A human hand has around 27 degrees of freedom. Most robots on the market today get by on 6. That gap is why AI can pass a bar exam and still fumble pulling a $5 bill out of a wallet. It is called Moravec's paradox, and nowhere does it bite harder than at the very end of a robot's arm.
At ICRA 2026 in Vienna, one of the more talked-about booths did not belong to a big-name humanoid. It belonged to a small Shenzhen company called Wuji Technology, showing off a robotic hand that costs a fraction of what the industry's gold standard costs, and a data glove built to solve a problem almost nobody is discussing publicly.
The Hand Problem Nobody Talks About
Modern AI is a monster in the digital world. Text, images, logic, it handles all of it well. The physical world is different. Picking up a coffee cup or plugging in a cable is something a 4-year-old does without thinking. It stays shockingly hard for machines.
There are two ways engineers have tried to close that gap. The first is tendon-driven design, where motors sit up in the forearm and cables run down to the fingers, mimicking how human tendons work. It looks elegant on paper. In practice, cables stretch, wear, and snap, and the gap between how a hand behaves in simulation and how it behaves in the real world grows fast.
The second approach is direct drive, where tiny motors sit inside the fingers themselves. It is much harder to engineer. When it works, every finger becomes its own self-contained unit, and the whole hand behaves more predictably.
Where Tesla, Boston Dynamics, and Figure Actually Stand
Tesla has heavily promoted the Gen 3 Optimus hand. What is missing is a clean, unedited demonstration of it reliably manipulating random objects the way a person would. There are plenty of polished clips and few raw "it just works" moments.
Boston Dynamics' electric Atlas takes a different path entirely. Its hands are deliberately simple, rugged grippers built for heavy industrial work rather than delicate manipulation. That is a smart trade for a factory floor and a poor fit for anything requiring fine dexterity.
Figure is arguably ahead of the American pack. The Figure 03 hand runs 16 degrees of freedom and pairs with the company's Helix AI model, already doing real work on BMW's Spartanburg production line. But hands are also the part of a humanoid robot that wears out first. Run these machines shift after shift and the fingers take the beating, turning replacement into a constant, expensive chore that rarely makes it onto a keynote slide.
Many Chinese humanoids keep costs down with a basic 6-degree-of-freedom gripper, most commonly the Inspire RH56. It is affordable and everywhere, and it is also coarse. Fine assembly and in-hand manipulation are out of reach for it.
The gold standard remains the Shadow Hand out of the UK, a tendon-driven design with 24 movements. It costs upward of $100,000, with some configurations closer to $300,000. It is also delicate enough to mostly stay in the lab.
The $5,000 Hand From a Motor Company in Shenzhen
Wuji Technology did not start out building hands. Founded in 2019 and backed by parent company Pan Motor, it began as a motor company focused on high-torque-density electric motors. That groundwork turned out to be the advantage that let it attempt direct drive, where most competitors avoid it.
The Wuji Hand 2 is a fully direct-drive, biomimetic hand with 20 active degrees of freedom, four independent joints per finger, and a thumb that properly opposes the way a human thumb does. There are no tendons to stretch or snap. Every joint runs at a 1,000-hertz control rate, and the design includes thermal management meant to keep it running through long stretches of use.
| Spec | Wuji Hand (Gen 1 / Gen 2) |
|---|---|
| Degrees of freedom | 20 active, 4 joints per finger |
| Weight | Under 600g (Gen 1); around 0.5kg (Gen 2) |
| Fingertip force | ~15 Newtons |
| Durability | Rated 300,000+ grasp cycles; internal testing past 1 million |
| Price | ~$5,000 (Gen 1 only; Gen 2 pricing not disclosed) |
Compare that Gen 1 price to $100,000 and up for a Shadow Hand. It is not an incremental improvement. It is a different category entirely, assuming the durability claims hold up outside a booth demo.
The Bigger Bottleneck: Data, Not Motors
A cheap, tough, dexterous hand solves half the problem. The other half is data. AI models learn from mountains of examples, and for text and images, the internet already provided that. For physical manipulation, the precise feel of fingers rolling a screw or peeling a sticker, that data barely exists anywhere.
Wuji built a data glove to capture it. The glove tracks a human hand across more than 500 touch points, streaming tactile and pose data at 120 frames per second. A person wears it, performs a task, and the system records exactly what the fingers did and felt. Because the glove captures motion in a general way rather than tying it to one specific robot, the resulting data is described as embodiment agnostic. In practice, that means the lessons can transfer across different hands and robots, not just Wuji's own.
My Take
Everything interesting here rests on numbers Wuji itself supplied at a trade show booth, not on a peer-reviewed test or a shipping product with a warranty. That does not make the claims false. It does mean "rated for 300,000 grasp cycles" belongs in the same mental bucket as any other manufacturer spec sheet until someone outside the company runs it into the ground.
What actually convinces me is the sequencing. A motor company spent years on unglamorous torque-density work before touching a finger joint. That is not the origin story of a hype machine. It is the origin story of people who solved the boring problem first, which tends to be the part that quietly outlasts the flashy demo.
The data glove is the smarter bet of the two products. Hands can be copied. A cheap, scalable way to harvest embodiment-agnostic manipulation data is harder to replicate, and it is the actual scarce resource in this race.
- Robotic hands, not AI models, are the bottleneck holding back general-purpose humanoids.
- Direct-drive actuation (motors inside each finger) avoids the calibration problems of tendon-driven designs like Tesla's Optimus and the Shadow Hand.
- Wuji's Gen 1 hand reportedly sold for around $5,000 against $100,000+ for a Shadow Hand, though Gen 2 pricing is unconfirmed.
- The accompanying data glove targets the deeper problem: a near-total lack of physical manipulation training data.
FAQ
What is the Wuji Hand?
It is a direct-drive robotic hand made by Wuji Technology, a Shenzhen-based company that started as a motor manufacturer. It runs 20 active degrees of freedom, with motors built into each finger rather than driven by cables from the forearm.
How does the Wuji Hand compare to Tesla Optimus's hand?
Optimus uses a tendon-driven design, while Wuji uses direct drive, which its makers say removes cable stretch and wear issues. No independent side-by-side test of the two has been published, so the comparison rests on each company's own claims.
Is the Wuji Hand available to buy?
The first-generation Wuji Hand was reported at around $5,000. Pricing and wider availability for the newer Wuji Hand 2, the version with the specs discussed here, have not been publicly confirmed.
Why do robotic hands matter more than AI models in humanoid robots?
Because manipulation, not conversation or reasoning, is where humanoid robots consistently fail today. A robot can pass a bar exam simulation and still struggle to reliably pick up a single object it has not been trained on.
Conclusion
None of this is confirmed by anyone outside Wuji Technology yet. No third-party lab has run the Wuji Hand 2 through independent durability testing, and no price has been attached to the version with these specs. What is worth watching is whether a small motor company that spent years on unglamorous actuator work turns out to have solved, quietly, the problem the bigger names have been demoing around for two years.
0 Comments