scott Walter · 00:00:00

That is the biggest pie we’ve ever seen, ever, trying to automate labor. And sim-to-real gap is almost zero. So locomotion is a solved problem. Humanoids will be deployed like five years after everyone stops laughing. And they’re the worst they’re ever going to be. They’re just going to get better and better and better.

00:00:23 — Why dexterous robot hands are still hard

Scott Walter · 00:00:23

And what I call the Fermi paradox of robotic hands. If they exist, where are they? And that’s always been the problem: you see them in the labs and then you go around and you’re like: I’m not seeing them anywhere. No one’s able to do them. And the question is: why is that the case? It’s extremely challenging to come up with something that has not only dexterity, something you can control, but also robustness. But what I was seeing during the teens is that robotic hands were suddenly becoming a possibility because machine learning made it possible. Before then, they were just really, really hard to control.

Scott Walter · 00:00:56

And that’s what the Shadow Hand proved: you can have this tendon-driven hand. And if you connect it to really good machine learning algorithms, you’d be amazed that you’d be able to do. The problem was that the robustness wasn’t there, the cost factor, everything else. So that’s kind of the preload that we’re seeing it. And why aren’t they becoming more and more available? Now, the big debate that is going on in hands in general.

00:01:18 — Tesla, Figure, and the tendon debate

Scott Walter · 00:01:18

And so people really started paying attention to robotic hands when Elon announced a 22 DOF hand that was going to go on Optimus and the revealing of that and everything like that. And after that it was like this Cambrian explosion of just robotic hand designs all over the place and everyone pursuing different designs. And for the most part, you could break it down into two categories, tendon-based hands and non-tendon-based hands, which most people would say direct drive. But even there, you have to break them down because you can never really put things in two different buckets.

Scott Walter · 00:01:52

They’re all going to go down there. But you can kind of look at it that way. And everyone’s pursuing it. And they all were very passionate about their designs and why theirs was the best approach. Some people would try one and they would kind of think about it again and then try a different approach. But about two weeks ago, Brett Adcock put out a post where he talked about the evolution of the hand designs at Figure and that the very first generation hand that they built was a tendon-based hand. And after attempting to do that he came to the conclusion that a tendon-based hand was basically an engineering dead end.

Scott Walter · 00:02:24

And he called it the biggest engineering mistake he’d ever made. And since then, he’s been pursuing hand designs that are non-tendon-based. And that kind of caused, let’s say, a very spirited discussion. And I’m glad it was out there because you got a lot of people that were very much on the side of tendons just going out and just saying he’s completely wrong. From right down to saying it’s a skills issue to explaining why they believe it and their particular thesis for it. As well as you get a lot of support from other people who said, yeah, we tried tendons as well.

Scott Walter · 00:02:55

It turned out to be really hard. Now, I’m not here to defend one position or not. I’m really here to maybe kind of present it to everyone so everyone can kind of come to their own conclusions because they both have their pros and cons. They both have the big engineering challenges in what they’re trying to do. And even within the tendon community, while they agree tendons the right approach, they will get in very spirited arguments about what’s the best approach or not. And you can see that 1X came out with their 22 DOF hand, I think.

Michelle Sun · 00:03:25

I think it’s 25.

Scott Walter · 00:03:28

Yeah, 20. I think it’s actually 22, 44 tendons. It’s kind of hard to know the count. They’re also throwing the wrist in there. So I usually don’t put the wrist in there. But they have what you would typically say is about 22 degrees of freedom of movement of the fingers. Quite complicated design.

00:03:43 — Routing tendons through the wrist

Scott Walter · 00:03:43

And the biggest challenge with all the hands is not having the tendons that can actually move the fingers and putting down the forearm, but figuring out how to get them through a wrist. The wrist isn’t in there. It’s an easier problem. I’m not going to say it’s easy. It’s not. But as soon as you do that you add this extra level of complexity that the engineering solutions are challenging. And bodies are fairly simple to model. But when we start getting into flexible stuff, there’s a lot of challenges. And we are doing it in our heads because we’ve been doing it for so many times that we know we can move our wrists and somehow keep our fingers constant.

Scott Walter · 00:04:19

But if we’re not doing that we will see that our fingers will kind of move as we are moving our wrist just because the length is changing. Now, in theory, you can solve the problem by having all your tendons go right through the center of rotation of your wrist. Okay? In theory, you could do that. The reality is there’s something called, I think it’s the Pauli principle. Two particles cannot occupy the same point in space at the same time. So you need room. It’s like you’re talking about before tendons. They can’t all occupy the same spaces.

Scott Walter · 00:04:51

So when they go around there, there’s just always going to be that. And then there’s this other problem is that when you’re doing it, sometimes if the tendon’s coming straight up and you’re going, that means suddenly you have to do this very sharp angled turn. And tendons do not like sharp angled turns. You need to have a kind of radius in there. So there’s big challenges to being able to do that. And there’s different schools of thoughts on how to do it. And there’s this thing called a Bowden tube. And everyone’s like: whoa, what’s that?

Michelle Sun · 00:05:15

Is that what 1X is?

Scott Walter · 00:05:20

Yeah, they’re using the Bowden tubes. And basically, if you’ve ever ridden the bicycle and look at your brakes and the brake cables, that outer tube on the outside, on the inside is a wire, which is effectively like a tendon. And the tube helps redirect that and maintain the same length. That’s the trick. That’s why you can turn your handlebars without suddenly the brakes engaging or disengaging or something like that. Or your gears changing or anything like that because you’re able to maintain the length through those things. Now, those are already engineering challenges. They have friction.

Scott Walter · 00:05:53

They take up space. They can snap as well. So within the hand community itself, there’s like a big debate on whether to do Bowden tubes or not. And the original Tesla hand that they showed at the 1010 event had them. And then the next time they had the public showing of them, they didn’t have them. And so everyone’s trying to figure out how do we route these things through there. And this is why you can see all the people that are building the direct drive ones have given up. They just say that is just a challenge.

00:06:23 — Tendon materials, flexion, and extension

Scott Walter · 00:06:23

The other main thing is what is the material you use for a tendon? If you use a cable like you see in a bicycle, the problem with that is the bending radius is not very good. And they tend to elongate over time. Anyone that’s had a bicycle knows you have to go and retune everything and make those adjustments because they’re going to stretch. And the other thing about metal is that anytime you bend it, there’s energy involved in the flexing of it, which means that’s energy going not to your hands. If you use something like rope or thread, it has much better properties that way.

Scott Walter · 00:07:00

It’s very flexible, very easy to move around. The radius of curvature is a little bit different, but they can fray and also stretch and everything else. So you’re worried about wear and tear. But there’s this magic material out there called Dyneema, which is used for climbing ropes and everything else is extremely flexible. It’s this aramid polymer, which is just kind of miraculous and doesn’t stretch very well and has very good abrasive properties. And it’s readily available. You can go on Amazon. You can just order it in any color you want. So anyone that is, an amateur building their own hands at home by using a 3D printer, and there’s plenty of open source hand projects out there, the Paloma project is one example.

Scott Walter · 00:07:39

You just put that in there. So it’s very easy. And they’re all using the same thing. Rock climbers use it for a reason because it’s like it doesn’t break and it doesn’t stretch. And also, it’s got this one great property that another kind of aramid known as Vectran doesn’t have, and that is it does not break down with UV light. Okay? So Vectran is very sensitive to UV. It will break down. But it has some properties a little bit better than Dyneema. And if it’s inside, ah, maybe it is that. So usually it’s two of those.

Scott Walter · 00:08:10

Or there might be some sort of, magic hybrids of these things where sometimes people might even put in stiffness. So by adding metal. One advantage of a wire cable: the problem with any sort of thread is that it’s really great in tension, and as soon as they’re trying to push on, it collapses. So I can only apply force in one direction. Bicycle cables, we know, work better in one direction, but they still work a little bit in the other. They will a little bit, but usually on the brakes, we still have a spring to help open them up.

Scott Walter · 00:08:42

The problem with a tendon-based hand is that you’ve got two ways of moving. One tendon gives you the flexion. And in order for you to open up your grasp, what they call extension, you need a second tendon, which is called extensor, to pull it back open. Or you need a spring return. And that’s not necessarily a great way to do it. And that’s because when you push on a thread, you’re not getting anything out of it. With a Bowden tube, you can do it a little bit, but mainly all you’re doing is taking the pressure off as opposed to pushing it back.

Scott Walter · 00:09:15

Unless it’s a bicycle cable and unless it’s human tendon. It turns out human tendons are an amazing thing because they can take pretty heavy loads and they actually can give a little bit. So our tendons are able to push back, not 100%, but much better than Dyneema. So there’s already a difference between what they’re using and an actual human tendon. Yeah. No one’s made that material yet because it has, the properties of a metal cable without being a metal cable. It has the properties of a Dyneema without being a metal cable. It’s like if someone could figure out how to make that thing, that would be a lot better because then you would be able to push pull.

00:09:50 — Direct drive, linkages, and degrees of control

Scott Walter · 00:09:50

And this is where the people that want to have what I call the direct drive are going to be in there. And there’s a couple ways to break down the direct drive. The first way is, everyone was trying to say, where’s the space to put our motors? If we can’t put them down in the forearm because we don’t want to have to go through this wrist mechanism. Well, it seems like there’s a lot of space in the palm. So everyone says, let’s go ahead and put something there. And the first generation Tesla hand did that but they actuated the hand actually with a tendon, which is a metal tendon.

Scott Walter · 00:10:20

And they realized all those things stretch and stuff like that. But the run link was very short and you didn’t have to run it through here. Others said, well, wait a minute. There’s another way that we can do that have the same thing with a tendon. And that is to use a linkage-based mechanism for that. So they created basically these mechanical links that go through there. And the advantage of a mechanical link is that it works in both directions, in tension and compression. So that means you only need one motor and it can give the same amount of grip strength as well as opening strength.

Scott Walter · 00:10:50

It can go both ways and it can lock in really tight and everything like that. Whereas if you’re using a tendon, you would need the second tendon or you need a spring return. In some cases, you might need a second motor unless you wrap it around like a capstan. But then it’s hard to separate the tugging from both of them because there’s a lot of complications here. Now, when we talk about degrees of freedom in joints, they’re two completely separate things. We have a lot of joints in our fingers. Our fingers have, you consider, four degrees of freedom.

Scott Walter · 00:11:21

Now, you might say there’s only really four IP joints. One of them is not called an IP joint. It has to be called the MCP down here. But it has two degrees of freedom. It can flex, and it can do this ab/adduction that everyone talks about. And then we have the other IP joints here. But for the most part, there are four ways that it can move, four joints there. The question is, how do you control that? And degrees of freedom really means more or less degrees of control. So if I have a single tendon through there with a single motor, I have only one degree of freedom.

Scott Walter · 00:11:49

And then what’s going to happen is there’s going to be coupling between them and my finger is just going to move like that. It’s very difficult to get different degrees of control without adding more motors for that. And so what ends up happening with a linkage-based mechanism is that you already have the coupling. And in many cases, there’s only four or five motors in there. So those fingers really only had five or six degrees of freedom, depending by the number of motors, even though they may have had like 15 joints, 15 ways of moving. And so what happened is that they would collapse down like that.

Scott Walter · 00:12:21

We assume they’re all identical. It’s like: nah, there’s a lot of stuff. So the thumb has way more kind of range, different kind of range of motion, more tendons than the other. I think our thumb has anywhere between eight and nine. And it’s like: well, what’s the answer? And it’s like: depends who you ask because it just shows how kind of nuanced some of these answers people are going to have. But typically, it’s more than what you would see in the others. And the thing with this closing like that is that we really don’t want our fingers to close this way where everything is coupled and there’s only one way of closing.

Scott Walter · 00:12:51

And that is, it means when you go to grab a bottle or something like that your fingers can actually miss. Try coming down and suddenly, we want to have the straight move and then have that clasping. So that’s something that’s important. Now, you can kind of do that with a mechanical linkage in that what you do is you have like an escapement mechanism, block mechanism. So it’s like the very clever that you can get it that when it’ll move like this and once it gets there, something kind of engages and now the fingers do this.

Scott Walter · 00:13:17

Okay. So you get this nice illusion that you have more degrees of freedom of control there and you really don’t. You’ve programmed it in. It means, oh, I can’t do this. I have to do that to close it. It’s the reality. But it’ll work in a lot of cases and you can get away with fewer motors and you do have that nice strength of being able to open things up really quickly. So that’s a lot of the advantages of this direct drive: it’s almost tendon-based, except the tendons are super rigid. One of the problems is when you get those mechanisms, it’s pretty easy to build linkages that will just give you this.

Scott Walter · 00:13:49

But once you do that you’re starting to have that same problem we talked about with the wrist. Okay, I get some mechanism that’s pulling up and down here. How am I going to build this linkage mechanism that not only is going to do this, but allows me to suddenly redirect everything along there. It can be done. It’s just that it’s a more complicated mechanical design and then the decisions of how you control the abduction. So we’ve got basically five ways of doing abduction. Do we have five motors or do we have a single motor? Those are the decisions everyone has to make.

00:14:18 — Why the pinky matters

Michelle Sun · 00:14:18

Yeah. And one of the things that stood out to me about the fingers is that for carrying heavy loads, right, I read that the pinky actually does a lot of work as well. So it’s more than we think and we don’t think about it when we do that.

Scott Walter · 00:14:35

Oh, way more than you think. The other thing that pinky does: there’s this big argument, and one of my podcast colleagues, Gustav Andersson, has pointed this out. I’ll ask you a question. If you had to lose one of your fingers, which one would you choose?

Michelle Sun · 00:14:54

Probably the ring finger.

Scott Walter · 00:14:56

Okay. Some people say the ring finger. Some people would go with a pinky. And it turns out he actually says it’s the index finger. I’m like: what? He’s a hand surgeon. He knows better. And part of it is because if you take out the ring finger or something like that you get like a weird gap. The pinky is more important than we think. And it turns out that our middle finger can take over for the index finger very easily. People are able to get that kind. That was, kind of shocking to me. And I only assumed it.

Scott Walter · 00:15:22

And the main thing about the pinky is not so much that it’s strong, but where it is, where it’s placed. When you want to grab something with authority, if I’m just grabbing it with these two fingers, you see it wants to move around. And if I get the others in there, it still can move around a little bit. And you’ll see this when, these robot hands are picking up a hammer. A lot of times, the hammer is really not in there. Well, it turns out when you get that pinky engaged in there, suddenly that thing is, just mechanically locked better than when the others are in there.

Scott Walter · 00:15:48

And a lot of it just comes down to simple physics, and that is you have kind of a fulcrum that whatever you’re grabbing usually is resting right in there. And then these others are trying to come around, but they don’t have as good a lever arm, right, because the distance of where that’s happening. The pinky comes down here, and the distance from here to here is so big. It doesn’t have to be strong. It gets it just from simple mechanical advantage of the distance.

Michelle Sun · 00:16:10

Yeah, it’s locking in that grip and stabilizing it.

Scott Walter · 00:16:14

And you see that with a lot of tools that you have, the way the grips are all set up. Is that pinky is in there that gives you that final amount of stabilization that is remarkable. And again, it’s just kind of a geometric thing. And why the nuance and the subtlety of what goes on with hands is so underappreciated until you start going in there. Now, this is the other thing that I invite everyone at home to do. A lot of times when you’re designing stuff, you have to make these prototypes and experiments. So, if you’re building a car or drones, you just got to go in there.

Scott Walter · 00:16:45

You can think about it a little bit, but you got to build the darn thing. When we start making hypotheses about how a humanoid works, you don’t have to build a humanoid. You are a humanoid. You can run the experiments on your own. And that means you can ask the question, what happens if I reduce the degree of freedom? What if this is removed? What if I try to do everything in the house without my pinky? So, just attempt to see what you can do and how important it is. Take, basically, medical tape, something like this, and just, bind some of your hands.

Scott Walter · 00:17:15

Or put on, really thick gloves. Restrict motion of, one of your shoulder joints. And just ask the question.

00:17:23 — Humanoid Games and real-world experiments

Scott Walter · 00:17:23

And so, a little bit of a cul-de-sac here: the humanoid games in China were fascinating to me. Because we get to see people running these real-world experiments. So, they had these highly optimized bots for running in the 100. And we could say, they’re not truly humanoid because they took away a lot of degrees of freedom. But I am so happy they did that because we are always getting in an argument about how important is that joint to running? How many degrees of freedom do we really need in the shoulder to be able to run effectively and stuff like that?

Scott Walter · 00:17:55

And you can run them in a simulation, and they kind of give you an answer, and you can intuit yourself. But there’s nothing like real-world data. And seeing how they did it, and then you begin to go, huh. And you might say, yeah, what’s important is, we finally understand the nature of a particular joint or particular axis to a degree that we’ve never had before. That we now can understand what’s a better way to optimize it when we want to build the complete one. So, I love the fact that they’re going through, and you have these very specialized events that are forcing you to answer that question in a way that can only be demonstrated with a real running experiment, if you want it that way.

Michelle Sun · 00:18:35

Yeah, for sure. And one of the races I saw that some run-up to humanoids were, leaning forward the whole way. And that was just, really fascinating and, different posture that usually we don’t think about running in, and they just experimented with it.

Scott Walter · 00:18:52

Yeah. And I think one of the best ones is, you may have seen it, is, everyone complains about, AI slop videos or, how people just use AI in a way that’s oh, it’s so bad. And you may have seen someone actually took one of those races and replaced them with humans. So, you saw the human runners, and I was this is great, because what more effective way of explaining what that running form looks like than to do that direct comparison of how they were leaning and everything else? The other thing is, we’ve got to remember, is, they are not built exactly like us.

Scott Walter · 00:19:26

They may look like us, but their musculature is very different. So, they have different power density. They have slightly different kinematics. The weight distribution is very different. So, we shouldn’t be surprised that they have this awkward running form. Now, I do not recommend that human runners run with their arms like that and lean forward, because you’re going to end up with, back pain. But it turned out that the engineers were trying to solve a particular problem, and it was, kind of overheating. And they ran, RL sims to see how we could do that to minimize the movement on one of the shoulder joints.

Scott Walter · 00:19:57

And that was the solution it came up with. And it turned out the solution was highly optimized, better than they all expected. And this is the problem with dealing with human priors, is that everyone’s, forcing their walking policies to look human-like. But it may be that a human-like walking or running policy is not actually the optimum. It might be the suboptimum for these creatures. And so, they come up with it. And I think someone mentioned it. It’s basically, it evolved the tail backwards. Because, velociraptors and others, they need the big tail back there to do that.

Scott Walter · 00:20:25

And effectively, they said they weren’t allowed to have a tail, but they kind of did it that way. It’s, yeah, that’s an interesting way of thinking the problem. So, as silly as that looks, again, for me, I am super excited because I’ve always wanted to know. These are these academic debates that we have all the time. No longer academic. We can point to something. And this is why he’s saying these races are going to be studied for years. Even though they’ll be surpassed by all sorts of things, when you go into teaching any sort of engineering class, you’re going to want to bring that thing up and just say, take a look at this.

Scott Walter · 00:20:57

Just like the Tacoma Narrows bridge disaster is still taught in every engineering class on structures and dynamics and everything else because it is so educational. And that 100 meter we saw, it was incredible because of the different phase transitions, again, pitter-patter, how they were going to accelerate, stand up, lean forward. What are you supposed to do? It’s, every track coach will love that as an example to show their sprinters, oh, this is why we want you to do this. Do you understand what we’re talking about now?

Michelle Sun · 00:21:29

Yeah, and I love that they actually have different games, just like the Olympics, right? There’s the sprints, there’s also the long jumps. So the makers can really, design very different robots to optimize for that and train different policies. So it really is pushing to the extreme: what is possible? And how do we rethink the whole architecture from, the joints where they’re put and how they run, how they jump?

00:21:57 — The humanoid decathlon challenge

Scott Walter · 00:21:57

Now, I pointed out already in a couple other podcasts, I’m going to do it again, and I’m putting it out there very, very soon, is I’m throwing down the gauntlet for the humanoid decathlon challenge. And that is actually a humanoid that can do the decathlon. And that’s very challenging because it’s not just about locomotion anymore. There’s a certain amount of manipulation. We can argue whether the degree of the manipulation is totally at the level of human dexterity. And I think you’ll be surprised. It’s pretty close. So, arm strength is going to be incredible, whole body coordination.

Scott Walter · 00:22:30

You’re going to need really good grip strength. You’re going to need, basically, some wrist motions to be able to do some of the events. So, I think it’s a really pretty good proxy of coming up with a humanoid, which is going to be much closer to a real human. And then we can start arguing about world records at that point. So, we would say, what is the form that’s able to do it? And if it’s able to do it successfully, if it breaks the record, then we can say that’s on par. Because Usain Bolt is able to do a lot more things than the robots that broke his record.

Scott Walter · 00:23:02

And one of them, he was able to stop without running into a wall. Okay. He was able to run the curve. And the other thing he pointed out is, not only did he get a gold in the 100, he got a gold in the 200, which meant he had to run the curve. And those bots really could not run the curve because they reduced the degrees of freedom to make sure they’re really good going straight. And the last thing is the last gold medal he got in was in the 4x100 relay, which means there’s a baton exchange, which means there’s a certain amount of manipulation.

Scott Walter · 00:23:30

You actually have to be able to receive the baton and hold on to the baton. And so, in order to get parity, to make it fair, we have to say: what do we do to force someone to build a humanoid that has all these required degrees of freedom, that has to do the 100, has to do the hurdles, has to do the high jump? Has to do the pole vault. The pole vault is going to be a test of strength, not just grip strength, but upper body strength. We didn’t see any robots doing pull-ups. A lot of them aren’t even strong enough to do a pull-up.

Scott Walter · 00:24:04

It’s like we shame how a lot of people can’t do a one pull-up, but the humanoids really, they don’t have enough strength to be able to lift that. So, when you start thinking of an athletic competition like that’s going to be the true test. And again, I’m going to put out the ground rules and everything like that. I hope next year at the games that they have a decathlon in there and they attempt it, and it may be that they struggle. Just like last year, they struggled in the games, right? It was comical. This year, it wasn’t comical.

Scott Walter · 00:24:33

It was like Formula One, NASCAR, kind of disasters going in there, but at the same time, incredible performances. So, you can see where you can go in one year. One year, I have a feeling these bots might struggle through a decathlon, but then they will learn so much that it’s possible. It’s possible in two years they could make an attempt at the men’s world record in the decathlon. So, that’s out there. That’s the challenge I want to have. Let these labs get started on that.

Michelle Sun · 00:25:07

Yeah, definitely. Well, then you can start thinking about building that. And I think we can really expand into all the Olympic sports, right? Like the team sports can be expanded into, and like that will be really fascinating.

Scott Walter · 00:25:22

Cycling, kayaking. You can go down and down the list. And eventually, there will be a humanoid that will probably be able to perform in all of those things. But kind of steps. There’s a lot of things that – strength and dexterity in the decathlon that most people agree is pretty challenging. But there’s not going to be the finesse like we’re not going to have it play the piano when it’s over. Maybe we should. It has to play the Olympic theme music. So, some things like that we know are maybe not fully encompassing, but we need some sort of proxy.

Scott Walter · 00:25:54

We already have the human data on decathlons and how challenging it is. So, let’s start with something we’re all familiar with. And again, exactly the same one. Not oh, we’ll do an event kind of like it. It’s like none of it’s the same. Right down to the long jump. You may notice the long jump, they were landing on the surface. The real long jump, you’re landing in a sand pit. And I have a feeling they didn’t want to land in a sand pit for a reason. And that is, sand gets in your joints.

Scott Walter · 00:26:19

And so, it’s like: no, no, no, no. If we’re going to do this, we’re going to do it with the exact same little ground rules.

00:26:27 — Tesla’s hand iterations

Michelle Sun · 00:26:27

Yeah. And do you want to go into the Tesla hand still?

Scott Walter · 00:26:33

Yeah, we can briefly talk about it. Because, again, we don’t necessarily want to talk about the hands that have already been talked about to death. So, the lead up of the Tesla hand is that it’s doing it very much like a human. And Elon talked about the advantages putting everything down the forearm. And there’s a lot of people that believe it. And now you’ve got to put tendons in there. And the idea is that you want full hand dexterity. You don’t want these hands that has minimal amount of degrees of freedom. But full hand degrees of freedom so you can play the piano.

Scott Walter · 00:27:02

So, the first generation one was seen at the Tevent. And it was like the prototype. And that was showing Bowden tubes. And if you look at it, you will see that as the wrist moves around that these tubes are kind of bending and going all over the place. And you’ll see the same thing on the 1X hand that they have that. They then abandoned that and came up with something else. And this is like their next generation hand. And there was a lot of patents that were filed on it. There were four or five patents.

Scott Walter · 00:27:27

And then when the patents came out, I did a patent review with Humanoid Hub on that. And one replied to Humanoid Hub’s post on that. It’s like: oh, we abandoned that design. Something like that. It didn’t work. So, a lot of the things that they were working on turned out to be kind of dead ends. And they’ve had to do something else. Of course, we don’t know what that is. And everyone else that’s building tendon-based hands is discovering the same kind of challenges and problems.

00:27:52 — Wuji, Sharpa, and motors in the fingers

Scott Walter · 00:27:52

And then if we go over to the other side of people who are in the direct drives, There’s two other ways of doing it. So, the Wuji hand. So, normally everyone says, let’s put the motors down here and either use linkages or maybe short tendons to be able to do that control. Wuji said, all right, what we’re going to do is we know how to make really small motors. And we’re trying to find space for where to put the motors. And everyone wants to put it in the palm. The problem is you do that. The palm gets pretty big.

