Can Robots Learn the Human Hand? A 20-DOF Dexterous Hand Enters Japan’s Physical-AI Race

When artificial intelligence moves off a screen and into the physical world, one of the hardest engineering problems is not the brain. It is the hand.

HumansX, a Tokyo physical-AI and robotics company, began official Japanese sales on Oct. 6 of Wuji Technology’s five-finger Wuji Hand 2 and its tactile Wuji Glove. The hardware is being sold as a data pipeline as much as a robot component: capture human finger motion and touch, retarget that motion to a robot hand, then use the resulting demonstrations to train imitation-learning or Vision-Language-Action models.

What 20 degrees of freedom actually buys

A degree of freedom is an independently controllable axis of motion. A useful human hand does far more than open and close: the thumb opposes the fingers, fingertips change orientation, contact points shift and force must be distributed across an object. Wuji Hand 2 is designed to give each of five fingers four directly driven degrees of freedom, including opposition and lateral movement.

That sophistication is unnecessary for many factory tasks. If every part arrives in exactly the same pose, a two-finger gripper or suction cup is usually cheaper, stronger and easier to control. Dexterous hands matter when the object changes shape, tools must be manipulated, an item has to be re-grasped, or the environment was built for humans rather than robots.

The harder product may be the training data

Physical AI does not become dexterous merely because a robot hand is mechanically capable. The model has to learn how the hand should move. Wuji Glove is designed for that data-collection problem, with 526 tactile sensing points at 120 frames per second, electromagnetic motion capture, claimed position accuracy within 2 mm and latency within 10 ms.

A human operator wears the glove and performs a task. Joint motion, position, touch and force information are recorded. Retargeting software maps the human hand onto a robot hand with different geometry. The robot can then reproduce the demonstration, and the data can feed imitation-learning or VLA models.

Japan’s policy is moving toward the same bottleneck

Japan’s Ministry of Economy, Trade and Industry has made AI robotics and physical AI an explicit industrial-policy priority in 2026. In June, METI and NEDO launched a multimodal foundation-model program intended to leverage Japan’s manufacturing and logistics data while building domestic AI capability.

METI’s GENIAC robotics projects make the data problem unusually concrete: five-finger humanoid hands, teleoperated contact data, screw-fastening, component insertion and other long-tail tasks are all identified as targets. Wuji Hand 2 therefore arrives in Japan at a moment when policy is shifting from conventional factory automation toward learning systems trained on physical-work data.

Japan is a robot powerhouse, but the axis of competition is changing

Japan remains one of the world’s central robotics countries. The International Federation of Robotics said Japan represented 38% of global robot production, while about 44,500 industrial robots were installed domestically in 2024 and the operating stock reached roughly 450,500 units.

Japan’s historical strength, however, is highly engineered factory automation: welding, painting, material handling and other tasks where the workspace is designed around the robot. Physical AI targets messier environments in which object location varies, materials deform and robots must share spaces built for people.

That shifts the competitive question from “Can the machine repeat one programmed path a million times?” to “Can it see, decide, touch and correct when reality changes?”

A research component, not a finished worker

HumansX lists Wuji Hand 2 at approximately 180 × 80 × 40 mm and about 800 g. Development support includes Python and C SDKs, ROS 2, and simulation models for MuJoCo and NVIDIA Isaac Sim.

That positioning matters. This is not a complete warehouse worker or factory employee. It is a component for universities, robotics companies and AI researchers building dexterous manipulation systems, teleoperation pipelines or humanoids. Vision, planning, safety, arm control, grasp strategy and system integration still have to be solved around the hand.

“World-class” remains a vendor claim

HumansX describes Wuji as a world-leading five-finger dexterous-hand maker and points to partnerships with Genesis AI and Lightwheel, as well as use of Wuji hardware in academic humanoid research involving Tsinghua University.

Those are meaningful adoption signals, but Japan.co.jp did not find a public, independent benchmark comparing Wuji Hand 2 with competing dexterous hands under identical conditions. Twenty degrees of freedom alone does not establish superiority. Real utility also depends on force, repeatability, impact resistance, thermal behavior, actuator life, communications, maintenance, price and task-success rates.

The glove may create the longer-lived business

HumansX is not limiting its Japanese offering to hardware sales. It also offers tactile and hand-motion data collection, dataset construction, setup support, ROS 2 and SDK training, teleoperation and retargeting guidance.

That expands the business from selling a device to selling the process by which a customer creates proprietary physical-AI data. Models will change, and robot bodies will change. Well-structured manipulation datasets may remain useful across multiple generations of hardware and software.

A Japanese integration layer for Chinese hardware

HumansX was established in June 2025 in Tokyo. Its representative director is listed in Japanese corporate materials as Yi Feng (易峰; Xavi Yi). The company argues that Japanese users struggle both to procure advanced overseas hardware with Japanese-language support and to complete the secondary engineering required after buying data gloves or robot hands.

That intermediary role reflects a broader shift in robotics. Japanese companies do not have to invent every component domestically. They may increasingly combine fast-moving Chinese hardware with Japanese manufacturing requirements, research institutions, safety expectations and proprietary shop-floor data.

Price is still missing

Neither Wuji Hand 2 nor Wuji Glove has a public Japanese price in the launch announcement. Standard lead time, warranty conditions and annual supply capacity are also undisclosed; HumansX asks customers to inquire for pricing, delivery and demonstrations.

The release also says adoption is progressing among major Japanese companies, but it does not identify those companies, quantities or use cases. That claim should therefore be treated as the distributor’s statement rather than evidence of large-scale domestic deployment.

The hand is where AI meets reality

Industrial robots achieved extraordinary productivity by reshaping factories around machines. Physical AI aims at the opposite challenge: put robots into environments designed for people and teach them to cope with variation.

In that world, the hand is not a peripheral. It is the point where visual understanding, planning and force control become physical action.

It is too early to know whether Wuji Hand 2 will become widely adopted in Japan. But the product pairing—a 20-DOF hand beside a 526-point tactile glove—captures the change in the robotics business. The competition is moving from building machines toward turning human skill into data that machines can learn.

Reporting cutoff: Oct. 6, 2026, 1:21 PM JST. Public materials reviewed do not establish Japanese pricing, standard lead time, warranty terms, annual supply capacity, durability life, named Japanese adopters or an independent head-to-head benchmark.