Scott Walter · 00:28:18

And they said, well, wait a minute. What’s sort of the biggest point of space in the finger? And it turns out it’s what we call the phalanx, or everyone would call it like the bone. And so, they said, why don’t we make the bones out of motors? Because they’re really big. The other alternative, which we see in like the Sharpa hand and origami and some others, is that to put it actually in the joint. So, build your finger like you would build a robot where every single joint is actually the actuator. The thing is, you don’t have that much space there.

Scott Walter · 00:28:48

And they’re looking at, how can we get something a bit bigger to get the torque density that we’d like to have? So, it makes sense to put there. There’s just one challenge with that. And like with everything, every time you come up with a solution, there’s like these other challenges you’ve got to get around. It’s kind of frustrating. By putting it right at the joint, the motor is spinning in the direction you want the joint to move. And that’s great. And you may or may not need a reducer depending upon your torque density. And what I mean by reducer: a reducer reduces the speed or something.

Scott Walter · 00:29:15

So, if something is spinning really fast and you don’t want it running really fast, you put a gearbox on there. With the advantage, not only does it slow it down, but it gives it more torque. And so, if you have a motor that’s spinning super fast and your finger might close, it’s like: no, no, I don’t want that. I want to slow it down. So, the reducer is there and also give you a bit more grip strength. And the reducer takes up a little space and everything like that. But that’s one way of doing it.

Scott Walter · 00:29:36

But with the Wuji hand, now we’ve got way more space. The problem is the spinning is along the axis of your finger and not perpendicular to it. So, you have to come up with some sort of transmission system to do that. There’s two ways. One is like a bevel gear design that anyone that’s kind of familiar with the differential in the back of their cars is that kind of design. The other is something called a worm gear. And I made a guess just based on kind of like the offset of where the actual joint is that they probably are using a worm gear.

Scott Walter · 00:30:07

And they are. So, the worm gear and the knuckle, they don’t necessarily show up. But I think there were a couple others that they were showing. So that works great for your IP joints. And the thing that’s nice is the distal, the very tip, your fingertip, they don’t have a motor there. Because the motor for moving it is in the one before, which is great. And then, the motor for this one is down here. And then, they get down to the MCP. And the MCP is always the biggest challenge because it has to move two directions.

Scott Walter · 00:30:36

Now, typically, the order of operations of those joints is that flex. And then, the abduction is always a joint before mechanically when they build it. Wuji decided to switch that up. And the reason they decided to switch it up is rather interesting. It might be partly because they figured that there’s an architecturally easier way to do it. And the other is the most important joint when it comes to like grabbing something, as you’re referring, like lifting up your luggage, is that joint. That’s where the grip strength is for trying to hold something, whether it’s a bag of groceries or anything like that.

Scott Walter · 00:31:06

And the other two, basically, they don’t provide very much except they make sure it doesn’t slide off your fingers. And so, that means you would like that thing to have a lot of power. And if you change the order of operations, then it becomes a bit more difficult. And they decided, we’re going to actually put a motor with three times the torque output as the ones that are up here. And we’re going to use that for flexion on the first joint. And so, they did that. And you’ll kind of notice as you look at that finger.

Michelle Sun · 00:31:33

Is that the circle, like the silver?

Scott Walter · 00:31:36

Yeah, you see that circle? That’s your abduction. That’s your abduction coming after the flexion. Right. And they had to come up with an interesting four-bar mechanism to be able to do that. But when I looked at it, they really understand the nature of the problem. And it’s like different than what everyone else did. Now, worm gears are really good at mechanical leverage, way more than most people would give them credit to. The problem is they’re so good at it, they’re very hard to backdrive. They work really well in one direction, but not so much in the other.

Scott Walter · 00:32:05

And you want torque transparency, and you want compliance and everything else. They’ve come up with a second version of a worm gear. There’s another type of modified worm gear that does give you backdrivability. But what it does is it reduces the forward advantage a little bit. It’s like everything. There’s always this compromise. But now they do have a little bit of that mechanical compliance in there. And that’s very important with hands and everything is we want backdrivability. We want them to be compliant. The biggest problem with a lot of these direct drive in the fingers is that you’re moving a ton of mass out into your fingers.

Scott Walter · 00:32:42

So the more mass you have in the fingers, the harder to accelerate, the harder to have compliance and everything else in control. And this is the argument that everyone that’s doing tendon-based hands, and especially like Kyber, which has a very good tendon-based hand, is that their fingers are just so light that they can move around really quickly. So, again, that’s the pros and cons. The other thing is that as you start scaling these motors down, unfortunately, the scaling problem doesn’t work as well as you would like. The motor will overheat very quickly. So not only do you lose mechanical advantage, they can get very warm very quickly.

Scott Walter · 00:33:16

And that’s a challenge. So the difference between Wuji, which is direct hand, and Sharpa, which is direct hand: Sharpa is going straight to the joints. And Wuji is doing it there. They still are putting some down here in the palm, but they have such good control of the fingers and the positions of them that their sim-to-real gap is almost zero. So every research lab wants to use them because the sim-to-real gap is gone, whereas with tendon-based hands, there’s still a lot of finagling. It’s still very challenging to do that. And that’s let’s say, their major advantage.

Scott Walter · 00:33:48

And then everyone will point out like all these other advantages. And that’s why you have these different camps that are arguing at each other of I can solve this problem. And then they’ll point out, yeah, but you can’t solve that one. And you don’t have the perfect hand yet as a result.

00:34:00 — Counting degrees of freedom and scaling hands

Michelle Sun · 00:34:00

Yeah, and Sharpa is the one that even the pinky has an extra degree of freedom, right? Is that what you’re just doing?

Scott Walter · 00:34:09

Yeah. That’s what they call the fifth metacarpals. Ask 10 biomechanists how many degrees of freedom the human hand has, and you’ll get 20 answers. Definitely. I’ve researched this. You’ll hear some people say that there’s 27 degrees of freedom in the hand. No. Now, someone who says that is probably a gamer who has a hand avatar in their gaming system with their VR and are counting the first six degrees of freedom, which is the position of your wrist and orientation in space, and with the hand DOFs. It’s like: that doesn’t belong there. So if you take that away, it means you’re getting down to maybe 21 degrees of freedom.

Scott Walter · 00:34:46

Some people might say there’s 22, depending upon the metacarpal. And what that means is that we need to do something called opposition. And that is our thumb being able to oppose other fingers and also doing it over on the pinky. And when it comes to the pinky, it turns out our pinky can kind of swing over because we have what they call this metacarpal bone down here that actually has a joint down there. And it can kind of swing over a little bit, as well as the fourth can a little bit. So there’s some arguments on whether you need both in there.

Scott Walter · 00:35:14

And that’s why you start getting into this, how many degrees of freedom that it has in there. And you can argue any number you want, but generally it’s considered to be somewhere between 20 and 22, whether you want to do that. When it comes to two degrees of freedom, I’ve heard approximately 20. And you’re like: how can they say approximately? It either is or isn’t because it’s integer, right? And it’s like: no, it turns out if you take your thumb and move your thumb, you’ll notice that none of your fingers move. If you take your index finger and move it, you’ll notice no other fingers move until you get to a point that suddenly the other one starts moving down there and there’s nothing you can do to stop it.

Scott Walter · 00:35:46

So that means there’s a loss of independence between them. And that means they start having fractional degrees of freedom. And when you add them all up, it’s something like 29.75 or something or 19.75. It’s a weird number. Now, the Sharpa hand and the Wuji hand have absolutely no coupling between them. So they are truly independent. They can move any one of their joints independent of any other. Tendon-based hands, they actually do have some crosstalk, apart from the movement of here and like another tendon’s rubbing up against another tendon. If it pulls on it, it’s possible.

Scott Walter · 00:36:18

So there is not full degrees of control as you think there’s a lot, but there’s still some subtlety that you have to tease out in your control algorithms to make sure you get your finger position where you want to have. And again, I’m devolving everything here. I would say one of the most interesting hand designs I saw recently was like a 1.3 meter hand. So I’m going to make this giant hand. And he made it like the Sharpa hand in a way, and that each joint was a big actuator. And when you get up to that scale, it makes total sense because you can size it in a way that it does not overwhelm the size of your design because you can get incredible torque density when you start scaling up.

Scott Walter · 00:36:54

And so a tendon-based hand almost doesn’t make sense when you make them big. So if you want to have like this big giant sculpture thing, just build them like a robot arm and don’t worry about doing tendon drives or anything like that or cable drives or anything like that. Well, the only alternative might be like a hydraulic-based hand because hydraulics are great when you start scaling up, like in excavators. And you can kind of fit them in there that suddenly they’re not dominating the size of whatever it is you have. When you start scaling down, suddenly your hydraulic pumps and everything become massive compared to what it is you’re trying to move.

Scott Walter · 00:37:29

So you can kind of look at it that way. And what ends up happening is as you scale down, unfortunately, the torque density just doesn’t go down the way you want, which is why you have to start considering something like a tendon-based hand versus direct drive at human scale. And it’s like right there, right at that point. If you went way smaller, really, really small to the micro level, Maybe direct drive makes sense. But kind of there’s like an area in there which you might say, yeah, remote actuation is the only way to do it.

Scott Walter · 00:38:01

And it seems like tendons make sense at that point, not so much hydraulics as you get bigger. Okay. This is more like chapters one and two as opposed to a very simple prologue of what you want to talk about.

00:38:14 — Allonic — braided hands and portfolio disclosure

Michelle Sun · 00:38:14

Yeah. Well, do you want to dive into Allonic? I think this is a really good segue to when you mentioned—I need to look up the 1.3 meter hand. And whenever we design something, I think designing for the extreme always teaches us something unexpected, right? So that’s why I want to dive into three different extreme designs that we don’t see a lot and they are very surprising. And maybe you can share what it teaches us in terms of a new perspective and what’s possible. So Allonic—you told me about it, actually. Tell me more.

Scott Walter · 00:38:48

First thing we want to do is that we’re talking about a lot of different hands and a lot of different hand companies. And Allonic is like the only one we’ll be talking about, which is actually in the RoboStrategy portfolio. And so I just want to put that out there, full disclosure. This is a RoboStrategy company and it’s a very unique design. As you can see, it’s a tendon-based design, but it’s a little bit different than your standard tendon-based designs. And that is that almost everyone is trying to figure out how to build a joint mechanically.

Scott Walter · 00:39:19

And they make a little tiny door hinge and they’ll have a tendon going on there that’s pulling it to open and close. And a lot of people have been studying different ways of being able to do that including Tesla was trying to come up with a novel way that they seem to have abandoned and gone back to the same thing that everyone’s just going to make a simple pin joint with a tendon there. Now, if you look at the human body, there’s only really one joint, which we can see a true mechanical analog to, and that is like our shoulder joint and our hip joints, which is a ball and socket joint.

Scott Walter · 00:39:51

Look at it and go, oh, that’s just, just like saying. But we’re all familiar with a door hinge. You look at the human body and I think even biology in general, I don’t think you will actually find biology having coming up with what you consider a pure door hinge. But there’s an analog in the way they work and that they have what’s called a rolling contact joint. They move around like that and the constraints are such that you can emulate it very easily with a door hinge or what we would call a pin joint or a cylindrical joint, just like that.

Scott Walter · 00:40:18

But the reality is it functions very different. So, again, when we go to degrees of freedom, I could argue that your IP joint has six degrees of freedom. You’d be huh, what? Well, the dominant one is in this direction, but it does have a little bit of wiggle in a bunch of different directions. And our body tries really hard to make sure there’s only one dominant direction. And a door hinge does a very good job of like eliminating all the other kind of motions, though it still has a little bit of slop depending on how tight it is there.

Scott Walter · 00:40:50

But the thing is, the door hinges mechanically lock it in to make sure that doesn’t move around. Now, if I have two things that are just doing rolling contact like that how do I keep it together? How do I keep it from falling apart? And the way the human body does it is we have these things called ligaments. So basically just imagine you get a bunch of rubber bands that are holding it on in there that will allow it to stretch. But you get the rubber bands in such a way that if you try doing that it just pops you right back in there.

Scott Walter · 00:41:17

But it’s a line that will allow you to do that very easily. And then you use tendons to do the actuation. So, what’s the difference between a ligament and a tendon? A tendon is basically muscle to bone attachment. And a ligament is basically bone to bone. And so you have that in there and they’re allowed to stretch a little bit. And what’s kind of great is that they act as spring sometimes. It’ll help pop everything in there. So, the human body has figured out how to do it in a miraculous way. And there’s also a little bit of a saddle joint in there to help keep a little bit of movement this way.

Scott Walter · 00:41:48

So, what Allonic has been trying to do is solve a couple of problems. We talked about tendon wear, right? The tendons wear out. And they fray all the time. Well, wait a minute. It’s like: us humans, we’re tendons. And we don’t seem to have that problem. And it’s like: we actually do because we have this thing called sleep. And when we sleep at night, our body is like regenerating and fixing everything that’s damaged. So, if you build like a tendon-based hand, you’re going to have a maintenance schedule. Every now and then, you’re going to have to replace the tendons.

Scott Walter · 00:42:20

Now, we don’t replace the tendons. We are able to miraculously repair the tendons. And if you really want to do the comparison, we want robots to go 24-7. And if we say, ah, the robots can’t go 24-7. The tendons are everywhere. We can go 24-7. Like, no, we don’t. It’s like: we work an eight-hour shift, and we’re exhausted. And if we continue to work eight-hour shifts nonstop, our tendons will fray and break. They need a break. And sometimes, they need a weekend or more. And if you’re not careful, you need surgery. And sometimes, you’ll have these cases that they never solve, that will end up having something with their carpal tunnel syndrome or something like that.

Scott Walter · 00:43:03

Or if they’ve ever had tendonitis. No, it takes a while. So we also have to be very careful how we play with the use of our tendons to keep everything nominal. So think about, a surgeon is known for his hands, right? That if they’re smart as he is, it’s all about his hands. And surgeons are not bricklayers. They will not do that. Because if they go out and start becoming bricklayers, they will lose their surgical skills in a way because they are stressing their mechanical components in a way that is just not going to work long term.

Scott Walter · 00:43:37

Okay, so what’s kind of the idea here with Allonic? Well, rather than repairing your hand, what do you imagine if you just dispose it and replace it again? Because a lot of times, it’s a question of cost. So when we are doing any sort of work, we put on consumables, gloves, to protect our hands. And then at some point that we’re out and we dispose of them, we put them on there because it’s cheaper to do that than expect our skin to repair itself all the time. And so, a lot of times the cost in a hand is where the actuation is.

Scott Walter · 00:44:07

It may not necessarily be out there. So the big problem with most hands is, you want to have them be dexterous, you want to have them be robust, and you want to have them be cheap. And it’s really hard to have all three. So the idea of Allonic is to go ahead and actually make them cheap by using traditional braiding technology. Now, we take our textiles for granted, but these machines are amazing. What they can do when it comes to any sort of textile, whatever they have come up with, that can just weave things so quickly in any sort of design.

Scott Walter · 00:44:41

We’ve seen embroidery and stuff like that. These machines can just do, right out there. And they recognize the same thing. It’s like: by using braiding technology that’s usually used to make ropes, they can also make fingers. And that part of it is that they do have a stiff member in there, which is a simple 3D-printed bone. So you feed the bone into the braiding machine, put these things around there, and you can determine from the pattern, and you can put in different fibers. So you can have, thicker or stronger or different kinds of fibers woven in there.

Scott Walter · 00:45:13

In some cases, you’ll see they do have, a different color fiber in there just to kind of prove you can do that. And that means you can start weaving all the complexity that would normally be very hard to assemble mechanically. And so it’s like: oh, certain kind of tendons, we can go ahead and do that. Oh, we need a couple ligaments here. Oh, I’ll put them in. What material do you want? So it doesn’t have to all be Dyneema. It could be cotton threads. It could be nylon threads. Whatever you want, which will give you the strength you want, give you the capability, the compliance you want in some places.

Scott Walter · 00:45:44

And the other thing, the difference between this hand and all the other hands I’ve had is that it feels kind of warm and comfortable. It’s soft. Everyone else’s hands are metallic or they’re plastic. And in many cases, their compliance they build in by putting a layer over it. So you can put a glove on it that makes your metal hand feel softer, but it still feels a little bit metallic. And so this, when it comes to actually coming into contact with people, will feel more human-like than many of those designs. So that’s the other thing that’s, super attractive about it and that they could just knock these things out really quickly.

Scott Walter · 00:46:22

And as a result of being able to do that you don’t care about the fact that your hand wore out. You just replace it because it now becomes a consumable that’s on the order of magnitude of maybe the normal kind of consumables you have.

00:46:37 — Allonic — cost, maintenance, and control

Michelle Sun · 00:46:37

So that’s definitely a very different approach. And you mentioned that there’s this trade-off triangle of robustness and dexterity. What’s the third? The cost, right? So this one, what is the BOM cost for this? And what do they sell it for?

Scott Walter · 00:46:57

So the BOM cost will probably be in the order of $100. So when they finally get the scale, this is going to be remarkably cheap. Who knows? It could be less than that. You’d love to really get it down the cost of gloves. I know some gloves are pennies, right? It takes gloves people throw on. There are other gloves in the order of a couple of dollars. So it’ll still be a little bit more. But in the grand scheme of things, if you can get a couple of weeks of wear out of this, it doesn’t matter.

Scott Walter · 00:47:23

And it depends on your application. It could be that this is going to last very, very long. But if you put it in a very dirty environment, or something where you’re very, very aggressive with it, it could be something that wears out very quickly. But at the same time, if you’re in an environment that has a lot of grit and everything, chances are the protective gear that the people are wearing is also very expensive. And it’s just so different and unique of everything else. There’s still challenges: how do you control this thing? Because it’s tendon-based.

Scott Walter · 00:47:54

But everyone is sort of getting over and around how to figure to do it. Because with enough ML, I think you can solve these problems. Everyone’s trying to simulate how their tendons are going to work. And figuring out everything to get a heuristic kind of control of it. And my feeling is, I think if you just start collecting enough data, enough ML data, your neural net will just know what to do. Just like we do. We know exactly how to manipulate our hands without even consciously thinking about it. And we’re not simulating or trying to model this with mathematics.

Scott Walter · 00:48:34

So sometimes engineers overthink it. And this is what I’ve always seen as the power of ML. If we just get enough episodes and just train it and everything, that will be the solution rather than saying, I have to have the super high-fidelity model to be able to control it. And Allonic, they’ve come out of stealth. They’re still in the startup phase. They have a lot of samples. We’re kind of bullish on them for that particular reason. But again, full disclosure, I might be a little bit biased here. Take everything with a grain of salt.

00:49:03 — Different designs for different applications

Michelle Sun · 00:49:03

Yeah, thank you. So this is a really great first look at a very different design. And one of the things that also stood out to me is that we may need different designs for different use cases, right? So for Allonic, it’s disposable, almost like a glove, high-use environment. And it may be for scenarios that are not really precise. Precision is not super needed. It’s like: oh, picking up things and things like pushing the cart around. And so maybe that tendon-based sim-to-real gap is not as much of a deployment blockage.

Scott Walter · 00:49:46

And just to point out, I’ve been looking at so many different ones, and everyone asks, well, which side of the fence are you on? And I’ve jokingly called you’ve got the tendonistas on one side and the drive cells on the other. Where am I? And I’m like: I think the TAM is so big and the application space is so big, they’re all going to win in particular niches. They will kind of self-select to the applications that they’re best suited for, and they will thrive in that very well. Because we have yet to figure out the hand that rules all hands.

Scott Walter · 00:50:17

We haven’t collected all the infinity stones yet when it comes to being able to come up with that. So you’re absolutely right, Michelle, that you’re going to see that tendon-based hands just make sense in certain applications and direct drive and other stuff like that. And the two sides will constantly argue over who has the best. But in the end, the customers will decide and they’ll find the applications and they’ll go in there and they’ll be very well. So that’s why I’m kind of bullish on that whole sector in a way is that they all will be able to find traction where it’s needed.

Scott Walter · 00:50:48

And then everyone’s like: oh, that’s a cop-out. It’s like: come on, choose one or the other. And it’s like: no, the reality is I think you don’t have to choose one. You only have to choose one if you decide to build your own hand. And then it turns out there’s a bunch of companies over in China that have like four or five different versions of hands that are all of the above because they also recognize the same thing. And then there are the others that are just like doubling down this is the only way forward and we want to put all our effort in.

00:51:15 — Daxo — up to 120 tendons

Scott Walter · 00:51:15

So the Daxo is really cool. Everyone argues over tendons and how many you need. And Tom Zhang is like: the more the merrier. In this case, I think he’s got up to 120 tendons. And then he thinks the more tendons makes it actually better. And he talks about it being more like infinite degrees of freedom, actually tendons and motors. So he actually has that many motors. They’re very small motors and he’s able to actually fit them in a space that’s about the size of a little bit bigger, kind of Popeye, but he’s able to get the form factor down there.

Scott Walter · 00:51:58

And he’s got so many tendons that he’s got a completely different control algorithm. Now, this was his first generation hand. And as you can see, it’s extremely flexible. I call it almost like sausage hands that there’s no clear kind of IP joint or it has way more IP joints than we’re used to. And that means as far as gestures and everything can do, it can do way more than a human hand can. And that was his first version. His second version, he actually has something that looks more like a human hand as far as IP joints going in there.

Scott Walter · 00:52:32

And Tom knows a lot about human anatomy and stuff like that. So he’s taken biomechanics. He’s also had like a startup that was working on a type of mechanical human hand. So he understands everything very well in the traditional approaches and was trying this other approach. And now he has sort of the normal number of joints. But again, he’s completely overactuated as far as the number of tendons. And what could be the potential advantage? Why would you even want to do that? Well, the human body actually has more tendons than everyone else is doing. So we think of that we have like a flexure and an extensure.

Scott Walter · 00:53:09

So Tesla did not have an extensor at first. They use a spring return. Now they’re putting one in there and they have two abductors in there. They were at first three. Now it looks like they’re going to be four and many others are four tendons per finger. Well, we actually have more than that in human anatomy. We also have these intrinsics internally that help us with the subtle control and movement. But we have more than one flexure anyways for grip strength and everything else. So looking at that well, wait a minute, maybe the more the merrier.

Scott Walter · 00:53:37

Because the more tendons you have, the more grip strength you can get when you need it. Because rather than having one actuator trying to pull on that finger, I can have multiple pulling on that. The other thing is that you have way more kind of weird combinations. So what a lot of people have been trying to do with the abduction is basically keep their cake and eat it too. It’s like they have two tendons because you need the two tendons for abduction. One pulling this way and one pulling that way. And then they’re running kind of a flexure through here.

Scott Walter · 00:54:10

But rather than having that flexure move this joint, they route it in such a way that it only does the PIP and the DIP. And what ends up happening is that if you pull both abductors at the same time, you don’t get ab or adduction. It actually forces you to get flexion. And that’s like a real clever thing. It’s like: oh, look, we can get this two for one here. And if I only pull on one, I get this and that. The problem with that is you actually lose a little bit of control of where it is, which is why you actually need an active extensor back there to give you the antagonism you need to put it in there.

Scott Walter · 00:54:40

And the way I want you to understand what’s happening here is that a lot of people are familiar with horseback riding. And you have reins on a horse and those leather reins are just like tendons. They pull, they work in one direction, but when you relax it, you can’t push it, right? You can kind of put it out there, but the horse is not going to feel that it’s only going to feel tension this way or that way. And we all know to make the horse turn its head this way or that way, we just have to pull on the rein this way or that way and how to return it.

Scott Walter · 00:55:09

And I want you to think of that just as abduction back and forth, back and forth. You can do that.

Michelle Sun · 00:55:14

So it’s like a pulley—there’s two.

Scott Walter · 00:55:19

It’s a pulley mechanism. Exactly. But only one of them actually controls and the other you just release to make sure you’re not putting the tension back. Now, if I pull on both reins at the same time, the horse’s head can’t go left and right. There’s only one thing you can do and that is it has to come back up. So that would be like getting in flexion. So you look at that and say, look, I get all the control I want. I can put the horse’s head wherever I want. If the horse is cooperating, what if the horse is like upset and it’s just like moving its head around?

Scott Walter · 00:55:49

Suddenly you realize you don’t have degrees of control to keep it where you want to because it’s fighting you. It’s bucking against everything you’re doing. And so what’s that got to do with fingers? Well, what happens when fingers come in contact with something? As soon as it comes in contact, I have to resist that. Otherwise, my finger is going to go. I want my finger to stay there. Well, as soon as I do that and start fighting that thing, suddenly I am inducing another movement somewhere that’s very hard for me to control because with just horse’s reins, I don’t have enough degrees of control.

Scott Walter · 00:56:17

You would have to actually add something else in there to be able to do that to give you that kind of stability. And these are the problems with minimizing the number you have. Now, if you go overboard and you put tons of them on there, suddenly not only can you get extra strength you want, you can get the stiffness when you need to have, and also different degrees of control because he’s got enough going to each group that he can move each pinky or each IP joint independent of another, which is something that’s very difficult to do when you have a minimum number in there.

Scott Walter · 00:56:49

It’s like you get that coupling that’s very hard to do. Now, most people can’t do it. Common’s an unusual person that I keep on saying, yeah, but people, can’t, you get this coupling and suddenly he’ll show me he can do it. I’m like: I don’t know how he does it, but he’s got like control of his fingers in a way that most people don’t.

00:57:07 — Daxo — grip strength, sensing, and proprioception

Michelle Sun · 00:57:07

But what kind of use case do you think this unlocks? Like this kind of hand can do that other hands that are less actuated cannot do?

Scott Walter · 00:57:16

Okay. Potentially, it could give you a lot more grip strength, which is the problem that everyone talks about is ah, we’re kind of lacking the grip strength, kind of dexterous control, resistance to force. But the other thing that’s rather interesting in here, and this is where it comes to ML, is that a lot of fingers, we’re trying to figure out how to get tactile sensing in there. Because this is something we haven’t really talked about is we’ve got all these movements. What about the tactile? It’s like: oh, that’s another episode. But it would be nice to kind of know where the touch is.

Scott Walter · 00:57:47

And there’s ways of doing haptics or sensing. One of them is that we actually have like the pressure sensors that we can feel something, which is amazing. But we also have this other kind of sense of where our fingers are, which is known as like proprioception. And we can also feel pressure and resistance that isn’t necessarily being felt through our fingertips, but we can kind of feel it in our forearms. So a lot of times, you’ll notice if you push on something like that you might notice that something in your forearm is tensioning up a little bit.

Scott Walter · 00:58:18

You can feel the touch there. So if you were to actually numb your fingers so that you can’t feel anything, you will notice you can still tell when you come into contact with something because that your forearm goes into tension. So that means there’s data out there about the contact you’re coming within. And it’s very, low quality signal, but it’s still signal and information. The other thing that all the tendon-based hands have, and this is a problem with the Tesla is because if you kind of go back to the Tesla image, you’ll notice there was like a lot of blinking lights and other things that were in there.

Scott Walter · 00:58:50

And you might want to know, what’s that all about? You see those green lights in there? Well, the problem again with a tendon-based hand is as it moves and comes into contact with something, I now have no idea what the angle of my joints are. With the direct drive guys, they know exactly. They can basically count revolutions on their fingers and they know where they are. They don’t actually need to have a decoder. I think they may have some built in there for some kind of redundancy. But the tendon-based ones, it’s like you’ve got no choice.

Scott Walter · 00:59:18

If you want to know where it is, you need some sort of feedback to be able to do that, which means ideally you’d like to do it just by measuring the tendon length or how much your actuator down here rotated. But the reality is, that’s kind of a guess on where it is, but the reality is it might be somewhere else and you want to know how to compensate it. Well, one way is I guess I just kind of look at it and estimate the angle. But in many cases, they need to actually have sensors there.

Scott Walter · 00:59:44

And a lot of those LEDs you’re seeing in there is because they have encoders built in there. And that’s basically the signal that lets it’s there and operating. I’m assuming the full-blown one won’t have that. So suddenly you’re having to build additional electronics. And like that means not only are you routing tendons, you’re routing wires through there and all these other things that kind of know what’s going on. It’s like: oh, so now I’ve got additional stuff that has to go there. So what Tom’s doing here is that he is able to tell that when you come in contact with something, that all the tendons, it’s basically kind of this idea of a tensegrity, that all of them suddenly feel attention and are reacting to that structure.

Scott Walter · 01:00:27

And the reaction to that structure, he’s feeling down here in the forearm because he has 120 motors down there. And those motors are set up that they have extremely good torque transparency. And he’s able to say, oh, this tendon, that tendon, they’re all feeling a certain force on here. And he runs this experiment enough times. Now, if you ask me to sit down and write the algorithm for that I would have absolutely no idea how to model. It’s a very difficult thing to model. Tom’s like: oh, we just run enough experiments and we know. And so he’s got it to the point that he can come up and he can touch his finger somewhere.

Scott Walter · 01:01:01

And he knows exactly where the point of contact is based on everything that’s down there because he has so much overactuated everything. So his philosophy there is that there are so many other things we can gain from doing it. Strength, cool proprioception that’s better than anyone else.

01:01:17 — Tacta — measuring how workers use their fingers

Michelle Sun · 01:01:17

Very cool. And I love that Daxo is taking the approach of maximalism, right? Let’s get as many actuators in there, as many tendons in there. And then the third example that we have is actually the opposite approach: let’s think about whether we even need five fingers. In fact, a lot of people that I’ve talked to talked about how many DOFs are really needed to handle like 90% of the tasks that humans actually do. It’s actually way less than 22 DOFs. So it seems like Tacta is an example that we can talk about also where they have five fingers, but then their launch actually showed.

Scott Walter · 01:02:03

They went along with that assumption. And again, there’s this huge academic debate over how many fingers you need and stuff like that. A lot of people like to weigh in on X about this all the time and say, well, I can do these things with three fingers and why do I need four and everything else? And again, there have been some academic studies showing that and most of the pushback they see from people is you can do a lot with three fingers and you kind of start getting out to your parade or frontiers that there’s still others that you need to have the other, have it.

Scott Walter · 01:02:36

So a lot of times it is very academic, but what Tacta did is they said let’s settle this debate and actually start measuring data with five finger gloves and they measured it from real workers in doing real tasks. So not some lab where they’re pretending to do stuff, but they actually suited up real people with a lot of tasks and they were surprised themselves to find out that the last two digits seem to be going along for the ride most of the time. And that meant that they believe that they can take over everything with just a three-fingered hand, because the others aren’t needed and they won’t have to worry too much about the body gap.

Scott Walter · 01:03:14

Now, some people say, ah, you really can’t do that. And part of the reason why we have five fingers is fatigue is that in order to generate the grip strength and everything, a lot of times we need those five finger hands. So many cases, the five fingers come in there more for a reason out of fatigue, not necessarily dexterity. And they realize this is our fingers don’t get tired and we can apply a little bit more force. So that’s why they’re able to say that they can get away without the index finger and without the pinky for that reason.

01:03:44 — Tacta — fluidic actuation and industrial arms

Scott Walter · 01:03:44

Now, Tacta is a lot of things. So the first thing is, I think they were kind of sitting there looking at the arguments and the mudslinging and everything going on between the two sides of people who are pro tendons and people who pro direct drive. And they’re like: maybe there’s a third way, a different way of being able to do the actuation. There’s kind of a hybrid between the two and they use something called fluidic tendons. And it’s a little bit like hydraulics and hydraulics have been used before. So, for instance, Clone is using hydraulics.

Scott Walter · 01:04:14

Sanctuary has done that. And the big question is: what’s the difference between hydraulics and fluidic? Well, usually with hydraulics you have one pressure source that creates a pressure and you have these micro valves that are able to open and close very quickly to be able to sort of let the hydraulic fluid in or out. And you can almost think of those as tiny actuators, but it’s more or less like open and close and figuring out how to do that really rapidly. So that’s been the trick that both Clone and Sanctuary, they have patents for things like that.

Scott Walter · 01:04:44

The other is that rather than having one single reservoir with these things that open up, why don’t you just have a syringe mechanism? So we all know how a syringe works, being able to bring up a fluid or to push it out. And so you just have a fluid and you have that actuated and you just push on it as necessary. And it also gives you a lot of control.

Michelle Sun · 01:05:02

Are the fluids currently in the palm area or where do you see it right now in this diagram? Well, in the forum.

Scott Walter · 01:05:10

Okay. So it is backed out somewhere else. So there’s a much bigger base. It’s not even in the forum, it’s in a base, but they’re going to be scaling it down. So right now it’s kind of bulky and it’s really meant to go onto a cobot. So this is the other thing. A lot of people talk about the humanoid distraction when you’re building a hand. A lot of people are trying to build a hand to go on the humanoid. And when you do that you have to do all these things. It constrains you so much on the size and the power and the weight and everything else.

Scott Walter · 01:05:36

But there’s all these robotic applications out there that just need a hand. They don’t have to be in a humanoid form. So we’re like cobots, industrial arms and stuff like that. There’s a lot of applications. If they had something like a human hand would be able to do a lot of it. And they’re not the only ones that have noticed that. Mimic has noticed that. Kyber has noticed that you can go down the list of some others. Like, wait a minute, let’s not worry about this distraction, the humanoid distraction and think about what would happen if we put it on a real robotic arm, which is already certified.

Scott Walter · 01:06:07

They have great precision. They already have a programmatic interface. We can source them really easy. Let’s do that. So that’s how they’re looking at that. And they’re not worried about how this is going to get onto an Optimus. They’re not trying to sell it to Optimus or to Tesla or to Figure or anyone else. They’re like: no, this is going to be going to other people. And they’re trying to solve that at hand problem. And the way they’ve done it is basically a kind of a cross between trying to be a bit direct drive and also be tendon-based.

Scott Walter · 01:06:37

And that instead of using tendons, they’re using a fluid to replace the tendons. And they’re using a tube, which you could argue is kind of like a Bowden tube in a way. So a lot of the routing is again done with these tube structures that will be able to go around but you can then have a wrist without the problem. Because again, if you’re having a tube going around there, you don’t have to worry about the length changing. And the fact that they have a fluid in there, the fluid doesn’t really care about length. It just cares about pressure.

Scott Walter · 01:07:03

And you can sense the pressure really quickly and react to that. So a lot of very interesting things about this hand design are using a different way of doing actuation.

01:07:12 — Tacta — gloves, data collection, and tactile sensing

Scott Walter · 01:07:12

But as you’ll see, Tacta is like kind of many companies in a way because it’s not just the hand and a few others are beginning to realize you need what literally is like hand in glove, right? You need to capture data. And there’s always this embodiment gap that everyone’s trying to figure out. Well, now how do I collect the data? Do I go egoistic, this and that and everything else. And a lot of times that’s not good enough. Do I go with an UMI glove? How do I do the mapping or retargeting from what my hands like?

Scott Walter · 01:07:42

Wouldn’t it be great if I could just slip this hand onto my hand and collect data with it? That’s the ultimate embodiment and or the ultimate kind of UMI. And they have a glove, which is very similar to the hand as far as the mechanical movement. And as far as the touch sensing, because that’s the one mode everyone says like with egocentric data, you can capture the movements, but you can’t capture the actual tactile feel.

Michelle Sun · 01:08:10

Is that the glove that you’re referring to on the screen?

Scott Walter · 01:08:13

Yep. So they have a glove, which is able to pick it up. And when they’re doing that they are also capturing what the tip sensors are going to feel. Now, the sensors they have all over, they came up with a way of being able to mass produce very high precision and high density taxels. So these are basically tactile cells. Think of them like the way we think of pixels or voxels, except for touch. And they have them in all the key locations, including the palm. So they know what it’s like to grab something. And I’m pretty much of that feeling that the data collection we have to do has to have a lot of modes and that a lot of the data, if it doesn’t have some sort of touch or force feedback, the data will be a very low quality.

Scott Walter · 01:09:03

It’s still usable. But what it means is you probably need a much larger data sets to be able to overcome that deficit. But if you can actually capture that mode right away, you’re way ahead of the game, way ahead. So they already were the philosophy. It’s like: we’re not just going to build a hand. We have to build the data collection thing to go with it. And we’re not just building a hand. We’re also building really good touch sensors that are going to be in there. So that’s some other part of the IP. That’s very interesting approach.

Scott Walter · 01:09:30

So it’s like the full stack of the hand in a way. And they’re also having to build the model that goes on top of it. So very serious company based in Palo Alto, actually just down the road from where the Tesla headquarters is. And when I visited them last week, I was really quite impressed with the operation that they have there. The number of engineers that they have, the equipment that they have in there is just wow, these guys, they kind of know what they’re doing.

01:10:00 — Tactile response rates and sensitivity

Michelle Sun · 01:10:00

Yeah. And one of the things that you talked about, the touch sensor part, I read that they have 400 hertz, right? So I think a lot of the numbers that I’m seeing is more than like the 50, 100 hertz. What does the 400 hertz unlock versus a lower number?

Scott Walter · 01:10:18

Yeah, that’s very important. We always hear these numbers that humans are able to react in like a 10th of a second. And I know that from like track and field, because they always said that it takes a 10th of a second for the runner to react to the sound of the gun. And that’s how they actually tell a false start, believe it or not, is that when the gun goes off, if the pressure that’s on there happens before a 10th of a second, it’s considered a false start, which is rather amazing, which means you cannot anticipate.

Scott Walter · 01:10:47

So it’s kind of based on human reflexes. And we always think about that. And a lot of times the update rates, I think our vision systems update also somewhere around 30 hertz, maybe it might be technically 20, but I think it’s considered somewhere around there. So everyone’s always built to that. And the assumption is that’s human reaction speed. But it turns out our touch sensing is incredibly reactive and it’s closer to a thousand hertz. So we’re getting a lot of information really quick. And it seems a lot of that is not so much going all the way back up to our brain, but it’s more like reflex action that we have.

Scott Walter · 01:11:23

So we have a lot of kind of edge compute or neural compute built into our hands and our reflexes. So you notice if you touch something sharp, you react way before you’re aware of it or something hot, and that’s because we are able to sense very quickly and they’re trying to get it down to that human scale level. The other thing they’ve done is that they can actually sense pressures that are a quarter of what humans can sense. So it’s even way more sensitive as well as way up on the high end. And as with a lot of instrumentation, either an instrument is really good at something fine or it’s really good at something big and it’s hard to get the two to be able to do both and they’re able to change.

Scott Walter · 01:12:05

That’s impressive. So they’re trying to get down to those levels. And that was new information to me, that there’s actually part of the human nervous system that is able to react faster than 100 hertz in the order of a kilohertz is amazing.

01:12:19 — Industrial use cases and customer discovery

Michelle Sun · 01:12:19

And one of the things that I’m trying to reconcile is that it’s really great that Tacta is tackling the hand space in a more deployment-centric view. Right. So instead of thinking to sell to humanoids, actually, why not? We just build a habit for the COBOS systems and go to the industrial setting. So in terms of this very fast tactile response and also sensing different pressure in a very precise way, what is the use case in an industrial setting? I get that at a home setting, if I touch something really hot, then I want to react.

Michelle Sun · 01:13:01

But in an industrial setting, what are some use cases that will be very useful?

Scott Walter · 01:13:11

They’re all sorts. The one use case I see coming up again and again and again is data centers and cabling and tables and data centers. And like they’ve probably had to plug in like an internet cable in the back of a laptop or something like that. Data centers are like that but on steroids and slightly different kinds of connectors. And a lot of times knowing when a connector is getting it in requires sometimes a lot of finesse because sometimes you can’t see it. So you have to be able to kind of react, tell from the field.

Scott Walter · 01:13:39

And then when it clicks in, know that you’ve got the right kind of click. So we are very good at doing that. And with the scale up of data centers, everyone is looking at trying to automate that. So that’s like kind of one example. You also have just like a lot of assembly, like consumer electronics and stuff like that. The finesse that kind of goes on. And then it turns out, see, this is the other thing that Tacta is able to do with their fluidic gloves. They can actually make smaller size hands. Most of the other hands, they’re bigger than most human hands.

Scott Walter · 01:14:11

So when I put my hand up there, it’s like the size of my hand. A lot of people complain about the size of my hand and higher. And in China, in a lot of electronics assemblies, most of the people who are working there are women. And a lot of times they’re there because they’re petite and their hands are much smaller at that scale that actually allows them to do it. I would have difficult time because my hands are both thumbs, kind of what it comes down to. And so they can actually scale it down to the size that will allow them to do that.

Scott Walter · 01:14:40

A lot of these very precise motions that are just very hard to automate in general. And again, what they’re seeing is that a lot of these assembly plants are people sitting at their desk. And all they’re doing is moving their arms. You don’t need the legs and you barely need a torso, but you need this fine dexterity control that’s out there. And the equipment is already there. It’s like: oh, they grab this tool. They go ahead and do that. They grab something else and be able to go. So they’re finding that. And this is what I encourage all entrepreneurs out there.

Scott Walter · 01:15:11

And I’m seeing a lot of them that are more like that. Actually get out to the deployment phase early. I would say even pre-deployment. Going out, meet your customers, find out what their needs are, and then design everything around that. And you will definitely be more successful way sooner than everyone else. If you’re sitting in a lab, everything looks great. And then you go to deploy. And then suddenly you’re discovering all sorts of things you did not know were going to be problems. And now you have to go back and redo it. So find your customers.

Scott Walter · 01:15:42

I’ve met a couple entrepreneurs, and it’s just great. They literally just go on a field trip. Someone will just pack up the car and drive around and just knock on doors of all these different like small companies to find out what they’re doing, what their manufacturing challenges are. And some cases find, the manager that’s there or the owner being really excited about, yeah, we’ve been wanting to automate this thing over here. These are the challenges, the problems, and the guys learn about it. And then they go back and they build something and then they get traction.

Scott Walter · 01:16:09

That’s the way it’s going to be. The TAM is so big. Don’t worry about building hands for humanoids. I am very pro-humanoid. It’s going to be there. But I don’t take the side that it’s going to be special purpose versus general purpose. There’s a huge spectrum in between there and that you’re going to need both because there’s just so many applications out there. And if you try to get kind of dogmatic that only one side is going to work and you think you’re going to dominate the whole thing, no, no, no, stay full.

Scott Walter · 01:16:44

That is the biggest pie we’ve ever seen, ever, trying to automate labor. And the tiniest sliver of that pie will keep you satisfied for your entire life. You won’t be able to eat that piece of pie. Let’s just put it that way.

01:17:04 — General-purpose bodies, specialized work

Michelle Sun · 01:17:04

Exactly. And if we think about humans, right, we’re general purpose, but we also have different professions. So a surgeon is different from an assembly worker. The skills, they’re all precision-type skills, but it’s a very different type of expertise as well. And the part that you mentioned really going out to the field and the deployment site, like it was really so true where when I visited factories in Shenzhen, I’m like: wow, these people, they literally sit there for 12 hours a day and they like have a basket on top of them. And then they move things and assemble and then move it to another basket.

Michelle Sun · 01:17:39

And so they really only get up to use the bathroom or get a sip of water. And so I think these are the insights that makers like Tacta would think about: hey, if we really want to be useful, maybe put the energy to make a really sensitive tactile sensing system. And then a very dexterous hand, and maybe think about how the movement is not as important as sometimes we want to think of.

Scott Walter · 01:18:05

Yeah. And the other thing to kind of point out is, this argument about, one humanoid can do everything. It’s like kind of the humanoid form factor can, but we’re going to see different versions of it for good reasons. And that is, we know, they’re different sized people and some of them excel in certain areas and others just because they’re a little bit taller, they’re a little bit smaller. Look at the Olympic Games. Gymnasts tend to be on the short side. Basketball players tend to be on the tall side. And a lot of it’s just physics.

Scott Walter · 01:18:41

If you’re tall, it’s really hard to do cartwheels. I can tell you that I cannot do a cartwheel. Okay. I am jealous of all these people I’ve seen doing cartwheels. So you’ll see these different embodiments make sense. And a lot of times it’s just that again, a surgeon kind of follows a certain path and that it may be that because they decided to not be a bricklayer, they have the hands to be a surgeon. But if they had made that choice earlier in life, they wouldn’t be able to be a surgeon because they’ve just kind of lost that finesse that you need.

01:19:17 — Why the next five years might be different

Michelle Sun · 01:19:17

Yeah. I think one last question that I have is, people have been saying that robotics is five years away for the past almost 30 years. And so what’s on the table now that wasn’t available 10 years ago that you think will truly unlock the next five years being that inflection point? Is that the edge compute side or is it more like the sheer amount of data and maybe it’s a bit of both?

Scott Walter · 01:19:54

Well, I think the most obvious thing is just machine learning and AI has changed everything dramatically. So you can do a lot more. First of all, the fact that we can just use that to control robot armors versus the old way, which was pretty much just that it would do the same thing. It didn’t have any intelligence to adapt to any changes in the environment. So that’s just opened up so many possibilities. Now, the reason why I say that’s important: there’s all these other things that are important. The other is just the interest in the field.

Scott Walter · 01:20:29

Ten years ago, no one was really interested in robotics. Now, that’s all everyone wants to do. And the more minds, especially brilliant minds, you start throwing out a problem like this, the more likely you’re going to solve it. The other has been the result of a lot of improvements in mechanics. And a lot of it comes because people are trying to solve the problem now. They’re trying to solve problems that could have maybe been solved before, but they weren’t interested in. So, making smaller actuators and stuff like that. We know how to make motors for the longest time, but no one really thought of making motors down to that scale that you could put in humanoid.

Scott Walter · 01:20:59

So, you saw Elon complain about that five years ago. That you could not find any actuators on the market for any price to do what you want. Now you can. It’s incredible. You go to Amazon, you can order and you get them a couple of days later. So there’s that. There’s the fact that locomotion is a solved problem, whereas like 10 years ago, it was still very challenging. So all these things are happening: this convergence of technology, new material sciences, new people trying to solve all those things. So what’s happened is that there’s just the excitement.

Scott Walter · 01:21:35

Maybe it’s just this big paradigm shift and the realization that humanoids are possible. Physical intelligence is possible. And this is something I’ve said before people. It’s like I’m a broken record on this. But one of my favorite quotes is from Arthur C. Clarke, where he says the space elevator will be built 50 years after everyone stops laughing. And I’ve paraphrased that to humanoids. Humanoids will be deployed like five years after everyone stops laughing. And I think everyone has kind of stopped laughing now. So that five-year we’re talking about—we’re seeing them already. Come on. They can do some manipulation tasks.

Scott Walter · 01:22:17

They’re just not quite reliable yet. They are going into some deployments. They are doing some of the low hanging fruit. And they’re the worst they’re ever going to be. They’re just going to get better and better and better. And five years from now, that’s a lot of time in humanoid space. Really, when I go back thinking about when we saw the Unitree bot doing the RL program walking, I think that’s ages ago. And I realized that was only 18 months ago. It was like January 2025. And so it’s like: oh, I thought that was, again, in my head, it’s like five.

Scott Walter · 01:22:51

But so what we’re seeing, and again, if we look at the humanoid games and also the mini marathon, the Beijing mini marathon, the contrast in 12 months is just unbelievable. And you’re going to see like that same step change next year in both of those. I’m really, really pretty sure you’re going to see it. So five years, yeah, we’ll be there. We’ll be there. And humanoids will be quite performative at that point.

• What task does the robot perform?

• Which part of the system is holding it back?

• What evidence proves it works repeatedly at the required economics?

I built the task-first Physical AI stack as my framework for answering these questions.

It maps 11 functional layers, from the part that touches the work, through actuation, power, intelligence. It can be used to locate a bottleneck in a system, to understand where a company competes, and to follow where the value accrues.

What is the physical AI stack

The physical AI stack is the set of hardware, software, data and operating systems that turns a goal into a reliable action in the physical world.

Most technology stacks are drawn from the enabling technology outward: chips, compute, models, applications. For robots, I find it more intuitive to begin with the task. I call this the task-first stack.

Start with a job such as placing a product in a bin. A gripper touches the product, an actuator moves the gripper, the power system supplies the required energy, the control software regulates the motion, planning finds a feasible path, perception locates the object and bin, a policy selects the action, and data trains the model.

There’s also compute, simulation and evaluation, which we’ll cover in this piece.

These 11 layers map the working system behind a robot task. Beneath them is the industrial foundation: energy, critical materials, manufacturing capacity and physical infrastructure.

The task-first stack changes the starting question. Instead of asking how advanced a model has to be, ask what has to go right for the task to succeed. The answer then points us to where the bottleneck is.

One task through all 11 layers

Recently Stanford student and OpenAI Robotics intern Thijs Simonian gave GPT-6 Astra an SO-101 robot arm, a paintbrush and a camera, then asked it to paint the Golden Gate Bridge. Astra planned one minute of actions at a time, reviewed the result and adjusted across attempts with feedback.

This is a good example of how the stack works and identifies the bottleneck. We will also go over 3 additional examples at the end of the piece.

For this Golden Gate Bridge painting task, the bottleneck is the control and actuation interface between Astra and the robot. The model (GPT-6 Astra) could recognize what needed improvement turn after turn, but the robot’s control interface, actuation and limited feedback limited how precisely the arm could execute the model’s judgment.

Don’t worry if the table doesn’t make sense to you right now. We’ll walk down the stack one layer at a time. After that, refer back to the table and it will make more sense.

1. What is a robot end effector? Task meets the world

A robot’s end effector is the part that touches the work. Common forms are a parallel-jaw gripper, vacuum cup, magnetic tool and multi-finger hand. Its geometry, sensing and materials determine which objects and processes the robot can physically handle.

The market spans task-specific industrial tooling and generalized dexterous hands. SCHUNK, Robotiq and OnRobot sell grippers and end-of-arm tooling. Shadow Robot, Wuji Hand and Sharpa develop multi-finger, or dexterous hands. I wrote about the designs and economics in Dexterous Hand Primer:

An end effector needs to manipulate objects at the required speed, force, tolerance, uptime and price. A simple gripper is sufficient for pick-and-place tasks. A dexterous hand widens the task set, but also adds joints, weight and cost.

Tactile and force-torque sensing sit close to this layer, or you can think of sensing as layer 1b. These sensors tell the robot whether contact is secure, excessive or beginning to slip.

What matters: what touches the work, and what tasks does that choice open up or rule out?

2. What are robot actuators? Turn command into force

Actuators convert electrical commands into movement. The actuator shapes the force, speed, precision, efficiency and thermal performance a robot can achieve.

Humanoid joints have to reconcile high torque, low weight, compactness, low backlash, reliability, cost and manufacturability.

I saw this layer up close in a precision gear factory in Songgang, Shenzhen. 26 wire-cutting machines ran around the clock. Staff inspected parts to tight tolerances. The joint’s eventual force, precision and reliability began in processes like these.

An actuator that is heavy, expensive or thermally constrained can limit the system even as the software improves.

What matters: can every joint deliver the torque, precision and thermal margin the task requires, at a weight and cost the robot can carry?

3. What powers a robot? Store and deliver energy

The power system provides energy to the actuators, compute, sensors and auxiliary systems. For mobile robots, this usually begins with a battery pack and extends through the battery-management system, power electronics, wiring, protection and charging interface.

Capacity determines how long the robot can work between charges. Pack weight competes directly with payload, while charging time and cycle life shape fleet utilization and replacement cost.

The power system can constrain runtime through charging frequency, battery swaps and thermal limits.

What matters: can the power system sustain the task’s peak load? What is the runtime of the robot?

4. What is robot control? Move toward the target

Control is the software layer that converts a desired motion into the continuous stream of motor commands that produces it. The control loop recalculates the required position, velocity, torque or motor current hundreds to thousands of times a second on a real-time processor. It has to absorb a late measurement, a small collision, a heavy object or a battery-voltage change while keeping the robot stable and within its limits.

Two common approaches are (1) PID, proportional-integral-derivative, the classical feedback loop: it measures the gap between where a joint is and where it should be, then corrects in proportion to that gap; and (2) model predictive control, or MPC, which simulates the effect of potential commands and picks one that satisfies constraints such as torque limits, balance, collision clearance or energy use.

Much of this layer is available through open-source libraries and vendor tools, including ros2_control, Drake and NVIDIA Isaac. Robot companies still have to integrate and tune the loops for their mechanics, tasks and safety limits. Companies such as Applied Intuition and Intrinsic also sell parts of the broader robotics software and tooling stack.

What matters: when something unexpected happens, what keeps the robot safe and on task?

5. What is robot motion planning? Choose a feasible path

Motion planning is software that selects a feasible sequence of movements, from the robot’s current state to a goal. For a manipulator, that may include choosing a grasp, or placing the object without colliding with the shelf. For a mobile robot, it may include route planning and task sequencing.

Classical methods include graph search, inverse kinematics, sampling-based planners such as rapidly exploring random trees, and trajectory optimization. MoveIt is an open-source manipulation framework. MoveIt Pro packages a commercially supported workflow. NVIDIA’s cuMotion library accelerates trajectory optimization on GPUs.

Planning and control are adjacent but serve different purposes. Planning decides what movement to attempt. Control makes the robot follow it in real time. This guide traces the dependencies from the object, so control appears first. During the task, information generally flows from perception to planning to control.

What matters: can the robot find a safe, efficient path for the current scene, and recover when that path fails?

6. What is robot perception? Estimate what is happening now

Robot perception turns sensor readings into an estimate of the robot and its environment. Cameras, depth sensors, LiDAR, force-torque sensors, tactile sensors, encoders and IMUs all contribute different information. A camera may identify a package. A wrist force sensor may reveal that it is snagged. Tactile sensing may show that it has begun to slip.

Perception answers a present-tense question: where are the object, robot and relevant contacts now? A model can use that information, together with instruction and goal, to decide what to do next.

Some perception functions are available as libraries, including object detection, pose estimation, visual odometry, SLAM and sensor fusion. NVIDIA Isaac ROS is a packaged example. For deployment, the robot still depends on sensor calibration, latency and edge cases of the target environment.

Control, planning and perception usually run on the robot. Perception and planning commonly run on the main onboard computer, while the fastest control loops may run closer to the joints.

What matters: what does the robot need to observe, measure and infer before it can act reliably?

7. What are VLMs, VLAs, world models in robotics?

Models interpret observations, reason about a goal, predict outcomes or generate actions.

• A vision-language model, or VLM, processes visual and language inputs. In robotics, it can interpret a scene and an instruction and produce a high level plan.

• A vision-language-action model, or VLA, combines vision and language inputs with an action output. Physical Intelligence’s π models, Google DeepMind’s Gemini Robotics and NVIDIA’s GR00T are examples.

• A world model learns how a scene may evolve. Given a current state and a possible action, it predicts the future state. NVIDIA Cosmos and World Labs’ Marble are different approaches in this broad category.

• A world action model, or WAM, is a world model that also outputs actions. Dyna Robotics’ DYNA-2 is one example. NVIDIA’s Cosmos 3 also includes policy variants that connect video prediction with robot actions.

Perception estimates the present, a world model predicts possible futures and a world action model ties those futures to executable choices.

The model is absorbing the stack. These models are increasingly taking on work that previously sat in the perception and planning layers. For example, a VLA may generate a high-level action chunk directly from camera input and an instruction, bypassing several hand-built perception and planning interfaces. It may also generate lower-level actions, but fast motor control still sits underneath it. A world action model can predict how candidate actions change the scene, which can support or replace part of a planner’s work.

What matters: which decision does the model make, and which decisions remain with conventional software and control?

8. What data trains physical AI systems?

There are 5 main types of robot-learning data:

• Robot demonstrations (or teleoperation data): A human controls a robot and records observations and actions together.

• Human demonstrations (or egocentric data): Humans record themselves completing tasks.

• Simulation: A virtual environment generates trajectories, labels and failures under controlled conditions.

• Internet and video data: YouTube and other online videos provide demonstrations at much larger scale than robot fleets.

• Deployment data: success, failure and recovery data from robots doing the actual work

I wrote about robot data here.

Volume is only part of the equation. The data also need to be relevant to the task, embodiment, sensor and environment. Failure data and variety of data are also key.

A whole universe of companies is forming around this layer. XDOF builds robotics data infrastructure and datasets. Mecka describes itself as a data, evaluation and deployment layer for physical AI, with a focus on egocentric human activity.

What matters: who can produce data that is relevant to this robot’s work, and how does that data improve the next release?

9. Where does physical AI compute run?

Robotics uses at least three kinds of compute:

(1) Cloud or datacenter compute trains models, processes fleets of data and runs large simulation workloads away from the robot.

(2) Onboard accelerated compute handles latency-sensitive perception, model inference and sometimes planning. NVIDIA Jetson and Jetson Thor are examples of this tier. Qualcomm, AMD and specialized edge-compute vendors also make processors for this tier. These computers need to fit the robot’s power, heat, weight and cost budgets.

(3) Real-time embedded compute executes fast control and safety functions on motor drives, microcontrollers or real-time processors. Texas Instruments, NXP, Infineon, STMicroelectronics and motor-drive vendors make processors and control components for this tier.

What matters: which computation must happen on the robot, and can the hardware deliver it inside the latency, power and thermal budget?

10. What is robotics simulation used for?

Simulation creates a controlled environment for design, training, integration and stress testing before a policy reaches the real world.

Simulation can generate variation that is expensive or unsafe to capture on hardware. It can also test an integrated stack before a physical robot is available at scale.

The sim-to-real gap is the mismatch between the simulated environment and the physical one. Contact physics, material wear, lighting, sensor noise are common sources of mismatch. A policy can perform well in simulation, and still fail in a real deployment.

MuJoCo is a physics engine widely used in research. Gazebo supplies open-source physics, rendering and sensor simulation. NVIDIA Isaac Sim connects simulation with NVIDIA’s broader robot-development workflow.

What matters: what did the team validate in simulation, and which real-world conditions remain outside that environment?

11. How should physical AI systems be evaluated?

Evaluation measures the gap between reality and expectation. For example, whether the system works across changing conditions, recovers from errors and makes economic sense. A useful scorecard includes task success, intervention rate, recovery behavior, cycle time, latency, energy consumption, safety and cost per successful task.

Evaluation also closes the learning loop. Failure helps inform the model training, data collection, hardware or workflow adjustments.

What matters: how do we tell that this system works in the target environment and business case?

Where does value accrue in the physical AI stack?

Companies create the most value when they solve the bottleneck holding back an important task.

Better models reduce the burden on hand-built perception and planning. Better sensing, end effectors, actuators and power systems make more behaviors possible or keep them running longer. More deployment produces more interaction data. That data then improves the model.

The cycle is better robot → more deployment → more data → better model → better robot.

The flywheel turns only as fast as its slowest layer. For one company that slowest layer is data collection. For another it is the hand, the actuator, the power system, edge compute, safety validation or the cost of running the robot at all.

The most valuable part of the stack can change as technology improves. Today, the main limitation may be the model and data pipeline. As models improve, the bottleneck may move to the actuator, sensor, or deployment network.

How can investors and operators use the physical AI stack?

Start with a specific task and ask 3 questions:

a. Where does the system learn? Trace who generates interaction data, who owns it and how it reaches the next version.

b. What limits the task today? Locate the constraint in the model, sensing, mechanics, power, reliability or unit economics.

c. What evidence would establish deployment? Ask for the operating conditions, task success, recovery behavior, cycle time, energy use and the economics of repeated operation.

The task-first stack breaks down a robot demo into a chain of parts that can be examined individually. Trace the task from physical contact back through the stack. It is then easier to see what the system depends on, where a company may have an advantage and where the robot is most likely to fail.

For paid readers, below I apply the stack to three tasks and surface three different bottlenecks: warehouse bin picking, inserting a connector and loading a dishwasher.

Applying the stack: three tasks, three bottlenecks

Read more

Welcome back to the 4-part series “the Business of Actuators”. This is part 2. If you’re new here, this series aims to cover the business of actuators. We go deep enough technically in order to understand the underlying factors driving the economics of actuators, and hence physical AI. Part 1 covered what is inside a humanoid joint: mechanical parts, the main designs, joint anatomy and the low-ratio vs. high-ratio design trade. Start there if you missed it.

This piece looks at how much an actuator costs. We work from an audited filing, current vendor listings and an independent teardown. One buyer’s audited price for a single driven axis is $202, and an independent teardown estimates a comparable joint near $167. [1]

And thank you for the feedback on the first post. If you are building actuators, robotics components or industrial automation systems, I would love to hear from you.

Sections

1. How Much of a Humanoid’s BOM Is Actuation?

2. The Right Unit of Comparison: Cost Per Actuated Axis

3. What Does One Actuated Axis Cost?

4. Why Retail Pricing Overstates What a Manufacturer Pays

5. What the $202 Benchmark Actually Tells Us

6. Data Provenance and Methodology

1. How Much of a Humanoid’s BOM Is Actuation?

One benefit of the current wave of Chinese robotics IPO filings is that we are finally getting more granular supply-chain data.

LeJu Robot, the company behind the Kuavo humanoid, filed for an IPO on the Shenzhen Stock Exchange’s ChiNext board under its fourth listing standard for high-growth technology companies.

Its prospectus reports direct-material cost of $19,587 per Kuavo unit, representing 78.9% of total build cost. [1] Most percentages below use that $19,587 direct-material denominator; a few use total unit cost or another filer’s own hardware-cost base, each labeled where it appears.

Start with the joint modules. Kuavo carries between 26 and 38 modules. LeJu’s audited procurement table gives a blended module price of RMB1,115, or approximately $156. That puts joint modules alone at 20.7% to 30.3% of direct-material cost. [1]

Now add the driver electronics. LeJu buys the driver boards on a separate line. Driver-board spend is 29.5% of joint-module spend in the same audited table, which puts $46 of drive electronics on a $156 module. The combined module and driver basket therefore reaches 26.8% to 39.2% of direct-material cost. [1]

Then add the dexterous hands. LeJu’s procurement data prices the pair at $3,370. Including the hands brings the full electromechanical drivetrain to 44.0% to 56.4% of direct-material cost. [1]

For a full electromechanical drivetrain including hands, the share is 44.0% to 56.4% of direct-material cost and 34.7% to 44.5% of total unit cost. [1]

Quick glossary:

• Direct materials are the parts and inputs that go into the robot.

• Total unit cost is direct materials plus assembly labor, manufacturing overhead and other costs allocated to the finished unit. For reference, 78.9% of Kuavo’s total unit cost of $24,825 is direct materials.

• Electromechanical drivetrain is the full motion path: joint modules, electronics and hands

• Joint module is just the mechanical assembly, motor and reducer.

This is why published actuator shares are all over the place. These calculations are using different inputs: the numerator, the denominator, and the robot. [1]

• Numerator: the actuator can mean at least three things: (a) the mechanical joint module, (b) the joint module plus driver electronics, (c) the full electromechanical drivetrain including the dexterous hands.

• Denominator: which dollar figure you divide by. Direct materials, total unit cost, hardware cost and ex-factory price are four different bases. The same $156 module is 0.8% of Kuavo’s $19,587 direct-material cost and 0.63% of its $24,825 total unit cost, before the numerator basket even changes.

• Robot: which machine produced the numbers. A full-size biped, a wheeled AMR and a quadruped carry different actuated-axis counts and different architectures, so the share moves with the machine even when the basket and the denominator are held fixed. Youibot’s 50.2% below states its numerator and denominator clearly, and it is still the wrong robot for a humanoid benchmark.

Our own earlier estimates of 40% to 50%, and separately 50% to 65%, labeled these boundaries too loosely. I am retiring them in favor of the basket approach above.

Drivetrain BOM share has trended up over time. Youibot, which makes wheeled and inspection robots and filed for an HKEX listing under Chapter 18C, reports drivetrain rising from 33.7% to 43.8% and then 50.2% of its hardware cost over three years. [2]

Youibot (a Chinese industrial robot company)’s fleet is still mostly AMRs and inspection robots. It launched two humanoids in 2025, a wheeled model in March and a bipedal model in August. Humanoid revenue stays inside total company revenue in the filing and reads as immaterial. As a reference, Youibot shows that the drivetrain can become a larger share of the hardware basket as robot capability increases.

For a full-size biped the share is higher. An independent teardown of the Unitree G1, which shipped more than 5,500 units in 2025, puts its joints alone at about 66% of that robot’s bill of materials. [8] That covers joints only, on a different robot than the Kuavo ladder above.

2. A Consistent Unit of Comparison: Cost Per Actuated Axis

Before comparing prices, we need a consistent unit. Robotics terminology makes this more confusing than it should be.

• A joint module is the physical mechanical assembly around a joint: typically the motor, reducer, encoder, bearings and housing. The driver is the power and control electronics.

• An actuated axis is one independently driven direction of motion: one commandable output with its associated motor and transmission chain.

• A physical joint can contain more than one axis. A wrist that both bends and rotates may contain two independently controlled axes in a single joint assembly.

• Actuator is the physical hardware component.

• Degrees of freedom, or DoF, describe the robot’s kinematic freedom rather than how many motors it contains. It refers to the total number of independent, movable joints that a robot can shift position or rotate.

Actuator count, actuated-axis count and DoF are often close to one another. But they are not universally 1:1.

Below are three examples:

Tendon-driven dexterous hand. To actively control 1 single rotational axis in both directions, 2 actuators are needed to pull in opposite directions, like a pulley. This is similar to bicep and tricep muscles for humans. (actuators = 2, actuated axis = 1, DOF = 1)

Swinging pendulum. A swinging pendulum has unactuated (passive) degrees of freedom (DOF), but zero actuated axes. (actuated axes = actuator count = 0, DOF = 1).

Soft robotics hands. Made of silicone, the material can bend in every millimeter of its structure. So, its DOF is theoretically infinite, while driven by a single cable motor (actuated axes = actuator count = 1, DOF = ∞).

For consistency, we use cost per actuated axis as the comparison unit in this piece. It gets us closer to the cost of adding one independently controlled direction of motion to a robot.

3. What Does One Actuated Axis Cost?

LeJu’s filings shed light on actuators volume pricing in China today.

Its audited prospectus reports Rmb40.3mn ($5.63mn), spent on joint modules in 2025 at a blended unit price of Rmb1,115, or about $156 per module. [1]

Dividing total module expenditure by that disclosed unit price implies roughly 36,000 modules purchased. LeJu buys driver boards on a separate line. Driver-board spend is 29.5% of joint-module spend, which puts $46 of drive electronics on each module. [1]

That brings the combined mechanical module plus driver electronics to:

$156 + $46 = $202 per actuated axis.

This is a useful high-volume benchmark. Zooming in to the data and recent statements, we found three insights about the pricing:

1) Price halved as LeJu’s purchasing scaled 100X. Over the two years LeJu’s annual module purchases went from roughly 370 to roughly 36,000, and its per-axis cost fell from $433 to $202, with nearly all of the decline on the driver-electronics line. [1]

2) $202 is a related-party price. LeJu owns 5.73% of the supplier making these modules, its finance chief sits on that supplier’s board, and roughly 83% of its comparable-product procurement comes from that related party, Wuxi Quanzhibo. [1] The filing says pricing uses a cost-plus arrangement and sits slightly below comparable products. [1]

3) Industry-grade joint modules on average cost $210 in 2025, to fall 10% in 2026 and mass-market pricing at $140. Quanzhibo, LeJu’s supplier, shared some details on pricing in July. Its CMO shared the 2025 industrial-grade joint-module average near Rmb1,500 ($210), expects module costs to fall roughly 10% in 2026, and the whole-robot mass-market pricing at Rmb1,000 ($140). [9] LeJu’s Rmb,115 blended price sits below that 2025 industrial average, which is consistent with the filing’s “slightly below comparable products”.

The filing gives no torque, mass, gear ratio or gear family for the $156 module, so it represents an average across joint sizes. One independent build-cost estimate exists for a joint of the same type. China Post Securities’ March 2026 teardown of the Unitree G1 built its joints up from components: Rmb1,000 ($140) for a small joint and Rmb1,500 ($210) for a large one, about $167 blended, driver board included. [8]

A merchant module price of $202 and an in-house build-cost estimate of $167 are referring to the same joint module (motor + reducer + driver + encoder).

4. Why Retail Pricing Overstates What a Manufacturer Pays

Now compare that $170-200 volume figure with what a robotics startup pays one unit today.

Current online-retailer listings, quoted in US dollars, from September 2026 include:

• DAMIAO DM-J8009-2EC: 9:1 low-ratio planetary architecture, 20 N·m rated / 40 N·m peak torque, 896 g, dual encoders, onboard driver — $385. [6]

• MyActuator RMD-X8-25: $450 without the USB-to-UART accessory and $480 with it. The selected version is listed “Without Driver,” so it is not directly comparable with DAMIAO without separately pricing the motor driver. [7][11]

• ZeroErr eRob 80T: 50:1 to 120:1 strain-wave architecture, 21 to 31 N·m rated and 44 to 143 N·m peak, 1.94 kg — starting at $1,440. The listing does not clearly specify driver status.

Pendulum Robotics’ public comparison table lists 130+ named rotary actuators with price, torque, dimensions, ratio, gearbox type and encoder configuration. The table includes a $325 DaMiao DM-JH11, a 52 mm, 51:1 harmonic actuator, and a $257 DaMiao DM-J6248P-2EC, a 76 mm, 48:1 harmonic actuator. These are public single-unit retail listings for finished actuators. The $202 benchmark above is an OEM volume-procurement cost. [7]

These are not apples-to-apples products. Each quoted price bundles architecture, size, electronics integration and purchase volume

The ZeroErr strain-wave module sells at 3.6X the price of the DAMIAO planetary module, and delivers less peak torque per kilogram at every ratio, 0.8 times DAMIAO’s. We covered the performance trade-offs behind those architectures in Part 1.

LeJu’s $202 combined module-and-driver cost is about 52% of DAMIAO’s $385 retail price for a broadly comparable low-ratio module. [1][6]

Single-unit retail pricing is a poor proxy for what a manufacturer pays at volume. LeJu bought roughly 36,000 modules in 2025 across about 1,000 robots. Unitree, which shipped more than 5,500 humanoids that year, designs and integrates its joint modules in house. [8] Part of the gap between $202 and a $385 retail listing is purchasing scale, and part is the related-party pricing already described.

5. What the $202 Benchmark Actually Tells Us

We draw two conclusions from the numbers.

Read more

Actuators can account for up to 70% of a humanoid robot’s bill of materials. A typical humanoid packs dozens of them. A design decision at one joint cascades down the robot.

The key question running through this piece is: where does a robot pay for torque?

More gearing shifts the engineering burden toward precision mechanics. Less gearing shifts it toward the motor, electronics and thermal system. Repeated across dozens of joints, that single design decision ends up deciding what the robot costs to build and which suppliers capture the value.

Actuators are complex machines. Over the past few weeks I have spent a lot of time learning how they actually work. I’m thankful to the authors of the many resources I have relied on along the way. They are cited in the Appendix.

By the end of this piece, you’ll know what is inside a humanoid joint, why it is built the way it is, and how that design choice redistributes value across the supply chain, from a small set of precision gear-cutters to a global market for rare-earth magnets.

This is Part 1. Subscribe to get the future posts in your inbox as they go live.

Sections

1. What’s Inside a Humanoid Joint?

2. Why Gearing Is the Key Architectural Decision

3. Five Actuator Metrics That Matter

4. What Gearbox Choice Tells You About the Supply Chain

5. Why QDD Changes Actuator Economics

6. Where Different Actuator Architectures Get Used

7. Where the Robot Pays for Performance

8. Acknowledgements

9. Appendix A: What’s Inside an Actuator

10. Appendix B: How to Read an Actuator Spec Sheet

11. Appendix C: How the Three Gearbox Architectures Work

1. What’s Inside a Humanoid Joint?

Before we look inside the actuator, I want to clarify three terms that repeatedly come up: joint module, motor driver and actuator. The industry uses them interchangeably, but they are actually different things.

A joint module is the mechanical assembly: motor, reducer, encoder, bearings, housing, and on some designs a brake. Two filings classify the components differently: one groups the encoder inside the joint module, the other treats it as a separate control component alongside the driver board.

A motor driver is the power and control board. It turns battery DC into three-phase current for the motor and closes the control loop.

An actuator is the broad category word that can cover both. Two vendors quoting an actuator can therefore be quoting two different bills of materials, one with the driver board inside the housing and one selling that board as a separate line.

For full details of each component, see Appendix A of this post.

2. Why Gearing Is the Key Architectural Decision

Electric motors and humanoid joints naturally operate in very different regimes. Motors are most compact and efficient when spinning fast. Humanoid joints need to move relatively slowly while producing turning force, or torque.

A reducer, or the gearbox, bridges the two: it reduces the motor’s speed for greater torque at the joint.

A higher reduction ratio gives the joint more mechanical leverage, but that comes with costs. Here are three of the penalties, and how each translates into economic trade-offs elsewhere in the robot:

1. Energy loss means bigger batteries and shorter runtime. Gears lose some energy to friction. A single planetary stage returns 95%-98% of the motor’s work, a strain-wave stage drops to 85 to 90%. A stacked two-stage planetary train (we explain more what that is in Appendix C) degrades further: squaring that range gives roughly 90% to 96%, an arithmetic estimate rather than a measured figure. Across dozens of joints, those losses mean more heat, shorter runtime, and potentially larger batteries or additional cooling.

2. Heavier joints increase BOM costs. Reducers contain precision steel components. And steel is very heavy. A heavier knee doesn’t merely cost 300g more. The hip actuator then needs more torque to move that knee, which means bigger motor, more copper/magnets, stronger bearings, more current and potentially a larger battery. Mass at each joint compounds through the robot.

3. High reduction lowers force transparency and dynamic response. High gearing mechanically isolates the joint from the motor. That makes the joint harder to move from the outside. It also makes the motor less able to feel external forces. One reason is reflected inertia: the motor’s rotational inertia is amplified through the gearbox by the square of the gear ratio.

This leads to one of the central design decisions inside a humanoid is, where do you pay for torque?

And it has implications of where the cost sits in the actuator.

At a simplified level, gearbox designs can sit on either end of the spectrum:

a) Smaller, faster motor + high-ratio reducer
b) Larger, high-torque motor + lower-ratio reducer

At the high-reduction end (a), more of the engineering burden sits in the reducer: precision gears, tight tolerances, specialized manufacturing and, depending on the architecture, multiple reduction stages.

At the low-reduction end (b), more of it shifts into the motor and electrical system: magnets, copper, higher current, electronics handling motor current and thermal management.

So, this design choice determines which components matter most, which suppliers are difficult to replace, and where value accrues inside the actuator supply chain.

3. Five Actuator Metrics That Matter

These are the five metrics to look at when evaluating an actuator claim, and what each tells us about the robot behind it.

1. Continuous vs peak torque: Can the robot sustain the demo workload?

Peak torque tells you what a robot can do briefly. Continuous torque tells you what it can do repeatedly in deployment.

2. Torque density: How much actuator mass is required to produce useful force?

Higher torque density means more torque for the same actuator mass. That cascades into lighter limbs, smaller upstream actuators and potentially lower overall BOM.

3. Backdrivability: How naturally can the robot absorb and sense external forces?

This matters for walking, manipulation and human interaction.

4. Backlash: How much mechanical uncertainty exists when direction reverses?

This matters for precision manipulation.

5. Shock tolerance: Can the hardware survive repeated real-world impacts?

This matters for reliability, maintenance and useful life.

The full list of actuator metrics is at Appendix B.

4. What Gearbox Choice Tells You About the Supply Chain

The three major gearbox architectures solve the same basic problem, but their manufacturing requirements create very different supplier economics.

Harmonic drives provide compact high reduction and near-zero backlash, but require specialized precision manufacturing. This includes grinding a flexpline that has to flex millions of cycles without cracking.

That manufacturing base has historically concentrated in Japan, and a few players such as Harmonic Drive Systems and Nabtesco. Chinese entrants, including Laifual and Leaderdrive, are now taking share, and the shift looks more like a maturing competitive base than a pure price war: Laifual’s most recent interim results showed shipments and margin both rising, not falling. We’ll cover more on Part 3 of the series.

Planetary drives prioritize efficiency, shock tolerance and manufacturing accessibility. The gears supply is not scarce; it can draw from a much broader conventional gear-manufacturing ecosystem. The challenge is to hold tight tolerances at high volumes. Low-ratio planetary is also the architecture behind QDD, which trades gearbox difficulty for another supply chain problem, covered in the next section.

Cycloidal drives serve high-load applications. The hard part to source is the eccentric disc’s precision profile, machined by a narrower set of shops than planetary gear-cutting, though a broader one than harmonic. They are less common in humanoids today, though at least one Chinese humanoid maker names a cycloidal joint as a core technology, so “heavier, so avoided” does not hold across every design.

The design, pros and cons of each gearbox architecture are covered in detail in the Appendix C.

5. Why QDD Changes Actuator Economics

If you follow humanoids coverage, you probably have come across the word QDD quite a bit. QDD, or Quasi-direct drive in full, is an actuator design philosophy, not a gearbox type. It refers to a system where a high torque motor is paired with a very low reduction gearbox (3:1 - 10:1). Quasi means almost: the motor is almost directly driving the joint, with just enough gearing to increase torque while preserving much of its responsiveness to external force.

QDD gearboxes are predominantly planetary due to their low cost and high mechanical efficiency (95%+), though low-ratio cycloidal variants are emerging.

The MIT Mini Cheetah actuator, developed by Ben Katz, became an influential reference design for modern QDD modules. Mini Cheetah used a single 6:1 planetary stage delivering 6.9 N·m continuous torque and 17 N·m peak. By contrast, traditional industrial robot wrists run gear ratios between 50:1 and 100:1 to statically lock loads. Even humanoid knees, which require higher torque to hold deep crouches, use low-ratio multi-stage planetary setups (eg, 22.5:1), staying well below industrial standards to preserve shock tolerance.

QDD changes where the robot pays for torque

Traditional robot architectures often pair smaller, fast-spinning motors with high-ratio reducers to generate torque at the joint. More of the engineering burden therefore sits in the reducer.

QDD flips this. More of the torque comes from the motor, which means, more magnets, copper, and more demanding electronics and thermal management, in exchange for a simpler, lower-ratio reducer.

By shifting the workload to a powerful, high-diameter motor, QDD delivers three mechanical advantages:

1. High backdrivability. At a low ratio you push the output and the motor turns. The robot feels external force through the joint itself, enabling soft compliance. In contrast, a high-ratio gearbox behaves closer to a one-way valve.

2. Low reflected inertia. Low reflected inertia allows a humanoid to react more quickly to disturbances, such as a stumble.

3. Shock tolerance. A walking robot hits the ground several times a second. Lower-ratio designs generally tolerate impact better.

That mechanical shift has economic implications.

• High ratio keeps the burden in the gearbox: precision-ground teeth, tight tolerances, and a small set of manufacturers who can hold them. High ratio buys precision at the cost of a narrow gear-cutting supplier base.

• Low ratio pushes the burden into the motor and its electronics instead: a wider stator, more magnet volume to generate torque directly, more copper in the stator windings, and a driver board built to handle it. Low ratio buys shock tolerance at the cost of rare-earth magnet and higher-current electronics.

Even though electronics is more complex, QDD can lower the total actuator BOM of a robot. Below are some reasons:

• Supply Chain Accessibility. Low-ratio planetary gearboxes are standard industrial components. They can be made by widely available gear-hobbing equipment. Harmonic reducers need a much more specialized supplier base. That being said, a lower-ratio design needs more concentrated magnets to make up for the less sophisticated gearing. China accounts for 94% of global sintered neodymium-iron-boron magnet production, up from about 50% two decades ago. [12]. We’ll cover in Part 3 of this series

• Cost & Efficiency Gains. According to SemiAnalysis’s June 2026 G1 teardown, low-ratio planetary gearboxes achieve 95% to 98% efficiency (vs. 85% to 90% for strain-wave) and up to 80% cheaper.

• Driver Localization. Public financial filings from Leju, a humanoid maker in China, show driver expenditure per joint module fell 83% in two years.

6. Where Different Actuator Architectures Get Used

Humanoids do not use a single “one-size-fits-all” actuator. Different joints optimize for different combinations of force, precision, impact tolerance, mass and packaging.

Legs and hips prioritize shock absorption, rapid acceleration, and heavy force production. To protect from ground impact, these joints lean heavily toward QDD planetary setups or high-force linear actuators.

Wrists and arms prioritize spatial precision, high torque-to-weight ratios, and compact form factors. They often employ harmonic reducers to maximize holding torque in tight spaces. Alternatively, QDD motors are placed centrally, then power is routed through tendons to keep the limbs light.

Current humanoid designs show how these trade-offs play out in practice:

• Tesla Optimus Gen 2/3 uses harmonic reducers across rotary body joints (shoulder, torso, wrist roll) for compact form factor and precision, while it uses linear actuators for high-force joints like knees and elbows.

• 1X’s NEO hand uses a tendon-driven Quasi-Direct Drive (QDD) architecture at approximately 5:1 to 15:1 gear ratios. Motors are housed in the forearms/torso to keep distal limb inertia minimal and limbs safe around humans.

• Unitree highlights its core architecture in investor briefings as a “high-bandwidth force-controlled quasi-direct-drive planetary rotary joint”. It combines a high-torque motor, low-ratio planetary gearhead, and rapid torque-control loop. (See Unitree’s 3-hour Investor Q&A for details).

7. Where the Robot Pays for Performance

This piece we covered the key question to ask when looking at actuators for humanoids or any robots: where the robot pays for performance. High reduction puts more of the burden into precision mechanical components. Low reduction shifts it toward motors, magnets, electronics and thermal management.

That architectural choice flows through the entire robot: how it moves, what it costs to build, which components become critical, and ultimately which suppliers capture value.

On Friday, we’ll put dollars against those choices: what a humanoid actuator actually costs, using audited filings and factory quotes, and why published estimates put actuators anywhere from 21% to 70% of a humanoid’s BOM. [1][4]

Acknowledgements

Thank you to Parth Ingle (@parthingle_x), Grant (@grantg07), and Sajiv Shah (@sajiv_shah) for technical review and feedback on earlier versions of this piece. Any errors or interpretations that remain are my own.

Appendix A: What’s Inside an Actuator

Below are the key parts inside an actuator and what they do.

Stator. The stator is the fixed half of the motor: a stationary ring of copper coils. When electricity passes through these coils, it creates an electromagnetic field.

Rotor. The rotor is the turning half. Permanent magnets (sintered neodymium-iron-boron) are bonded to a steel structure. As these magnets interact with the stator’s magnetic field, they force the rotor to spin.

The motor is built primarily from laminated electrical steel, copper wire, magnets and epoxy.

Reducer. Also called the gearbox, it reduces rotational speed and turns that into greater turning force, or torque, at the output. It is typically built from hardened steel components. There are three main types of reducers used in robotics, which we’ll cover later in the Appendix. The choice of gear family can be a major driver of joint cost.

Encoder. The encoder tells the robot’s computer what angle the joint is at, how fast it is moving, and which direction it is turning. It can use a magnetic ring or optical disc read by a sensor. Some joints use dual encoders, one on the motor and one on the output.

Bearings. Bearings support the actuator’s rotating components and carry loads between the actuator and the robot’s limb. They typically consist of rolling elements, such as balls or crossed rollers, running between two hardened races.

Driver board. The driver board turns commands from the robot’s main computer (e.g., an edge compute device in the chest or head) into controlled electrical power for the motor. It typically contains power transistors, a microcontroller and connectors.

Housing and brake. The housing holds the actuator assembly together and helps manage structural loads. It also has to deal with heat generated by the motor. Some joints include a brake to hold position when required.

Appendix B: How to Read a Humanoid Actuator Spec Sheet

Let’s cover the key terms to understand when reading an actuator’s spec sheet. One way that’s easier for me to understand these mechanical terms is to link them back to the human form and think about our own joints, muscles and movements.

Gear Ratio (or Reduction Ratio) (N:1) is the mechanical trade-off between speed and torque inside the actuator. It is usually written as N:1 (e.g., 9:1, 40:1). A high-speed, low-torque electric motor turns a gearbox. The gearbox slows down the rotation at the output shaft by a factor of N, while multiplying the motor’s torque at the output by approximately N, before accounting for efficiency losses.

Gear ratios are all about mechanical leverage.

• High ratios (40:1 to 100:1) function like a low gear on a bicycle. They provide greater mechanical leverage for producing and holding torque, but reduce output speed.

• Low ratios (3:1 to 9:1) function more like a springy, dynamic athlete’s limb. They are responsive and fast but provide less mechanical leverage, so the motor needs to do more of the heavy lifting.

Peak torque vs. rated / nominal torque (N·m). Rated / nominal torque is the torque a motor can produce continuously without exceeding its thermal limits. Peak torque is the higher torque a motor can produce for a short period before heat becomes the limiting factor. A large gap between peak and nominal torque means the actuator cannot sustain its headline peak performance continuously.

Torque is raw rotational strength. A bodybuilder’s bicep generates torque to curl a 100-lb dumbbell. It answers: “How much turning force comes out of this joint?”

Torque density (N·m/kg) is turning force per kilogram. It is an important metric in robotics because mass at the end of a limb costs torque at every upstream joint. Vendors rarely list this number directly; you calculate it by dividing torque by total actuator mass. Today’s merchant joint modules typically yield 45 to 75 N·m/kg at peak. Always check continuous torque density too, as thermal limits keep sustained performance below peak figures.

Think of an elite rock climber vs. a big powerlifter. A rock climber might be able to lift less absolute weight than the powerlifter, but can generate high force relative to body weight. In the same way, torque density tells us how much rotational force an actuator produces relative to its mass.

High torque density is important for robots because a heavier actuator requires more torque just to move the limb itself. It is also a major engineering challenge: producing more torque generally requires more motor material, which adds mass.

Backlash (arcmin) is mechanical slop, often measured in arcminutes. One arcminute is one-sixtieth of a degree. Turn a gear train one way, then reverse it, and the output can briefly remain stationary while the teeth cross the gap between them and begin transmitting force in the opposite direction.

Think of backlash like looseness in a joint. Too much can cause instability and reduce accuracy in fine tasks like threading a needle.

Very low backlash is useful for precision tasks, although some gear designs require small clearances for lubrication, thermal expansion and reliable operation.

Backdrivability describes how easily an external force applied at the output can drive the motor backward through the transmission. It can be characterized by the amount of output torque required to move the joint. A highly backdrivable joint can respond more naturally to external forces.

The analogy: a backdrivable joint is like a relaxed arm. When someone pushes your wrist, your arm yields smoothly. A poorly backdrivable joint behaves more like a rigid arm that resists being moved from the outside.

Backdrivable designs are useful for human interaction, force control and dynamic movement. Less-backdrivable designs can be useful when stiffness and holding loads are more important.

Shock tolerance describes how well an actuator survives sudden impact loads, such as when a robot’s foot hits the ground during walking. Vendors may characterize it using peak or momentary torque relative to the rated figure. The merchant modules reviewed above range from about 2x to 4.6x rated torque.

Appendix C: Comparing Gearbox Architectures: Harmonic, Cycloidal, and Planetary Drives

Below are the three main gearbox architectures used in robotic joints.

Harmonic, cycloidal and planetary gears all transfer rotational power and reduce speed. The difference is in how they do it. The three architectures make different trade-offs between positioning accuracy, shock tolerance, weight, efficiency and manufacturing complexity. Below, I’ll use the same human-anatomy analogies to help visualize each architecture.

Harmonic Drives (Flexing Cartilage & Tendons)

• How it works: A harmonic drive uses a flexible, thin-walled steel cup that deforms inside a rigid outer ring to transmit motion.

• Anatomy analogy: Think of flexible cartilage or tightly tensioned tendons. The architecture relies on controlled elastic deformation rather than only rigid gears rotating against one another.

• One-liner: Premium precision at a cost.

• Strengths: Very low backlash and high reduction ratios (often 50:1+) in a compact, lightweight package. This makes harmonic drives useful for applications such as precision industrial and surgical robots.

• Drawbacks: Higher unit cost, specialized manufacturing requirements and lower tolerance for shock loads than some alternative gearbox architectures. The flexible component is also subject to fatigue over its operating life.

Cycloidal Drives (Interlocking Vertebrae)

• How it works: A lobed disc moves eccentrically inside a ring of stationary pins or rollers, distributing forces across multiple contact points.

• Anatomy analogy: Think of interlocking spinal vertebrae. Loads are distributed across multiple contact points rather than concentrated on a small number of gear teeth.

• One-liner: The rugged option for heavy loads.

• Strengths: High shock tolerance, high load capacity and low backlash. Multiple contact points allow the gearbox to distribute sudden loads across the mechanism.

• Drawbacks: Cycloidal drives tend to be heavier and bulkier and can be more complex to assemble. This makes them more suitable for high-load applications where ruggedness matters more than minimizing joint mass.

Planetary Gearboxes (Skeletal Ball-and-Socket Joints)

• How it works: A central “sun” gear drives multiple “planet” gears inside an outer ring gear.

• Anatomy analogy: Think of a shoulder or hip joint: a rigid structure that distributes force across multiple contact points.

• One-liner: The scalable, cost-effective workhorse.

• Strengths: High energy efficiency (often 95%+ for a single stage), good shock resistance, and an established manufacturing base using conventional gear-making equipment. Low-ratio planetary gearboxes are commonly used in Quasi-Direct Drive (QDD) actuators at ratios such as 3:1 to 10:1.

• Drawbacks: Planetary gearboxes have mechanical backlash. Stacking multiple stages to achieve higher gear ratios can increase backlash, friction, weight and complexity.

Share Core Matter

Appendix D: Sources Cited

[1] LeJu / Lejoin Intelligence ChiNext prospectus, audited by Tianjian, procurement tables pp. 1-1-187 to 1-1-188.

[4] SemiAnalysis, “China’s Unitree Will Dominate Global Robotics,” 2026-06-09, read via Core Matter’s own capture of the piece.

[12] International Energy Agency, “China’s share in rare earth magnet production, 2024,” Rare Earth Elements report, April 2026: https://www.iea.org/data-and-statistics/charts/china-s-share-in-rare-earth-magnet-production-2024

What we cover

• Why fingers break before motors or gears in a robotic hand

• The move from tendon-driven transmission to rigid linkages in Prensilia’s product line

• A 300,000-cycle grip force test protocol and what it does not tell you

• Underactuation, degrees of freedom, and degrees of actuation explained

• Bill of materials breakdown for a 3-motor dexterous hand

• Customer mix across research, prosthetics, and industrial buyers

• Manufacturing scale from hundreds of units a year toward volume production

“Fingers are the parts that break the most, because you have impacts with objects, you have unexpected movements from the robot.” — Francesco Clemente, Managing Director, Prensilia

“We are talking about thousands of cycles, not really millions of cycles, with the tendons.” — Francesco Clemente, Managing Director, Prensilia

“The motors account for maybe thirty five percent, more or less, of the total cost.” — Francesco Clemente, Managing Director, Prensilia

Francesco Clemente on LinkedIn
Prensilia Website
Watch on YouTube

Subscribe now

Every episode, I sit down with the founders, researchers, and operators building physical AI worldwide, tracing the full stack from components and supply chains to what actually deploys. See all published episodes here

Chapters
00:00 Francesco Clemente, Managing Director, Prensilia
02:22 Prosthetics and research customers
04:46 Broken hands and reliability
07:09 Loaded versus unloaded cycle testing
11:59 Failure modes in fingers, gears, and motors
14:19 Motor heating and cooling
16:42 Tendon maintenance and anchoring
21:26 Weight and robotic-arm payload
23:48 Degrees of freedom and actuation
26:14 Underactuation and adaptive grasping
31:01 Motor current, torque, and heat
33:29 Grasp taxonomies and motor count
35:52 Abduction and hand design
38:17 Tactile sensing
40:38 Bill of materials and actuators
45:22 Industrial use cases
50:04 Competitors and grippers
54:42 Manufacturing scale

Core Matter is an independent research practice covering the physical AI value chain

Michelle Sun: Fingers are the parts that break the most. And that connection is a weak point that is gonna break. The motors are bottlenecked, and having a lot of motors inside of a small volume, it’s easy to look at spec sheets, but you have to understand that Francesco Clemente, managing director of Prensilia. Prensilia builds robotic hands out of Pisa, Italy. It’s fun out of set and in a school of Advanced Studies in 2009. Started off in prosthetics, today their hands are on humanoid robots, factory arms, and human wrists. Prensilia’s Mia hand runs three motors, closes in 280 milliseconds, which is faster than a human hand, and has survived 300,000 cycles at full grip force in testing. It sells for $10,000 to $15,000 compared to the Chinese hands at 5k. On the spec sheets, the Mia hand might look overpriced and under threat. What they’re seeing in order volume says something different. Today we’ll talk about what’s on the ground in the dexterous hand markets.

Francesco, welcome to the show. Prensilia is 17 years old today from 2009, and you are a spin off from Sant’Anna, started way before robot hands were cool, and definitely before the humanoid boom. So what did the company look like in 2009 and who paid the bills before humanoids existed as customers?

Francesco Clemente: Yes, thank you very much for having me today. I’m very excited to be here. The company, Prensilia, was founded, as you said, by researchers of the artificial enzy area that were working at the BioRobotics Institute of the Sant’Anna School of Advanced Studies. They were already working as a researcher to on the development of robotic hands and they were using developing those tools for their own research needs. Then other researchers that they were collaborating with started asking for accessing those tools, and basically here it how it came the idea to spin off a company in order to allow other researchers also to use these devices that were developed in the lab.

So the first customers were really other researchers, so universities and research centers around the world. And that let’s say research background is still with us. So today we are still working, collaborating a lot with research centers, and this is also one of the reasons why we have the specific versions of our robotic hands for research activities.

Michelle Sun: That’s amazing. And so the Mia hand came out of a prosthetic. How is the medical version similar or different from the robotics version? Did you have to change anything for the robotics customer?

Francesco Clemente: Yes. Apparently robotic hands can be used, let’s say, for prosthetics as well as for robotics today, but these two worlds are relatively different because you have very different specifications from one market to the other one. For instance, in prosthetics, you have a very clear bottleneck on the ability of the person to control the prosthesis. So you have a certain amount of degrees of freedom of movement that the person can control in a reliable way, so it doesn’t really make a lot of sense to have a 20 degrees of freedom robotic hand that can do any gesture, because the person cannot control that complexity.

On the other side, you are also limited by the weight and by the size of the device. Of course, besides prosthetics, you have specific sizes in prosthetics that you try to match. That is like small, medium, and large, I would say. The prosthesis has to be that specific size, it cannot be any size. So today we see robotic hands that have very different sizes, because of course you attach them to a robot, so it’s not super important if the robot itself and the robotic hand match in size perfectly, let’s say. But this is very important in prosthetics.

And of course another point is weight, because the person has to actively carry the prosthesis. The comfort is very important in that case, so you cannot really have a prosthesis that weighs two kilograms, because the person will be tired after one hour, let’s say, of using the prosthesis. So you really have to go down in weight in order for the prosthesis to be comfortable.

Michelle Sun: When we first spoke, you mentioned that at ICRA in Vienna earlier this year, people have been showing up at your booth carrying broken hands from other vendors. Tell me that story again. What had failed and what were the customers looking for?

Francesco Clemente: Yes, so we were at ICRA and we were meeting a lot of people at our booth. A lot of them were researchers or engineers from companies that were looking for robotic hands. And some of them were saying, okay, we are looking for robotic hands that are robust, that can be used also outside of the lab, because we were trying some other hands from competitors, but they didn’t work. We bought them because of the low price, but then we realized basically that they were not reliable for what we were going to do.

So I think that this summarizes a little bit what the status of the market is today, because there is a lot of competition, a lot of hype around robotic hands, and everyone is really working on these devices, so the demand is going higher and higher, but people are somehow approaching these devices for the first time. So they don’t really also don’t know exactly how to use them, what are the specs that are important. I mean, it’s normal, because these devices are complex. And for us that are working in the field since more than fifteen years, it’s clear what you are going to look at. But as you said before, robotics and also prosthetics was, let’s say, a niche market for robotic hands in particular a few years ago. Now everything is exploding, so people have to be accustomed also to understand the specs and what they are getting for the money.

Michelle Sun: I wanna double click on the point you make about there are things that are not on the spec sheets, right? So there’s degrees of freedom and you can also see the comparative price pretty easily. You tested Mia to three hundred thousand cycles at full force. Walk me through that protocol, because 300,000 loaded cycles and a million unloaded cycles are very different claims, and on the spec sheet it’s not easy to distinguish. Yeah, talk me through how to read through the spec sheets more efficiently and tell the differences between different hands.

Francesco Clemente: Yes. One thing that prosthetics has taught us is that we want to develop a robotic hand or a prosthesis that then can be used outside of the lab by patients and users. And also engineers, it has to be robust. Okay, so it’s easy to look at spec sheets that basically report the number of degrees of freedom, which is the kind of measure of the dexterity of the robotic hand that you’re buying, and everyone is looking into that today because they’re looking into fine manipulation protocols and solving manipulation at a higher level.

And of course you want to have something that is similar to the human hand in terms of movements that you can perform, but you have to understand that that comes with a cost associated. So the device becomes very, very complex. You either have a very bulky forearm that hosts all of the motors, with tendons that drive the fingers and the joints, or you have very small motors inside of the hand, inside of the joints of the fingers. And this means that the performance will have to be low because of the size of the motors that you’re using. So reliability is something that you have to compromise, also with dexterity, let’s say.

So we really pointed to have, tried to achieve, high robustness with our devices, and this is the reason why we basically developed a test protocol where we would run several thousand cycles at full force in order to really understand if people could use the device at the specs that are reported in the spec sheet. Okay, because if you say, okay, my hand can perform this type of grasps and this amount of force, 50 newton grip force, you have to know how many grasps you can really do at that amount of grip force, you know? Because you can test it at no load. But of course that’s not a measurement, or let’s say a parameter, that tells you in the real world how much the device is going to last.

Michelle Sun: If we do the math for, let’s say, the cost per grasp cycle, the Mia hand is ten to fifteen thousand dollars and the hand at the ICRA show probably around five thousand. How many cycles does that other hand last? And on a per grasp cycle basis, which one is actually cheaper?

Francesco Clemente: Well, it’s very difficult to tell a number for competitors, let’s say, because you really have to get the device and test it to know for sure, let’s say. But the thing is that there are two points. One is the grip force that you can get with those devices, that is relatively lower. So in general you also have maximum performance that’s capped because of how the device is designed. So if you have more motors, usually you have a lower grip force, this means that also for single grips you have a lower performance. And on the other side, of course, if you run at full grasp, you probably get a shorter and shorter lifespan.

So I can say for the competitors, okay, they will last 100,000 cycles or whatever, but the problem is that, what they reported to me is the fact that the performance was too low to carry out a large amount of tasks. And also, in the tasks where they were using the hand, they were still breaking, even if those tasks were relatively light, let’s say. So also moving cans, empty cans, let’s say, would be a relatively problem for those devices.

Michelle Sun: When you say that a hand starts breaking, are we thinking about the motors overheating, or actually like some of the fingers fall off? Or tell me more about the failure modes when it comes to, at the end of the 300,000 cycles, or in a competitor’s hand, a hundred thousand cycles, what would be the breaking point of the dexterous hand?

Francesco Clemente: Yeah, there is a lot of stuff that can break in a robotic hand. So one of the reasons why we try to limit the amount of motors in joints, it’s also to reduce complexity in order to have less stuff that can break, let’s say. But in general, of course, fingers are the parts that break the most, because you have impacts with objects, you have unexpected movements from the robot, so they are the most exposed part, I would say. So it’s important to have a device that allows you also to replace relatively quickly the fingers or the fingertips in order to service the hand in a relatively quick way. And this is, I would say, one problem that we also have and that we solved in this way, basically with a quick-release mechanism that allows you to replace the more distal part of the finger in a quick way, from a user perspective.

On the other side, of course, when you run a lot of cycles, you start to have wear of the components. So we have gears inside of our device, and these wear out because of friction. So at some point you will have to replace some components. First, I mean, usually you don’t have a dramatic, sudden break of the component, but you start having more mechanical play in the transmission. And then at some point, of course, the component is worn out too much, so that it basically breaks.

Another point that is very important is related to the motors. So the motors are the bottleneck in the mechanical transmission for several reasons. One is the one that you were mentioning before, so one is overheating. So electric motors heat a lot, and having a lot of motors inside of a small volume basically means that you need to have some sort of cooling system.

So if you look at the shadow robotic hand, that is this anthropomorphic robotic hand that was available since twenty years, they have a large forearm where they put the motors, and then with the tendon system they run the fingers, they have fans on the forearm in order to pull down the motors. So this is a problem that was there since all the time, let’s say, so it’s not a new problem, let’s say. So if instead you are putting the motors in the fingertip or in the joints, the problem is that you don’t really have a space to put fans there. And the motors, the smaller they get, the more they heat, because the surface area of the case of the motor is smaller. So the capacity to transfer heat outside decreases a lot with their volume and surface area, let’s say. So what happens is that sometimes you have to stop using the device in order to wait for the device to cool.

This is again one other reason for us to use less motors, bigger ones, so that they have less heating issues, let’s say. Just very quickly, another point is the gearhead of the motor that also limits the torque that the motor can provide to the system. So another point that can break is really the gearhead of the motor. So if you are pushing the motor to the limit, the structural components of the motor can fail at some point.

Michelle Sun: So it seems like there are four key areas that the hands can fail. The fingers can fall off on its own. And then there’s the motors wearing off after too many cycles. And the motors overheating is another one. And the last one is about the gears.

Francesco Clemente: Right. The gearheads. So the gears that are inside of the motor, let’s say, or coupled directly to the motor. So the last one is the gearheads within the motors being worn out because of too many cycles.

Michelle Sun: It’s interesting you mentioned Shadow. That’s actually my next question. You spent a decade building tendon-driven hands, the IH2 Azzurra, 11 degrees of freedom. And then you decided to remove tendons. Shadow used tendons for 20 years and half of the industry roadmap is still using tendons. What was the reason behind this design decision? Is that the cycles, tension lost, or is it really expensive to service?

Francesco Clemente: That’s a very good question. So tendons, we think tendons are great because they provide you a lot of flexibility in terms of design flexibility. So you really can move the motors around, because you can use the tendon to transfer motion from the motor to the fingers. So you have the ability to have more freedom of where to put the motors inside of your device.

Also, one of the reasons why they were used a lot is when you have a lot of motors, so that you can really store them in a more convenient way, and transfer motion with tendons in a very simple way around. But on the other side, the main problem of the tendons is, first, the fact that they usually break. I mean, tendons themselves are not a problem. I mean, tendons are also used in airplanes for controlling some components of the wings, so they’re not a problem by themselves. But the problem is that when you put them inside of a robotic hand, you have a very small amount of space available to make a tensioning system that is robust enough to be used correctly for a lot of cycles.

And specifically, tendons need a lot of maintenance. So the problem is again not with the tendons themselves but with how you anchor the tendon to the transmission. So the anchoring points are usually weak points. You can make a knot, you can use sleeves in order to connect the tendon to the mechanical transmission, let’s say, and that connection is a weak point that is gonna break. So even if the tendon is, let’s say, rated to hold one hundred fifty newtons, after a few cycles it will hold probably a hundred newtons. So they’re not very good with a large amount of cycles.

So we are talking about thousands of cycles, not really millions of cycles, with the tendons. And this is one of the main reasons why we moved to gears and links. Of course, gears are more reliable on this scale, I would say. They have the disadvantage that you have less freedom, because you cannot have a gear of any size, of any shape, you cannot transfer motion at any angle, let’s say, so you are more constrained. So you can do less, let’s say, with these tools.

Michelle Sun: So where does the anchor points fit? So you mentioned that there are all these tendons, and then there are points that are the failure points, right? The anchor points are usually where the problems start to appear. Are they within the knuckles area in each of the fingers, or where do you usually hide the anchors?

Francesco Clemente: We have one hand that is called IH2 Azzurra that is driven by tendons. So these tendons are anchored to the mechanical transmission on two sides. On one side is in the palm, because in the palm we have the motors, so we have a mechanism that basically pulls the tendon. So on one side the tendon is connected there in the palm. And on the other side is at the fingertip. So basically the tendon runs through the joints of the finger and then anchors to the back of the fingertip, so that when you pull the tendon, the finger will close. And when you release the tendon, the finger will open because of a spring inside of the joints. Usually, let’s say by generalizing, one point is before the joint that you want to move, usually close to the motor, and the other point is beyond the joint that you want to move. So at the fingertip, or let’s say between the metacarpophalangeal joint and the proximal interphalangeal joint, if that joint is the one that you want to move.

Michelle Sun: I noticed that the Shadow hand with the forearm units included is around four point three kilograms. Versus the Mia hand, with three motors inside the palm, is around a tenth of the weight. Walk me through the math when it comes to the weight and the performance that comes with it.

Francesco Clemente: Yes, of course. So weight is of course one of the important parameters when you want to design a robotic hand in general, so specifically for prosthetics, but in general for robotic hands, you need to set a certain weight. Of course, the Mia hand and the Shadow hand are very different devices, no? So three motors versus 20 plus motors. This comes with a set of different specifications and complexities of the device.

For us, the motor is one of the heaviest components that is inside of the hand. So one way to reduce the weight of the device was to limit the amount of motors that were inside of the device. So what does this mean? Basically, in this way you are setting the limits for the use cases in which you can use your device. Okay, so if your device weighs five hundred grams, it means that you can integrate it also in robotic arms that can carry up to three kilograms. So you will lose half a kilo for the end effector, but you still have two point five kilos for the application. You can use it in prosthetics. On the other side, if your device weighs five kilos, four kilos, you need at least a robotic arm that has a payload of ten kilograms, because if you use a robotic arm with a five kilogram payload, you basically don’t have any additional payload for the task that you need to do.

So it really sets the basis for the set of applications that you need to perform, that you can perform with your device.

Michelle Sun: Speaking of applications, so would you say that the Shadow hand is designed for a different set of applications versus a Mia hand? And what would those two categories be?

Francesco Clemente: Yes. So we designed the Mia hand specifically for prosthetics, and then we reengineered the device, modified the device in order to have also an industrial version. So we have two versions of the device, again, one for prosthetic users and one for industrial settings, so for automating processes in the industry. The Shadow hand is more of a research tool. So far as my knowledge goes, I’ve never seen that robotic hand used in an industrial setting. Of course it’s not usable in a prosthesis, as a prosthesis, but it really targets research settings, where you have capabilities to control the device that are, let’s say, not normal, I would say, because the device is very difficult to use, because you have to control 20 degrees of freedom all at once. Of course it’s becoming more and more simple thanks to AI. So AI is simplifying the use of more complex tools also in more standard settings, I would say. But as far as I know, that amount of complexity is still something that is used more in research settings.

Michelle Sun: That makes sense. And we’re getting to the audience favorite term, degrees of freedom. So everyone is obsessed about how many degrees of freedom, like our human arm has, a hand has, and the tool that you gave me last time was active versus passive degree of freedom. Tell me more about that, and how to read a spec sheet when it says, hey, there’s twenty degrees of freedom on this hand. How many motors are actually there? How many joints are just following along with the motor? Walk me through this vocabulary set of degrees of freedom.

Francesco Clemente: Yes, there is a little bit of confusion there, because degrees of freedom is a general term and then everyone is using it in a bit different way. One way to look at it is using degrees of freedom versus degrees of actuation, where degrees of freedom is number of joints and degrees of actuation is number of motors, or, as you said, passive and active degrees of freedom. So passive ones are the ones that are not controlled directly by a motor, while active ones do. This is the difference between the two.

So why do we have this difference? Because, for instance, if we take a robotic arm, we generally have six joints, six motors, one joint, one motor, and it’s very simple, let’s say. On the other side, in robotic hands, since several years, researchers have been exploring other ways to develop robotic hands, and one way is through underactuation. So underactuation means that the number of joints is larger than the number of motors that drive those joints. And the main reason for this is exactly to reduce the amount of complexity of the device, in some way also to increase the performance. Okay, because one way is the number of movements that you can do, but then you again have grip force, speed, weight, and all of the other performance metrics that you have to compare with and also kind of balance, let’s say.

In this case, what happens if you have less motors than joints, it means that these joints have to be coupled somehow together. Okay, so for instance, in Mia we have three motors that are used to actuate the flexion extension of the thumb, the flexion extension of the index finger, and then the last three fingers are coupled together. So you have three joints, but basically only one motor that controls those joints. And this limits the complexity of the device. This is one way, relatively simple way, to control different joints with a motor, basically with a rigid mechanical transmission.

On the other side, in Azzurra, for example, we have one motor per finger, but one finger has two joints. So we have one tendon that basically wraps around these two joints per finger, and then is used to control both joints simultaneously. This underactuation allows the finger to also wrap around the object and provide some flexibility. So rather than having a rigid transmission between these two joints, we have the tendon and springs inside. This basically allows for the proximal phalanx, in case the proximal phalanx encounters an object, the second phalanx can still continue to flex and adapt to the shape of the object. This flexibility is actually called underactuation and is, let’s say, a smart mechanical system that allows you to still have some adaptability of the shape of the hand to the object without having to actively control all of the degrees of freedom, so all of the joints independently.

Michelle Sun: Got it, got it. So there’s underactuated, when it comes to underactuated hands, then that means that the degrees of freedom is more than the number of motors, the actuators. Which sounds great when it comes to like less overheating, less weight, and less complexity. Is Azzurra also underactuated?

Francesco Clemente: Yes. Because it has eleven degrees of freedom and five motors.

Michelle Sun: I see. And then Mia hand, how many degrees of freedom, and three motors?

Francesco Clemente: That’s six degrees of freedom.

Michelle Sun: Six. Okay. And in theory that sounds great, right? Why wouldn’t all hands do that? Like what is the reason for people to wanna have, is there something like overactuated hands, or, you know, just what’s the opposite of an underactuated hand?

Francesco Clemente: The opposite is, let’s say, fully actuated. So for each joint you have one motor. And this, of course, is the simplest way to develop a robot. So for each joint you have an actuator, and you also have this direct driving, you know, let’s say. The problem, let’s say, with this, and specifically with this setup in a robotic hand, is the compromise that you have to make while choosing the motor and while choosing the performance, let’s say, of your hand, as we said before. You basically have two choices. You end up either having very small motors inside of the joints, because the joints and the fingers are very small, or you need to use some kind of mechanical transmission, like tendons, to have a remote actuation system and then transfer motion. Of course, this allows you to use larger motors, but the device becomes relatively bulky and heavy.

Michelle Sun: And so for Mia hand you use tendons that connect the motors to make it more degrees of freedom than the motors?

Francesco Clemente: No, no, for Mia hand we use rigid mechanical transmission, so linkage.

Michelle Sun: Azzurra hand?

Francesco Clemente: Yeah, that’s the third, yeah.

Michelle Sun: Azzurra, you use tendons?

Francesco Clemente: Yes, correct.

Michelle Sun: Yep. Yeah, makes sense. And you told me that a lot of these motors run five, ten minutes before they need a cooldown. How does a buyer look at the spec, is it possible to look at the spec sheet and know about this mechanism, like the cooldown performance? Is that something that people just don’t publish on the spec sheets?

Francesco Clemente: Yeah, this is very tricky. The problem is that it’s difficult to know from a spec sheet. The reason is that you don’t usually publish heating performance, dissipation performance, of your robotic system. Okay, this is linked to the fact that motors are designed with several parameters. They can run continuously at a specific, when they produce a specific torque. So if you provide a certain nominal current, this current can be provided continuously, and the torque that is generated can be generated continuously. So the motor can run without any stop or without any problem.

The fact is that electrical motors can be provided also with more current than this nominal one. The maximum current that you can provide is called stall current. So when the motor is not moving, you can provide the maximum amount of current, and the motor can generate the maximum amount of torque. Of course, you want to have this torque, because it’s much larger, so two to twenty times larger than the nominal one. So in order to have meaningful performance, you need to exploit these larger currents and torques, but the motor cannot handle these currents for a long amount of time. So this is the reason why you have to stop using it at one hundred percent after some time, because the motor itself cannot dissipate all of that heat that is generated.

So this is something that is well known in the field, and this is the reason why robotic arms are so big, you know, relatively to grippers, because the motors are, let’s say, oversized with respect to the torque that they would actually generate. So they are used at, let’s say, thirty percent maximum performance in order to avoid these issues and be able to run the robotic arm continuously. On the other side, robotic hands are very challenging at the moment. So we don’t really have technologies that allow us to use robotic hands at 30%, 20% of their maximum performance. So it really hits with these heating issues that come up all the time, I would say.

So for a customer it’s very difficult to know this, and I think that you really learn this by experience. So as a researcher that was working with these devices for several years, I’ve seen a lot of these hands, in general also grippers, more advanced grippers with very small joints, run into heating issues, and this is something that we were trying to design for during the years, but it’s something that’s not given for granted.

Michelle Sun: If you were designing the hand for a humanoid that ships ten thousand units or more in a couple of years, how many motors does it have?

Francesco Clemente: That’s a very difficult question. So the question of scalability is an important one. Again, we think that a system, an engineering system, in order to be optimal for the task, has to be sized for that task. I mean, you can design a robotic hand that has 20 motors that of course can perform a lot of movements, but it’s very difficult to scale from an industrial perspective. Okay, so our idea also when developing the Mia hand, and also with new devices that we are developing, is to study grasp taxonomies, because there are several research studies in the field that study grasp taxonomies. We basically study the shape of the human hand while doing tasks, and say, okay, we have the cylindrical grip, the pinch grip, the lateral grip, that are basically grips that are used more often or less often during activities of daily living.

And try to engineer a system that basically covers most of the grasps that are used in doing the activities of daily living. So in this way, you design a device that can perform, let’s say, 90% of the activities of daily living, reducing cost and complexity by 80%, I would say. So I really think that the Mia hand has some limitations, so the fact that you only have three motors, also the fact that these motors can be used all for the thumb position, you only have two positions for the thumb position, so whether it’s opposed or open, you cannot control it actively. It’s a little bit of a limitation for some of the tasks that are required, but probably five motors are a number that I would bet for a robotic hand that is cheap enough to be scaled also from a manufacturing point of view and also a robustness point of view.

Michelle Sun: Interesting. So Mia hand has three motors. Where would the two additional motors be added?

Francesco Clemente: Okay, one could be, of course, on the thumb opposition, because at the moment, I said before that the Mia hand has three motors, one on the flexion of the thumb, flexion of the index, and then the last three fingers are coupled together. Actually the motor that moves the flexion of the index is also coupled with opposition of the thumb, so that with a single motor and a special patented mechanism you can perform the cylindrical grip, but also you can oppose, you can open the thumb, so that you can perform also the lateral grasp. So this movement is coupled with this movement. So when the index finger extends completely, you can move this other joint, let’s say. We would probably decouple those in order to be able to add these independent ones, so that you have a little bit more flexibility and also have a little bit of in-hand manipulation capability.

Probably the other one I would still put in abduction rather than flexion, because there are also studies in the research literature that basically say that having a hand that is able to have abduction movements and two to three degrees of freedom in the flexion of the fingers is equivalent to having a twenty degrees of freedom hand without abduction. So abduction is very important, because it allows the fingers to align, in order to have, so the thumb with the long fingers to align, in order to have a much more stable grasp.

Michelle Sun: I like that. It’s a live design session right here. And you highlight a really important point when it comes to what is the optimal performance level that is also easy to make industrial, on an industrial setting, in mass production, and balancing all these different factors, not just on the spec sheet, like what it looks nice when it comes to the numbers of degrees of freedom. And there’s been a lot of talk about tactile skin, tactile sensors. What are your thoughts about that? They’re not easy to make. They’re very technical and they can be quite bulky as well, depending on what type of tactile sensor that is. Is tactile sensing a prerequisite for these dexterous hands, or do you see that as a luxury?

Francesco Clemente: Probably the best answer from my side is that I don’t really know. I don’t know how much we know about this. So there is a lot of activity going on that demands tactile sensing and very rich tactile information that you use to do dexterous in-hand manipulation. And this, of course, is a specific, let’s say, activity in the manipulation, and in trying to get to something that resembles the human hand. How much this is needed for the tasks that we have to do in, let’s say, an industrial setting, I don’t know. But of course I understand the point of people that want to go deeper into that.

In our devices, we try to have also a balanced approach in this way. We have different amounts of sensors. Some of them are simpler ones. These are basically load cells that are inside of the fingers. So the fact of having the sensors inside of the fingers and not on the fingertip, on the external part, makes them more robust, because not being at the interface with the object, the sensor does not wear out during grasping, so the sensor lifespan is increased. On the other side, it’s more difficult to get position information, so spatial information about the location of the contact point between the finger and the object. For this, you really need a matrix of tactile sensors that are on the fingertip, and this is something that you cannot really go around.

We have partnered with Touchlab, which is a UK company that develops some of the sensors that tie inside of our hands, and that work very well. For instance, we have developed the Mia hand with these sensors, with a matrix of tactile sensors on the thumb, and this allows for richer, let’s say, information to come out from the hand. From a deployment point of view, there is a lot of discussion about the fact that if these sensors are required in, let’s say, a teaching phase of the robot, and then when you deploy the robot, maybe you need simpler sensors, or no sensors, in order to perform the task in a continuous way. But this is something that is still to be discussed and understood.

Michelle Sun: Interesting. Did you mention that the Mia hand doesn’t have tactile sensor on the fingertip, but on other parts of the hand? And did you mention that there’s also sensor on the thumb?

Francesco Clemente: Yes, correct. So the Mia hand has load cells in the thumb, index, and middle finger, at the base of the finger. And then this part, or, yes, at the proximal part of the fingers, I would say, and this allows for the sensors to be sensitive on the whole finger. So it’s not important where you’re touching, the sensor will still pick up deformation. But as I said before, on the other side, you don’t know the location of the contact point.

Michelle Sun: Very cool. If you take the Mia hand apart on a table with all the parts that go into it, motors, gears, sensors, machine parts, can you walk me through, let’s say, the three actuators, like how many, what’s the percentage of the bill of materials that it costs, and what is the most expensive line item in the hand?

Francesco Clemente: Yeah, the most expensive line item, I will say that the motors and the force sensors, so those are the two most expensive ones. Then the mechanical components are, of course, accounting for probably 70% of the cost. So most of the cost is mechanical transmission and the frames of the device. And then you have the electronic components, the PCBs, those ones really are much cheaper to do also at a larger scale, so they don’t really account for a lot of cost in the bill of materials. But of course the motors and advanced sensors, for instance the ones that are a matrix of tactile sensors, I would say, are the most expensive ones.

Michelle Sun: For the semi percent of the transmission related costs, how much of that would be the three actuators and the sensors?

Francesco Clemente: On top of my mind I would say that the motors account for maybe thirty five percent, more or less, of the total cost. Well, we don’t have a lot of motors, so for hands that have a lot of motors this number goes up a lot.

Michelle Sun: I mean, thirty five percent is actually not that out of the range when it comes to like the numbers that I’ve been seeing, like forty to sixty percent of a humanoid is actuators. I’m assuming another bulk of it would maybe, another twenty percent is the touch sensors. Is that how you’re seeing it?

Francesco Clemente: Yes, more or less.

Michelle Sun: And where do these motors and gearboxes come from? You mentioned that you work with Touchlab for the touch sensors. Do you also work with a European supply chain when it comes to these hand scale actuators, or are you sourcing from Shenzhen?

Francesco Clemente: No, we use suppliers from Europe for the motors, because we basically could not find a replacement for these motors that was keeping pace with the performance. So we never switched to other solutions.

Michelle Sun: I see. So the European supplier that you have is able to give you off-the-shelf actuators that can be replaced pretty quickly. Or did you have to do modification in-house to fit the motor for the Mia hand?

Francesco Clemente: No, no, we use off-the-shelf actuators, actually. So usually the motor suppliers have a catalog of motors and gearheads that you can combine together. So basically what we do during design is combine motors plus gearhead, in order to have the best combination to get the torque and the speed that is needed for the application. We have a little bit of customization that we sometimes discuss with the suppliers, but it usually only gets down to the size of the output shaft, or the shape of the output shaft, in order to be able to connect more easily the motor to the rest of the mechanical transmission.

Michelle Sun: Who actually buys a Mia hand today? Who are your customers? How do they split between prosthetics, industrial sales, humanoids, and research labs?

Francesco Clemente: Yes. So at the moment we started selling Mia hand as a research device as well. So before it was also CE marked. Most of our customers are still in the research space. We CE marked Mia hand at the beginning of twenty twenty five. So at the moment the hand is CE marked both as a prosthesis and as an industrial gripper. I would say that we still have at the moment more customers from the prosthetic side, and the industrial side is coming out later. I would say fifty percent research, thirty five prosthetics, and fifteen percent industrial. We are now pushing a lot on the industrial side, because we believe that’s an early adoption market in that case, because industry is not really used to using robotic hands, but they’re accustomed with grippers, but we see rising demand for more flexibility. So we want to push on that.

Michelle Sun: Do you have any visibility in terms of what kind of industrial task your customers use Mia hand for? And what are the customer profiles of these industrial customers? If you were to point at one particular deployment task out in the world and say, hey, this is what actually pays the bills, and what you see as the biggest growth area for Mia hands, which one would it be?

Francesco Clemente: Yes, we are very excited about two use cases. One is testing of human machine interfaces. So all of those objects and goods that have a human machine interface that is designed to be used with the human hand, so by humans, like the interior of cars, so in the automotive sector, at the moment you have at the end of the line a person that goes inside of the car and tests the interior of the car. This can really be automated with a device like Mia hand, that is lightweight and can fit together with a robotic arm inside of the car, and be used to test all of these components. But also like medical devices, like ultrasound machines, ekographs, that have keyboards and mice, that have these interfaces that are relatively advanced, not just simple buttons, but big buttons, levers, a mouse, and also a touch screen. So here we see that Mia hand can be used because of its flexibility, and you cannot really use a simple push-button gripper, let’s say.

And the other one is HoReCa, so hotelier, restaurants, and catering. So serving food and preparing food. In that case, we think that both the flexibility of the device and also the aesthetics of the device are very important, because you have to interact with very different objects while serving food and preparing food. On the other side, you are also in contact with the general public, so this is something, so also aesthetics is very important. We have worked with partners in the past in order to develop robotic cells that prepare ice cream or serve you food, and this is also something that we look for.

Michelle Sun: With Oversonic, the Italian humanoid customer, they use your hand on the robots. Walk me through how that actually happened, from the first email to the first production order. How long does it take, and how many units do they eventually buy?

Francesco Clemente: Yes, Oversonic is an Italian company that develops humanoid robots. So there are not a lot of Italian companies that develop humanoid robots, and there are not a lot of Italian companies that develop robotic hands. So it was relatively simple for us to get in contact. I think that one of the first contacts we got was at an industrial fair in Bologna, so we met them there, and we started discussing possible collaborations, because they found the development of the robotic hand to be very difficult, and we already had a device that could be fitted to the robot. So they were happy to discuss collaboration with us. So they have, at the moment, different use cases.

The one they are exploring the robotic hand, so Mia hand, the most, is with their medical version of the robot. So they have a medical version of Robee, that is the name of their robot, that is used inside of hospitals to support rehabilitation therapies inside of the hospital. So also in that case you need, it’s not an industrial setting, but you need flexibility in order to interact with patients, interact with different objects, rehabilitation devices, let’s say. They were looking for somebody that was resembling the shape of the human hand and that would provide the flexibility in the grasping and manipulation that Mia hand could provide. In this way basically we found this use case that was interesting for them, and they were basically deploying some of their robots in their first pilots inside of this use case.

Michelle Sun: Amazing. Seems like this medical and hospital use is a recurring theme that you mentioned, about like ultrasound machines, and then with Oversonic as well. If we zoom out on the whole dexterous hand market, Psionic also works on prosthetic hands like yourself, and now their Ability Hand sells into robotics, also around like fifteen to twenty thousand dollars. Inspire Robotics got a lot of humanoids using them. And then Shadow, you mentioned, sells mostly into research institutions. And then there’s also Robotiq in Canada that owns the industrial grippers segment. Who do you see as your competitor, and who’s in a completely different business that just looks similar on the form factor?

Francesco Clemente: From an industrial point of view, I think that from the people that you’ve mentioned, let’s say that the Shadow robotic hand is a little bit different with respect to the other ones, because as far as I know, as far as what I see, they’re very much focused on research only. So they’re not selling their device to be used in industry and in prosthetics. So that’s relatively different. And also the hand is very much different from the devices that we develop, so there is also a matter of really different specifications between the two.

Of course, the Ability Hand is the closest one to us, so that’s, of course, a competitor for us. I think that our devices are a little bit different, with different specific strengths and weaknesses, but definitely we are playing in the same space. Robotiq, that you mentioned, is, of course, one of the largest gripper manufacturers in the world. They focus a lot on standard industrial grippers. So again, Robotiq or Schunk or these kind of people, so they have a very strong market in the industrial settings, let’s say. What we would like to do is complement their offering with something that is more flexible. Okay, so our competitors, meaning that probably the clients in the industrial settings that they have and the clients that we have are the same, but we are offering something different.

So let’s say that the device that we offer is complementary with respect to the standard gripper, because the standard gripper is something that is great if you need no flexibility, but you need very high performance, so very fast, very strong, very repeatable, and that’s it. But you only can grip, let’s say, a single object. While we are providing something that is more flexible, still industrial grade, but lower in performance. So we are not able to match the grip force and the speed of a gripper, because that’s a very specialized tool.

Michelle Sun: And it seems like it’s a different performance and different use case as well. So when it comes to grippers, it’s about picking one thing to another place mostly, versus when you mentioned about the use cases that you see, the two main ones with HoReCa and also the human interface, is more about more delicate objects, or you touch some screens or moving a joystick in the ultrasound machine. So is that how you think about it? Like the gripping part, like pick and place mechanism, belongs to the gripper form factor, versus the Mia hand is able to do something that is more about, for example, touching a screen or testing the automotive interior, that needs more of a grip, finger based grip sensation, like action.

Francesco Clemente: Yes, this is one way to think about it. So it’s the specificity of the gripper, so the fact that it’s very much designed for pick and place and for moving objects quickly, while Mia hand, of course, it’s more flexible, so that you can do different things in a specific use case. But not only that, the flexibility comes down to the flexibility of the device. So on one side the gripper is usually designed to be used, so from the start to the end of its lifetime, for the same activity. So you design the fingers of the gripper in order to pick a specific object, because you want it to grip that thing and do that without changing it. While on the other side, a hand, the fact that it’s a hand, gives the device more flexibility. So this device is designed in order to be also relocated to different tasks, so not doing the same thing over and over again. But let’s say, okay, today I need to do these tasks, in two months I need to do something else, and this is something that is enabled by the flexibility of the device.

Michelle Sun: My last question is, when you think about owning that market when it comes to complementing the industrial grippers and supplying to these industrial use cases at scale, what are the biggest blockers when it comes to the manufacturing side? Do you see yourself needing to find, like, a different supplier to build a Mia hand at tens of thousands or hundreds of thousands at volume? Walk me through that future.

Francesco Clemente: Yes, of course, this is something that we are looking into at the moment. So at the moment, Prensilia is basically a design firm. So we have a strong engineering team that designs all of the components, and we outsource the production. And on the other side, we get all of the components, assemble, and do the quality tests in-house. And this allows us to ensure a good quality of the device. Of course, this can run up to a certain scale, let’s say hundreds of hands per year. But if we would have to go to thousands of devices per year, then we need to probably change approach. We are already looking into this in two ways, let’s say: redesigning some of the components in order to use manufacturing processes that allow us to scale production from a manufacturing point of view, but also partnering with partners that help us, let’s say, with assembling part of the device, in order to be able also to scale manufacturing capabilities on our side. So we think that we are not gonna solve this by ourselves, but we are now partnering with strong partners that are specialized in manufacturing, in order to be able to scale production and meet the demand that seems to be needed for the next years.

Michelle Sun: Awesome. Well, thank you so much, Francesco. This is amazing, to chat with you about everything from the design and the mechanism that goes into dexterous hands, but also the market that you’re seeing on the ground. Thank you so much for joining today.

Francesco Clemente: Thank you. I mean, it was a pleasure.

Subscribe now

Subscribe now

Subscribe now

Robot models are not yet like LLMs. We do not use them daily for work or at home. That is why the releases are harder to visualize.

In 2026, every week, or multiple times a week, we see new robot foundation model releases. Many of them claim a GPT-3 moment for robotics. They are all impressive. At the same time, terms like “in-context learning” and “scaling laws” can feel abstract to many (myself included!).

When I read each release, I want to know these four things. What does it actually do, in a humanly relatable way? How does it compare with previous models of the same company and other models out there? What does it unlock? And what’s next?

This format is inspired by the WIRED series that explains complex ideas through different levels of detail. Two of my favorites are its pieces on sleep and machine learning.

In this piece, I share my mental model for answering those questions, one release at a time. Here’s the structure:

• Explain it to a 5 year old

• Explain it to a smart 12 year old

• Explain it to an undergraduate student

• Explain it to an expert peer

• Which bottleneck is it solving

• What it is trying to solve next

I organize by release date, most recent first. I will make this a series as new models ship, so we can compare them over time. If you are new here, subscribe to get the future pieces in your inbox.

For ease of referring to the robot, let’s call it Otto, like autonomous :) These models are Otto’s brain. Each section below gives Otto a different one. “Showing Otto” means giving the model a short recording of someone performing the task. It is closer to uploading an example into a chatbot than to retraining a robot from scratch.

We will go through Skild AI’s S1, Generalist AI’s GEN-1.5, Dyna Robotics’ Dyna-2, Sunday Robotics’ ACT-2, Xiaomi’s Xiaomi-Robotics-1, Robbyant’s LingBot-VA 2.0, and Physical Intelligence’s π0.7.

NVIDIA’s GR00T N1.7 (July 7) expanded the underlying stack rather than demonstrating a new way for robots to learn. We omit it here to focus strictly on task-learning breakthroughs.

Every number in this piece is as reported by the companies themselves.

As always, comments and feedback are greatly appreciated.

Skild AI, S1

Release date: August 25, 2026

To a 5 year old. Dad records a video of himself making pancakes. Otto watches the recording once. Then Otto tries to make pancakes on its own.

To a smart 12 year old. Previously, teaching Otto a new task meant collecting hundreds of examples. Then running a training session for that one task. With S1, Otto learns to make pancakes from one video of Dad doing it. No upgrade to S1.5 or S2 needed. The same model handles each new task, one video at a time.

To an undergraduate student. In the GPT-2 era, each new task needed its own fine-tuning run. Want a model to sort movie reviews into positive and negative? You gather labeled reviews. Then you train. GPT-3’s breakthrough was that examples could go in the prompt itself. Give it “sea otter: loutre de mer” and “whale: baleine.” Then ask it to translate a new animal. No training run. S1 does the physical version of that. The video demonstration is the prompt. The robot executes it with the same model, weights unchanged.

To an expert peer. Skild calls this in-context learning for robot policies. S1 conditions on a video demonstration at deployment time. It does not update its weights. On unseen tasks, Skild reports 66% cumulative per-step success, versus 9% for its language-conditioned VLA baseline, after pre-training on 100,000 hours. One in-context video performed like about 380 post-training episodes. Skild says those episodes would require 50 to 100 hours of teleoperation.

Bottleneck it’s solving. Teaching a new task is expensive. It used to mean a technician’s week of robot operation. S1 changes the unit of teaching to a single recorded example, weights unchanged.

The GPT-3 parallel. GPT-3 made a new language task cheaper to specify. Users could put examples in the prompt. Before that, only machine learning engineers could train a new model. After it, anyone could teach one. GPT-4 and the models that followed did not reinvent prompting. They made it reliable enough to build products on. S1 is reaching for the same shift in robotics. Deploying a new task takes one demonstration instead of a retraining run. Over time, the same model accumulates a long task list.

What it is trying to solve next. One caveat on the numbers first. Skild grades tasks step by step. The score counts the steps Otto completed. If Otto stalls, a human can nudge it forward so the later steps still get scored. So 66% is the share of steps completed. It is not the share of tasks Otto finished alone.

The open questions:

• Can one demonstration reliably produce a clean, unassisted run?

• Does performance hold across unfamiliar objects, layouts, and robot bodies?

• Can a one-off demonstration become a robust production system? The system is more than the robot. It includes how demos get recorded and checked, which tasks get accepted, and how failures get caught. Deployment needs a whole system to work.

Generalist AI, GEN-1.5

Release date: August 19, 2026

To a 5 year old. Show Otto a three-second video of Dad twisting open a jar. Otto tries to twist open the jar.

To a smart 12 year old. Otto can start opening jars from one very short example. Give it a few more minutes of Dad doing the same thing, plus a quick update of the brain. Jar-opening success goes from about 59% to 83%.

To an undergraduate student. A chatbot can pick up a pattern from examples you paste into the conversation. Show it two invoice lines coded the right way. It codes the next one the same way. No retraining. That behavior has now shown up in robotics. GEN-1.5 can learn physical skills on the spot, without being trained on that skill explicitly. It offers two modes. One short physical example for immediate behavior. Or a few minutes of data and a handful of updates for a small, lasting adjustment.

To an expert peer. Generalist describes GEN-1.5 as in-context learning with optional few-step fine-tuning. One 3 to 12 second demonstration produced 59% average success across 10 tasks, with a reported plus-or-minus 10 percentage-point spread. One gradient step on one minute of data reached 66.5% on a held-out task. Ten steps on about five minutes, or roughly 50 demonstrations, reached 83%, with a reported plus-or-minus 9 points. The pre-training set contains 270,000 hours of UMI data and grows by 10,000 hours per week.

Bottleneck it’s solving. Teaching a new task is expensive, and this cuts the bill. Task-specific data and compute drop to minutes and a handful of updates.

What it is trying to solve next. These are short, simple tasks: jars, zippers, pouches. The open question is whether the same pattern extends to long-horizon work.

Dyna Robotics, Dyna-2

Release date: August 10, 2026

To a 5 year old. Dad wears a head camera and records himself chopping tomatoes for dinner. Otto watches many videos like that. Then Otto tries chopping with its own robot hands.

To a smart 12 year old. Dyna teaches Otto how the world changes. When a knife cuts a tomato, the tomato splits into slices. Otto learned all of this from watching videos of people. It can translate human hand movements into robot hand movements.

To an undergraduate student. Most robot models are VLAs, vision-language-action models. They look at the scene, read the instruction, and output an action. Dyna-2 is a WAM, a world-action model. It also predicts how the scene will change next. The actions come out of that same model. Understanding how the world evolves helps Otto pick a better next move, especially somewhere it has never been.

To an expert peer. Dyna calls Dyna-2 a world-action model, not a VLA. It uses a video-diffusion backbone with a mixture of transformers, flow matching, and DiT layers with causal masking. Dyna reports more than 1 million hours of egocentric human video for pre-training. Its evaluation covers 14 benchmark tasks across 3 embodiments, with 39 tasks in total. The reported customer-site pass rate rose from 46% with DYNA-1 to 87% with Dyna-2.

This matters for a second reason. Dyna’s pre-training data is ordinary human video captured on head-mounted cameras. No robot gloves. No teleoperation rigs. That source is cheaper and far more abundant than robot data. In Dyna’s experiments, the video prediction is what made human-to-robot transfer improve as data scaled. Their WAM reached 1.55x the success rate of their previous VLA.

Bottleneck it’s solving. Robot data is scarce, and teleoperated demonstrations are the expensive kind. Dyna learns world dynamics from large-scale human video instead. If that reliably transfers to robots, we need far less of the scarce stuff.

What it is trying to solve next.

• Does human video keep improving robot performance on very different robot bodies, not just human-shaped ones?

• How much robot-specific data is still required after a million hours of human video?

• Can the 87% customer-site result be reproduced by independent operators and across more sites?

Sunday Robotics, ACT-2

Release date: July 17, 2026

To a 5 year old. Otto folds laundry. It folds in homes it has never visited. It folds clothes it has never seen. Even when clothes fall on the floor, Otto picks them up and keeps folding.

To a smart 12 year old. Otto practiced folding laundry in Sunday’s lab. It practiced until it could fold 99 times out of 100. Usually that kind of practice only works in the practice room. A new home breaks it. ACT-2 is the exception. Otto folds laundry in homes it has never visited, with clothes it has never seen.

To an undergraduate student. ChatGPT gets better at your style of writing after a few corrections. It carries the lesson into the next conversation instead of forgetting it. ACT-2 does the physical version. One fine-tuning example teaches a new behavior. The improvement generalizes to unseen homes instead of sticking only to the practice room.

To an expert peer. Sunday uses “solve” for reliable performance across a declared scope at a stated adaptation cost. For ACT-2, the declared task is laundry and the stated per-home adaptation cost is zero. Sunday reports 99 successful folds out of 100 in its lab evaluation. The claim to test is transfer: the same behavior is reported in unseen homes and with unseen clothing, without collecting new data or fine-tuning in each home.

Bottleneck it’s solving. Skills usually break in new environments, so each new home would need its own data collection and fine-tuning. ACT-2 needs neither: no data collection or fine-tuning per home, with training data from people wearing robot gloves.

What it is trying to solve next.

• Laundry is one task. Vacuuming, toy tidying, zippers, and coffee are still in training.

• Does the reliability survive months in a real family home, not just evaluation visits?

Xiaomi, Xiaomi-Robotics-1

Release date: July 16, 2026

To a 5 year old. Dad loads the laundry into the washing machine. He uses a handheld robot hand. A small camera on the gripper records everything. Otto watches many recordings like that. Then Otto loads laundry with its own hands.

To a smart 12 year old. Many people recorded themselves doing chores. They used a small handheld gripper with a camera. Then the robot practiced the same chores with its own body. The recordings are not tied to one robot. So the same model can learn the tasks on different robot bodies. In Xiaomi’s tests, under 10 hours of examples per task reached a 75% average success rate.

To an undergraduate student. This is similar to training ChatGPT on a large amount of general material. Then giving it a smaller set of examples in the format it will actually be used in. Xiaomi pre-trained on broad human manipulation data. Then a smaller amount of robot data taught the model to follow language instructions on real robots.

To an expert peer. Xiaomi-Robotics-1 uses embodiment-free UMI trajectories for pre-training, then real-robot data for post-training. Xiaomi reports 100,000 hours of UMI data across more than 1,700 scenarios, followed by more than 7,200 hours of real-robot data from homes and cross-embodiment datasets. Under 10 hours of demonstrations per task produced 75% average success, versus 40% for the baseline. Under 40 hours produced 85%, versus 53% for the baseline.

Bottleneck it’s solving. Robot data is scarce, so Xiaomi splits collection in two: pre-training on portable-gripper data anyone can record, then post-training on real robots. It also publishes how much demonstration data each task took.

What it is trying to solve next.

• Do these numbers hold outside Xiaomi’s own tests?

• Does the handheld-gripper-to-robot transfer hold across more robot bodies and harder tasks?

• What does a demonstration hour actually cost? Recording is one part. Annotation, setup, and supervision are the rest.

Share Core Matter

Robbyant (Ant Group), LingBot-VA 2.0

Release date: July 9, 2026 (arXiv, revised July 16; peer-reviewed at RSS 2026)

To a 5 year old. Otto carries a glass of water across an obstacle course. While it takes one step, it is already getting ready for the next one. It does not stop to think between steps.

To a smart 12 year old. If Otto stopped after every movement to think about the next one, it would be slow and jerky. The water in the glass would slosh out. It would lose the race. LingBot predicts what will happen next while Otto is still moving. Then it checks what actually happened and adjusts.

To an undergraduate student. There is no clean chatbot parallel here. ChatGPT generates one token after another. The next token waits for the last. LingBot-VA 2.0 predicts and executes at the same time. It predicts future video states and actions in parallel with what the robot is doing.

Picture the AI autocomplete in email apps. It drafts the rest of your sentence from context while you are still typing. LingBot does that, except the sentence is a physical movement. Otto keeps moving while it drafts.

To an expert peer. Robbyant describes LingBot-VA 2.0 as an autoregressive video-action model with closed-loop control. It uses a semantic visual-action tokenizer, a shared observation-action latent space, a sparse mixture-of-experts backbone, and asynchronous inference. Future visual latents are predicted while the current action executes. The model re-grounds each rollout on the latest observation. The paper is peer-reviewed at RSS 2026 and does not publish a task-cost number.

Bottleneck it’s solving. Robots pause between thinking and acting. Plan too slowly, and the robot is slow and jerky. Act too fast on stale predictions, and the robot keeps moving after the environment changes, eg. a chair gets shifted, something rolls into its path. LingBot predicts and acts asynchronously in a closed loop, so Otto keeps moving smoothly while staying responsive. Real environments do not pause while the robot thinks.

What it is trying to solve next.

• Can the robot keep making accurate next-move predictions at speed, over long tasks?

• Can it react when objects around it move unexpectedly?

Physical Intelligence, π0.7

Release date: April 16, 2026

To a 5 year old. Otto has one toolbox. Not a separate box for pancakes, one for laundry, and one for making coffee. After dinner, Otto cleans up the kitchen. It picks different tools from the same box for each job.

To a smart 12 year old. Before, robots often had a separate specialist model for each task. One for opening the drawer. Another for picking things up. Another for wiping the counter. With π0.7, Otto uses one brain for the whole kitchen cleanup. It matches or beats the separate models.

To an undergraduate student. Instead of having one AI for translation, another for trip planning, and another for financial analysis, you use Claude for all three. π0.7 attempts the physical equivalent. One model controls different robots across different tasks. The analogy is not perfect. Claude produces text in one interface. π0.7 still has to translate instructions into safe movements through a physical body. That translation is called grounding. It connects words like “wipe the counter” to the actual forces, paths, and motions that work for this robot, in this kitchen.

To an expert peer. Physical Intelligence frames π0.7 as “a step-change in generalization.” It uses multimodal prompts with language instructions, world-model-generated visual subgoals, speed and quality metadata, and control-modality labels. One π0.7 model matched or beat π*0.6 task-specific RL specialists in the company’s controlled experiments. The post does not state the total training-data scale or the cost of adapting to a new task. That missing number matters when comparing it with the models that publish demonstration hours.

Bottleneck it’s solving. When every task needs a separate model, deployment becomes a model-management problem. Different weights, interfaces, update cycles, and failure modes to track. π0.7 is one general model across tasks and robots, which simplifies that stack. It also shows a robot attempting tasks it was never taught, like operating a kitchen gadget it has not seen before.

What it is trying to solve next.

• Can it do genuinely unseen tasks reliably, not just recombinations of familiar skills?

• How well does grounding hold when the scene, the objects, or the robot body change?

• How much does it cost, in data and in compute, to adapt the unified model to a new task?

Where this leaves us

One note on the tool behind several of these models. UMI, the Universal Manipulation Interface, is a portable gripper with a camera. A person carries it while performing a task. It captures the person’s movement and the scene at the same time. The original open-source design costs about $370 in parts. The 3D-printed gripper is $73. The GoPro camera and accessories are $298, per the paper that introduced it. That is cheaper and easier to scale than teleoperating an actual robot. Teleoperation needs expensive hardware, setup time, and skilled operators. Sunday built its own variant: sensorized gloves at about $200 a pair.

In summary:

• Xiaomi-Robotics-1 pre-trains Otto on humans using handheld grippers, then adapts to new tasks with a few hours of real-robot demonstrations

• S1 lets Otto learn from one example

• GEN-1.5 enables Otto to learn in seconds and minutes for quick tasks

• Dyna-2 teaches Otto from large-scale human video by predicting the environment

• ACT-2 lets Otto do a task reliably in homes it has never seen, with no per-home training

• π0.7 lets Otto do many tasks with one general brain, not a collection of specialized brains

• LingBot-VA 2.0 predicts what happens next to help Otto move quickly and smoothly

The releases are moving the frontier in different directions, but they leave the buyer’s main question partly unanswered: what does it cost to make one new task reliable in production? I’ll dig into that in a follow-up piece in the coming weeks.

In this episode we cover:

• Why fresh produce still has no automated harvester

• The labor math underneath agricultural robotics

• Physics that replaces chemicals, and vision that reduces them

• What actually breaks a machine in a field

• The two-year payback every agricultural machine has to clear

• Who manufactures and services these machines

• Turning farm jobs into ag tech jobs

Danny Bernstein on LinkedIn
Reservoir Farms Website
Watch on YouTube

Subscribe now

Every episode, I sit down with the founders, researchers, and operators building physical AI worldwide, tracing the full stack from components and supply chains to what actually deploys. See all published episodes here.

Chapters

• 00:49 Danny Bernstein, Reservoir, and a technical community for agriculture

• 00:49 Silicon Valley builds seven of everything, agriculture waits for one

• 03:15 Driscoll’s, Taylor Farms, Dole, and a $4.8 billion county

• 05:29 Not AI for luxury, and the three gaps worth building into

• 05:29 Precision surgery by robot, fruit picked by hand

• 07:50 400,000 advertised farm jobs and 182 domestic applicants

• 10:11 What actually breaks a machine in a field

• 12:36 Sonoma, and the plan to cover the top ten specialty crops

• 12:36 Washington, tree fruit, and the Cosmic Crisp patent

• 14:58 Two years to return the cost of the machine

• 17:14 Service teams, dealers, and Andros Engineering

• 21:50 Merced College, and building on a teaching farm

• 24:12 YC gives you tokens, Reservoir gives you a tracto

We are performing precision surgeries using robots, but we still hand harvest fruit. And there were over four hundred thousand of these jobs that were posted in twenty twenty five and only a hundred and eighty-two domestic applicants for those positions. These are jobs that Americans are simply don’t want to do sitting where you’re at today, what are some biggest gaps that you’ve seen that you wish more people are building? And not AI for luxury or

AI’s for abundance, but it’s really AI for resiliency. Why combinator gets a million dollars in tokens from Oken AI, we get access to a tractor.

Danny, welcome to the show. Thank you, Michelle. This is super fun. Thanks for coming out to Salinas. Yeah, I know this is a great day to be out here. So we’re out here, Salinas, 24 acres. 40 acres. 40 acres of you know, like different machines running on the field. So take me back to where this idea first started. Yeah, so Reservoir is a technical community for agriculture, and that’s kind of a new thing. And and so what inspired me is that I actually spent close to 20 years in Silicon Valley.

How impactful technical communities were toward accelerating big discoveries. And you’re in San Francisco and there are technical communities kind of everywhere, whether it’s Y Combinator or around UC Berkeley, around Stanford. And I started to study basically how were technologies developed for the physical world where these technical communities may not exist, and we have significant technical opportunities in areas like agriculture or forestry or land management.

But like where are the startups and how do they connect with each other? And that is like the, you know, 30 second version of kind of how we came up with this idea, which is what if we built like the most robust technical community around Ag and what if we did it in the most important growing region, Archieville, in the United States? Tell me more about Salinas, right? How did you pick this part of California? What’s so special about it? Yeah, so I spent close to twenty years in Silicon Valley. I was part of a startup that sold to Google.

And I was at Google for 10 years and I spent a couple of years at Microsoft. And what really defined my time in Silicon Valley was sort of the behavior of how Silicon Valley developed products, which is one company comes out with a thing, and that thing is the new thing. Let’s say it was Zoom or it was quad code or whatever the first thing is. And then every company basically comes up with their version of that thing. And they basically are competing around what is essentially a solve problem.

And then if you flip it around basically and you look at what are the areas that don’t have that kind of activity, where you don’t have seven companies competing to solve one challenge. And agriculture is like that, where we have a bunch of really significant gaps in technology. And we’ll talk about it. But like an example would be automated harvest for fresh produce. Simple idea. Automated harvesters, automated robots for harvest, specifically in fresh produce where the output is more delicate.

We have practically no solutions. So you contrast that with Silicon Valley, where they’re like seven different ways to vibe code or seven different ways to do video calling, or seven different ways to message someone or email someone, right? We have an abundance of solutions we have to choose. You flip that around and you look at an area like agriculture where you don’t have that abundance. So we started calling it technology as resilience. And then if you look at basically agriculture in particular, and you were to hone in on fresh produce.

which is on micronutrients supply, like our most valuable calories are our micronutrients. And that’s fresh produce. And Salinas ends up being really the fresh produce capital of the United States, if not the world. Yeah, so Monterey County is a five billion dollar ag economy and Salinas is really the center point of that. So where we’re sitting right now, we are within driving distance of the headquarters of Driscoll’s, of Taylor Farms, of Dole Vegetables, all these like incredibly

massive vegetable and fruit production companies. So where do you build the technical community around those products? You do it here. So you’re really in the backyard of all these like the center of all these different agricultural powerhouse. When you talk to like companies like Drisle or, you know, even John Deere that is like in the space for a long time, like how do they view like ArcTech and the autonomous

wave that’s coming to this industry. Yeah, I mean they view it like a imperative, like it’s a must happen. It’s not a nice to have, it’s a must have. And it’s because of labor and input related costs. So chemicals, inputs, fertilizer, all the sort of ingredients of farming are becoming more expensive and more regulated. And then the labor of of agriculture is becoming increasingly expensive and scarce. We cannot pack each of those.

And so then they basically view technology as the only way out of those challenges. There isn’t really an alternative to technology. We’re not going to suddenly have millions of people who want to get into farming. It’s not going happen. We’re not going to suddenly have a pipeline of new chemicals that are going to be better than what was previously there because of anything we have more and more sort of chemical resistance in our crops now. And we haven’t had significant novel new solutions in chemistry come into ag in a very long time. It’s like something like two in the last thirty years.

It’s the view of growers, of operators, of agribusinesses, even of communities, that technology is essential. And so it’s viewed as one of the areas of AI that are not AI for luxury or AI sort of for abundance, but it’s really AI for resiliency. And that’s a different story altogether. And you mentioned about how there are so many gaps in

you know, how AI is solving in agriculture. It’s different from like vibe coding, having different so many different ways to play around with that. Tell me more, setting where you’re at today, what are some biggest gaps that you’ve seen that you wish more people are building? Yeah, sure. So I we really put them in three categories. The first is using physics to replace chemicals. The second is using AI or machine learning or vision.

to reduce chemical use, basically more targeted application of chemicals. So chemical replacement, chemical reduction. And then the third one is automating the jobs that nobody wants to do. And we have a lot of really hard jobs in agriculture right now. So we’re in Salinas, it’s about 70 degrees today. But if you drive two hours east to the Central Valley where we’re in the middle of harvest or kind of in the middle of tree fruit harvest, it could be 105 today. We do not have any automated harvest technology in Stonefruit.

We hand harvest peaches, we hand harvest table grapes. All of those sort of incredible fresh produce items require hand harvest still in 2026. we are performing precision surgeries, you know, using robots, but we still hand harvest fruit. Those are the three categories. Physics to replace chemicals, AI, sort of machine learning and machine vision to reduce chemical use and automating the hardest-to-fill jobs. Remember hearing a case where, you know, with these harvests.

being so such a labor intensive and still dependent on like human going onto the field is actually pretty dangerous, right? Like they get heat stroke and, you know, hours and hours under the sun. It’s actually like not safe to work so many hours. And so that’s where machines can really increase our quality of life and also lower the cost of Yeah, these are jobs that Americans simply don’t want to do. There was a the American Farm Bureau performed a study in twenty twenty five. They looked at

The 400,000 jobs that were posted were basically farm workers, and there were only 182 domestic applicants. Wow. So it’s 0.04% of the 400,000 jobs that were posted received a domestic applicant. The reason they know that is because there’s a visa that is our for foreign sort of the the for the farm worker that’s brought in from a place like Mexico to the United States to do that work. It’s called an H-2A visa.

And it’s how we actually perform harvest today is using imported labor that stays in rural communities for ten, eleven months in specific housing called, you know, H-2A housing or farm worker housing. Each individual H-2A job is posted publicly to ensure that there isn’t a domestic applicant for that particular position. And there were over four hundred thousand of these jobs that were posted in twenty twenty five.

And only a hundred and eighty-two domestic applicants for those positions. It’s wild. And so basically you’re saying that even though we have an unemployment in this country and some of our rural communities had nine, ten, eleven percent unemployment, these are jobs that Americans are simply not even applying for. And there are a variety of reasons why. It’s hard work. We also have a gig economy now that if you don’t want to do outdoor work, you can just drive a car and pick up, you know, Doordash.

So there are different ways to make ends meet now that didn’t exist in previous generations. So yeah, people do not want to work on farms. These are hard jobs. And technology is really for markets like the United States or Japan or Australia where there’s a premium on labor and labor is expensive. Technology is the gap. It’s that’s the solve. Yeah, and definitely in not just America, like globally, you’re seeing people abroad sit in front of their computers or on their phones or driving car.

Delivery, there’s definitely this gap that machines can help bridge. You sitting out here seeing all these machines running on the fields, what are the things or conditions that really break a machine? Is that the dust or is that the humidity, vibration, or you know, what are the surprising things that you’ve seen that are you saying yes to all of those? But if we take a step back, the robotics talent, the physical AI talent, if we think about where it’s originating from, it’s mostly from

PhD labs or from computer science schools or from places like Google or Nvidia or whatever, you have individuals that are, you know, in Palo Alto, in San Francisco, in Cambridge Mass, and if they decide tomorrow that they want to start a agricultural robotics company, like where do they begin? And so a lot of what we did to design Reservoir Farms is we actually thought about that persona. So I’ll very often say to the team, let’s think about the roboticist in Palo Alto and what she or he is looking for in a startup space.

And that’s how we design a Reservoir Farm. So we have forty acres behind us. It has leafy greens, so vegetables and strawberries. Right now those forty acres are it’s sort of farmed like an R and D farm or a test farm. So we have a few acres of iceberg, then a few acres of romaine lettuce, then we have five acres of strawberries, which is easily the most expensive to farm, like exponentially more expensive to farm. They it red gold. We can talk about that. And then we have some brassicas, we have broccoli and cauliflower.

So in one farm they can access behind us, you can access several different ultra high value commodities to be able to build your machine. They can get a lot of what those sorts of conditions are. Our particular farm is flat. It’s actually laser leveled. Most of the farms in Salinas Valley are laser leveled. If you go thirty minutes north, Castroville or or Prendale, you’ll see a lot of strawberry fields that are on s pretty severe slopes. And a lot of vineyards are on slopes and sometimes

winemakers actually love, you know, kind of the the challenged vineyard, you know, like that the this these Pinot Grapes had to struggle and therefore you can taste this or that. So there are a lot of different conditions that introduce a lot of variety and a lot of complexity. Sometimes you’re in a GPS-denied environment. Sometimes you have very low connectivity. So we see a lot of different things that come up. mud, muck, mire, all the above exist here. So this is like really a proving ground for physical AI. And there’s a feeling in ag in particular.

that if you can make it here, you can make it anywhere. And the vision is to have not just in Salinas, but also in like wine country and tell me more about that. So we opened Reservoir Farms in Salinas in March. And then a month later we opened our second location in Sonoma County. So Salinas is two hours south of San Francisco, but only about an hour south of San Jose. So it’s very accessible to NVIDIA to Google, etc. And they’ve all been here, which has been great.

And visited us. And then we’re in Sonoma County, which is an hour north of San Francisco. So we have startups that are splitting their time between YC and Founders Inc. and Reservoir, which is pretty cool. And that makes sense, right? Because YC, last time I checked, didn’t have a farm. So it’s useful to them. So yeah, we opened in Wine Country. We have fifteen acres of Pinot Grapes there. And it’s a similar structure that a startup can get onto our farm and be testing in rows and vineyards. And our plan is to

Roll out farms to cover at the very least the top ten US specialty crops by farm gate value. So wine grapes, table grapes, almonds, leafy greens, brassicas, carrots, tomatoes, blueberries, strawberries, and maybe I forgot a few more onions than they’re onions. And so then a startup that is thinking about working across ag, which is fragmented by crop, which is part of the complexity of this market.

they can work with Reservoir and then in one year or what a year and a half, they can actually access lots of different crops and test in lots of different environments. So Bonsai Robotics, which is one of the startups here that our fund actually invested in, they in the same week tested at our wine country location on vineyards and then tested on strawberries here in Salinas using the same machine. It’s pretty fun. So you’ll cover the whole supermarket. Yeah, we like to yeah we’ll open in Washington State sometime in the next six months.

And that will be tree fruit and blueberries. And we love Washington because it’s really like the intellectual and industrial epicenter of tree fruit and blueberries. The Cosmic Crisp Apple, which is my favorite apple. The Cosmic Crisp apple is legally by rules around the intellectual property, only can be grown in Washington. So yeah, so it’s licensed by Washington State University, which has the patent, and they actually license it to Washington growers.

And they could only grow it in Washington, I think in Clint 2032. Wow. But it’s an amazing, amazing apple. So then if you think about where would you build a startup in your startup space or where would you build a technical community that was focused on tree fruit, you would of course be in Washington, which is 60% of the apples in the United States. And then blueberries actually like there is extraordinary efficiency for blueberries in Washington relative to pretty much every other state. They grow more more inexpensively.

And similarly to these other environments, there is a lot of automation gaps. We don’t have hand harvest. We don’t have automated harvest in tree fruit. We don’t have automated harvest in fresh blueberries. So there’s a lot of opportunity, a lot of like big like really meaningful and high value tech gaps in these areas. While we’re doing our preparation, we talked about the operational layer of Archtek, right? And so there’s so many things that

goes in like beyond building the machine, beyond the road model and actually getting machines on the ground and keep them working. So talk to me more about that and what you’ve observed and what the gaps have been. Yeah, I mean the ab tech CEO has a lot to think about for sure. I think it really attracts it’s a very high demand sort of job. We tend to see more senior CEOs here because of that, not exclusively, but because of it, it requires a lot of resilience, a lot of humility. I mean there are a couple things. One is that the most

Critical thing is that they’re delivering a machine that really must deliver ROI within two years, maybe three, ideally one. So they are dealing with a very discerning customer that knows their numbers and is really kind of commanding a pretty high bar. It is a much more sort of sophisticated and informed sort of customer than I think a lot of roboticists are gonna be accustomed to. And that’s exciting. I mean, I think it’s a big challenge, but it’s exciting. And if it and if you can make it work, like

Carbon Robotics, for example, their laser weeder, they’re now doing a hundred million ARR. It’s great. So that’s a company that is is starting to scale in their other examples. But an AgTec CEO needs to think about the logistics of their machine, how they get it from here to there. They have to think about the servicing of those machines. Are they going to be servicing in-house with their own team? So Carbon Robotics, for example, is built an in-house service team. Bonsai is beginning to build one. And then but Bonsai also has what’s called dealer relationships.

So there are ag dealers that sell, may sell John Deere Tractors, and then they also sell ag tech as well, sort of more emerging ag tech. So they have to think about those sorts of channel partnerships, both for sale and service. They of course have to think about manufacturing and a manufacturing supply chain. So a lot of companies will do their own manufacturing for like maybe the first two or three years and sort of manufacture out of a warehouse, or they’ll work with specialized manufacturing firms or actually domestic.

that they come in a contract with, like Andros Engineering, which I’m sure you’ve never heard of. But if you go to Andros, which is in Passa Robles, our friends at Andros, you will see them hand assembling half a dozen different extraordinary AgTech machines that are all software enabled, AI enabled, and they’re building them. And there’s another one, GK Machine, which is in Oregon, which is doing something similar. And so you have these very specific manufacturing pipelines.

And addition to all the supply chain considerations they have to think about. So it’s a very opsy job. And then they have to think about how are they marketing their products and how are they pricing their product are they doing? Is it lease based? Is it service based? Is it, you know, selling the machine, etc.? So it is demanding that way and I would say I have a lot of respect for the ad tech CEO. It’s it’s not for the faint of heart for sure. I think one of the things that really stood out to me is how we’re farm is

currently operating like very closely with the community college around here. And I think that also touch upon the shifts in the labor market that we like as more autonomy is going to the industry, like what does the future like architect labor market look like? And what do we need the skills to be? And how do we train up the next generation of people operating on the field? No, it’s a big one for sure. And and we’re not the only ones talking about it, but

We think about we are building a technical community for agriculture and we’re doing it in place. Like we’re doing it in rural regions of the United States. So rural Washington, rural Arizona, rural California, and then we’ll go beyond that, rural New York, rural Florida, et cetera. And it’s difficult to imagine convincing the San Francisco engineer to want to be in place here. Like their ties are there, they’re part of that ecosystem. They’ve gotten accustomed to creature comforts of being in San Francisco. So when you think about

Who is going to be very affinitized towards rural California? Well, it’s the people who are from there. And so we have a labor pool here. We have a we have a high tech labor pool here that we need to work with the community to to train up. So we have to think about what are the hiring needs of a Carbon or a Bonsai or a Verdant Robotics. We have to work with the community colleges that are in region. It’s also the California state schools. It’s a little bit with University of California as well.

And we have to have a really like ongoing, I’m talking like weekly, sometimes daily dialogue about the sort of opportunities that go between those firms as we transition ag jobs to ag tech jobs. It’s one thing to be servicing and implement on a tractor that’s dumb, that’s just like mechanized. It’s like you just let’s say it’s like a big, dumb tilling machine, no offense to a tilling machine. But then going into something that’s smart, that’s AI enabled, that might have, you know, twenty four GPUs on it or something like that. Like we’re talking a

another level of complexity. We have a machine out there that’s an automated harvester that’s pre-commercial that has eight robotic arms. How many cameras? I don’t know, but like maybe a few dozen cameras. So there we’re talking about like a different different level of service orientation. So what we did is that when we began to build Reservoir Farms, the concept for it, which was two years ago, one of the first conversations we had was with the local community college here, which is called Hartnell College. It’s been around for over a hundred years.

And they are basically a feeder into agribusinesses here. Those have been historically like pest control advisors or you know, agronomists, they might be folks who help you figure out what kind of chemistry you should be using on your farm or figuring out kind of how to problem solve around disease control or weed control. But now those same sort of pipelines of agricultural talent need to be thinking about AI enabled machines and robotics with a hundred different degrees of freedom.

So that’s a really different thing. So how do we do it? We basically set an MOU with them about how we work together. And then fast forward to today, we have eight interns with our startups here from Hartnell College. We have eight community college interns at a deep tech, high tech startup incubator, which is incredibly cool. And so we view it like that there’s a strategic role of inclusive innovation in physical AI. If you’re building physical, if you’re building AI for the real world and you’re doing it in place.

You have to be thinking really early on about your labor pool. And we encourage companies to think about it. But then we as the sort of convener, as the technical community, we, you know, have to think about it as well. So we’re learning a lot, but it’s a it’s a neat and new mode. Yeah. And it’s really about kind of like mixing the like chemistry, like learning about the

the pesticide side of things, but also like how it interacts with like the autonomous and the world model side of things to use. It’s real. It’s like community integration. Exactly. So in the Central Valley of California, which is about two hours east of here, we are launching a version of Reservoir there next year. And we decided to launch it actually on a community college campus. Wow. So there’s a school called Merced College, which is Merced is basically two hours south of Sacramento and then two hours north of Bakersfield.

So then this really substantial stretch of California’s inland, Eastern California, let’s say, middle to eastern California, you have several of the largest ag economies in the country. So Kern County and Fresno County are both eight billion dollar ag economies annually. They’re massive. And so we thought, you know, we should probably build a Reservoir there. This is tree nuts, table grapes, processed tomatoes, carrots, etc. It’s like a really, really big part of the agricultural industrial sort of state.

We thought a lot about what would be the ideal scenario for building Reservoir there. And we decided that doing it at a community college farm made a lot of sense because community college farms are multi-crop because they’re trying to train students. So on one farm you have tree nuts, table grapes, citrus, stone fruit, processed tomatoes, melons, peppers, etc. You would never have that in an industrial farm. So then a startup can be thinking, How am I building my machine? Not just for one crop and one task.

Each new crop a startup adds is more TAM for them. For sure. So it’s like you can think add a crop, add TAM, add a task, add TAM. So this particular farm in the Central Valley is perfect for us. And rather than building, let’s say, you know, ten miles from that farm, we worked out a public-private partnership with a college to build Reservoir Farms right there. So we’re doing some different things and working on it a lot. I’m curious in terms of these inters working at your farm and also with all these startups that

being debated by your community, like what are the things that they are learning that they cannot learn in a a community college inside their classrooms? Yeah, it’s a great question. I mean they’re they’re learning about the very specific intricacies of of machine maintenance and repair. They’re beginning to see sort of hands-on design decisions about machines. They’re also learning the equipment and sort of how it’s used in the real world. So we have site managers and site coordinators that are employed by Reservoir.

all of which are are ag engineers, mechanical engineers that that know the way around a shop. And so they’re getting that pretty extraordinary hands-on experience. But a lot of times the startups themselves are. So if you think about what are the characteristics of a startup founder in ag tech, they may be a software wonk, but they may not be a machinist. And so very often we are teaching the re startup founders how to drive a tractor, how to work a forklift. One of the unique things about Reservoir is that we have tractors here.

for them to build on and test on and add sensors to and do all the things that they would need to do in an ACTEC context. I think we’re the only startup incubator in the world that has access to John Deere tractors for as part of our perks. You know? So like you know at White Combinator gets, you know, what you know a million dollars in tokens from OkenAI, we get access to a tractor. So I think it’s pretty unique. Take me five years out, right? So where do you see worse of a farm at? Like how many farms would you have and what are some

New use cases that you’re excited about. Yeah. I mean, I think we’ll certainly have farms across the American West to cover the largest crops. We’ll have expanded nationally. I think we’ll have also established Reservoir Farms and sort of the major international hubs that have some characteristic chal similar characteristics of challenges. So basically every first world market has a labor shortage. If they are also growing their own specialty crops growing if they’re in agricultural market, like sort of exporter versus importer, they have significant challenges. So it makes sense to do Reservoir elsewhere.

And there’s a lot of the due tech ecosystems that have formed around those areas, like Australia and the Netherlands and Israel and other places. So it just makes a lot of sense for us to be there. We would hope to see several solutions that are plugging gaps and harvest automation. I would think by like 2030, 2032, that we’re automating more and more of harvest, that we have automation of berries and leafy greens and vegetables and brush. So that’s a big one. I we also want to see the scaled adoption of sort of physical tools to replace chemistry.

So that’s UV light for disease control, that is laser weeders, that’s electric weeders, and a bunch of other novel solutions that we’re expecting as well. So that area, which is basically using atoms instead of chemistry in order to control disease, control weeds, that’s good for the world. It’s also necessary because we’re gonna see increased regulation around the use of chemistry, and it’s gonna become increasingly expensive to do chemistry at scale. And so

That’s something that we’re tracking and you know you’ll have to come back hopefully sooner than I. So awesome. Yeah. Well, thank you so much, Danny. This is awesome. I appreciate it, Michelle. Thank you so much. Thank you.

This morning NVIDIA announced Jetson Orin Nano 2, the entry tier of its robotics compute line. Put simply, Jetson Orin Nano 2 is a distribution strategy: NVIDIA’s move to lock in the physical AI value chain from the beginning of a product’s lifecycle. We will discuss in detail later in this piece.

Before that, we will first go over an overview of NVIDIA’s three-computer architecture in physical AI, and how the company is positioning for the next wave of growth in robotics and distributed computing. While NVIDIA frames its physical AI stack as three computers, two further layers come into play when building a working robot. Below training sit the data operations layer, and actuation and motion control. NVIDIA partners on both instead of building them, and we explore why. The entry tier is also becoming a crowded place, so we close on competing offerings from Qualcomm, AMD and others.

Sections

• What is NVIDIA’s three-computer architecture?

• What is a Jetson, and what do the numbers on the spec sheet mean?

• What did NVIDIA announce on August 25, 2026, and what gap does it fill?

• How does Jetson Orin Nano 2 compare with Thor and the rest of the line?

• Which layers of the physical AI stack does NVIDIA not sell into?

• Why doesn’t NVIDIA make robot actuators?

• Who competes with NVIDIA at the edge?

• What this means for an investor

• Claim ledger

What is NVIDIA’s three-computer architecture?

Deepu Talla, NVIDIA’s Vice President of Robotics and Edge AI, framed it this way on the launch briefing:

“Physical AI ultimately is the largest opportunity in front of us. And we’ve been working on this problem for over a decade. And we are starting to see now real uptake because of improvements in model accuracy and intelligence.” [1]

AI is moving into the physical world: factories, warehouses, farms and city infrastructure, and each of those settings needs models that run on the machine, not in a data center. NVIDIA reports more than 3 million developers building on its robotics stack and more than 10,000 companies on Jetson, either shipping a product or developing one. [2]

NVIDIA divides physical AI development into three steps, and assigns a computer to each. Train, simulate, deploy. A builder trains a policy for the robot to perform a task, tests that policy in simulation, then deploys it onto the machine.

Why does a robot need more than one computer? A robot itself is not one computer. Policy inference, perception, sensor fusion, limb-local control and safety monitoring are different workloads with different latency budgets.

A beginner’s guide to NVIDIA’s physical AI stack

Here is what NVIDIA sells at each step.

Train: DGX. Integrated GPU systems that live in a data center or are rented from a cloud provider. Training a robot policy on large volumes of video and telemetry carries the heaviest compute demand of the three steps.

Simulate: OVX and RTX Pro. Graphics-heavy systems built for physics simulation and rendering at scale. This allows a policy to be tested across thousands of virtual attempts before touching hardware. This is super important as robot hardware is costly to replace and breaks easily.

Deploy: Jetson Thor and Jetson Orin Nano 2. The modules that go onto the robot itself. These have to be small, fast and power-efficient, because a watt spent on compute is a watt not spent on runtime.

Around the hardware, NVIDIA gives away a large body of software. Two model families run across all three computers:

Cosmos. Open-weight generative world foundation models. This also includes tokenizers that compress video into the units a model predicts over, and a pipeline that turns raw video into training-ready synthetic data. It sits alongside Omniverse on the simulation computer.

GR00T. A robotics foundation-model family and data pipeline. It is trained on DGX, validated through NVIDIA’s simulation suite, and executed on Jetson.

For simulation:

Omniverse. A development platform for 3D simulation and digital twins, built on OpenUSD as its scene format. Omniverse Kit is the SDK.

Isaac Sim. A robotics simulation application built on Omniverse Kit, adding robot models, sensor models and ROS 2 support.

Isaac Lab. A robot-learning framework on top of Isaac Sim, for reinforcement and imitation learning at scale. It replaced Isaac Gym and ORBIT.

For deployment:

Isaac ROS. In NVIDIA’s words, “a collection of NVIDIA CUDA-accelerated computing packages and AI models” for building robotics applications. [3] GPU-accelerated ROS 2 packages, in plain terms.

cuVSLAM. A CUDA-accelerated stereo visual-inertial SLAM library, renamed from ELBRUS, that estimates a robot’s position from stereo cameras and an optional IMU. [4]

What is a Jetson, and what do the numbers on the spec sheet mean?

Jetson is NVIDIA’s product line for the deploy step, the modules that go onto the robot’s body. It is the embedded line of NVIDIA’s physical AI chip offering, built for machines that move: robots, drones, cameras and industrial equipment.

This is confusing, but Jetson Nano and Jetson Orin Nano are different products. Jetson Nano launched in 2019 on the Maxwell architecture and was the entry board. Jetson Orin Nano launched in 2023 on Ampere. The line goes Jetson Nano, TX2 for drones and small robots, Xavier for autonomous machines, then Orin and Thor. [5] And Orin is a family, with Orin Nano at the entry, Orin NX in the middle and AGX Orin above it.

A cheatsheet for compute terminology

TOPS or TFLOPS, with the precision attached. Trillions of operations per second. Precision is the number of bits used to store each number the model works with: FP32 uses 32 bits, INT8 uses 8, INT4 uses 4. Fewer bits means a smaller memory footprint and faster arithmetic, the trade-off is accuracy. The same chip produces roughly double the TOPS at INT4 that it produces at INT8.

Dense or sparse. A sparse figure assumes the model has been pruned so the hardware can skip zeros. Vendors tend to quote whichever number is larger. It’s important to check which density they’re referring to.

Memory capacity and bandwidth, in gigabytes and gigabytes per second (GB/s). For language and vision-language models, memory bandwidth, not raw compute, sets the limit on speed. This is why a part can have impressive TOPS and still produce tokens slowly.

Power envelope, in watts for the whole module. For a robot, this is battery life, and it converts directly into runtime between charges.

Form factor and module compatibility, in mm. Whether a new module drops into the existing carrier board determines whether an upgrade is a firmware update or a hardware redesign.

Supported-until date, calendar year. How long the vendor commits to shipping the part determines the length of amortization years on a design, which sets the cost structure of the chip.

What did NVIDIA announce on August 25, 2026, and what gap does it fill?

NVIDIA announced Jetson Orin Nano 2, the entry tier of its offering, on August 25, 2026. The specifications are below, all marked preliminary and subject to change on NVIDIA’s own materials: [6]

Compute: 78 TOPS INT8, Ampere GPU
CPU: 8-core Arm Cortex-A78
Memory: 8 GB LPDDR5x, 120 GB/s
Power: 15 W to 40 W, entire module
Compatibility: full-stack NVIDIA software, form-factor compatible
Availability: 1H 2027
Price: not disclosed. No price appears in the press release or the briefing deck.

Ampere vs. Blackwell. Ampere is NVIDIA’s 2020 GPU architecture. Blackwell is the current architecture, released in 2024. Here are the full Jetson generations: Nano on Maxwell, Xavier on Volta, Orin on Ampere, Thor on Blackwell. [5]

Orin Nano 2 is the first new entry-tier design in 3 years. Talla confirmed Orin Nano 2 is new silicon, still on Ampere, with the GPU optimized further for inference. [7] This means NVIDIA paid for a new chip design on a 5 year old architecture, instead of bringing Blackwell down to the entry point.

Jetson Orin Nano Super, launched December 2024, used the same Ampere silicon as the original Jetson Orin Nano from 2023. A software update increased performance from 40 to 67 TOPS, and lowered the price from $499 to $249. [8] Orin Nano 2 is an actual new chip at that tier.

Before Orin Nano 2, NVIDIA’s compute offering was strong at the high performance end and stale at the entry end. Built on Blackwell, Thor put a data-center-class part on the robot at 2,070 FP4 TFLOPS, 128 GB and 40-130 W. [9] The entry tier stayed on Ampere, at 40 TOPS on the original Orin Nano and then 67 on the Super. Both were too slow for a VLM.

NVIDIA shared real time model performance for Orin Nano 2 against Thor. Against a 20 tokens-per-second line the company labels real-time interactivity, 5 of the 10 models tested clear it. Every model at 8B parameters or above falls short, at 10, 12 and 14 tokens per second. The 4B tier straddles the line, at 17, 18, 21 and 21. The 2B tier clears comfortably, at 34, 39 and 42. Cosmos Reason 2 8B, NVIDIA’s own robotics reasoning model, produces 14 and misses NVIDIA’s own threshold. [10]

How does Jetson Orin Nano 2 compare with Thor and the rest of the line?

Orin Nano 2 is the entry offering of NVIDIA’s edge compute Jetson family. It replaces Jetson Orin Nano Super and shares its form factor. That means a builder can move to it without redesigning the carrier board.

On NVIDIA’s own benchmarks, Thor delivers 3.8x to 6.1x the throughput of Orin Nano 2. Thor executes 31B and 35B models above the real-time line. Orin Nano 2 tops out around 4B. [11]

Qwen 3.5 4B: 18 to 68 tokens per second, 3.8x
Gemma 4 E2B: 34 to 133, 3.9x
Gemma 4 E4B: 17 to 95, 5.6x
Qwen 3.5 9B: 12 to 73, 6.1x

Thor is the brain. Orin Nano 2 is the part for a gripper, a sensor head, an AMR base, an inspection box beside a line. NVIDIA demonstrated two Reachy Mini robots operating from a single Orin Nano 2, each executing a small language model, speech recognition and vision at once. [12]

Which layers of the physical AI stack does NVIDIA not sell into?

While NVIDIA frames its physical AI stack as three computers, there are more than three layers involved in building a working robot. Below training, there’s the data operations layer, the collection, curation, labeling and fleet observability that feeds everything above it. And underneath the machine sits actuation and motion control, the motors, reducers, drivers and the control loops that move the machine.

NVIDIA partners on both layers. Deepu Talla said on the launch briefing:

“So NVIDIA, we build these three computers, we sell these three computers, we build the acceleration libraries and reference models and software on top of these three computers for data generation, for training, for testing, and lastly, for runtime deployment. And then we work with literally every robot company on the planet, whether they’re building a physical robot, whether they’re building a robot brain, whether they are in the field of actuation or sensing, whether they’re providing just data collection services, it doesn’t matter what layer they’re in the ecosystem, we partner with every layer of the ecosystem. NVIDIA does not build robots, but we enable every company building this technology.” [13]

Talla names both missing rows himself, actuation and data collection services, 40 seconds after saying which layers NVIDIA builds and sells. He shared that actuation is becoming an extremely important part of the solution and that NVIDIA partners with several actuation companies. However, he didn’t name any companies. [14]

This gives us a rule that holds across NVIDIA’s physical AI strategy, and we will call it the silicon pull test: NVIDIA occupies every layer that creates demand for its own silicon, and partners on every layer that does not.

In data operations, fleet observability and semantic search over robot logs do consume GPUs. Those GPUs sit in a data center. NVIDIA already sells them through DGX and cloud providers. NVIDIA keeps giving away models and simulation, as long as it seeds and expands demand for its silicon. If data curation moves onto the robot, it creates a new compute socket on the machine, then the layer becomes NVIDIA’s territory.

Why doesn’t NVIDIA make robot actuators?

An actuator converts electrical power into controlled motion at a joint. A robot-grade one packages a motor, a reducer, an encoder and a driver into a single unit, and a humanoid carries 25 to 50 of them depending on hand complexity. [15]

Three reasons why NVIDIA doesn’t make robot actuators today.

1. Actuators fail the silicon pull test. Actuators do not create demand for new GPUs.

2. Actuators and chips have structurally different margins. Manufacturing reducers requires high capex and high fixed costs. Harmonic Drive Systems makes the strain-wave reducers that go into precision robot joints. In its FY2026 annual report, gross margin came in at 30.4%, up from 26.7% the year before. [16] Nabtesco reported its Component Solutions segment, precision reduction gears, at 10.6% operating margin on ¥45.5bn of sales in H1 FY2026. [17] NVIDIA’s gross margins are in the 70%s because of its fabless business model. [18] These are structurally very different businesses.

3. Cost per unit for actuators does not come down with scale. NVIDIA’s cost per unit of compute decreases with every process and architecture generation. That improvement comes from economies of scale over a fixed cost. It lets NVIDIA cut price per unit of compute, widen the market and hold gross margin in the 70%s at the same time. A harmonic reducer is the opposite. Its costs are materials, grinding tolerance and fatigue life. These are all variable costs. Increasing volume does not change the cost basis. Silicon can achieve annual price reductions. Precision mechanics cannot. That’s why we see a player of HDS’s scale achieving 26.7% gross margin, and break-even operating profit.

That being said, actuation increasingly requires real-time motion control, and that control needs silicon. Which may be NVIDIA’s playground.

Who competes with NVIDIA at the edge?

The entry tier of edge computing has become a crowded space. Qualcomm is shipping now. AMD released a Thor-tier product in July with a decade-long supply commitment. Hailo and Horizon are selling into the tiers on either side of Orin Nano 2. The common theme is that competitors are putting real-time motion control silicon on the same board as the AI accelerator, while NVIDIA keeps that function on a separate part.

Read more