No robot required: Japan’s AIRoA-0 learned two-handed chores from people holding grippers

Most robot brains learn the slow way: a person teleoperates a robot arm, the arm records what happened, and the clip goes into a training set. Japan’s AI Robot Association, known as AIRoA, just released a model built on the opposite premise. On November 6 it published AIRoA-0, or A-0, a robot foundation model whose manipulation training came entirely from people holding grippers in their hands. No robot was needed to collect the data.

That detail is the whole story. If useful robot data can be recorded anywhere a person can carry a handheld tool, the bottleneck shifts from how many robots a lab owns to how many hours of human work it can capture.

How the data gets made

The capture device is called YUBI, short for Yielding Universal Bidigital Interface. A person operates a pair of handheld grippers while YUBI records video and the grippers’ trajectories, with VR-based pose tracking keeping tabs on where each hand is. The hardware design came from Toyota Motor Corporation’s Frontier Research Center and the collection software from AIRoA, and both are open source.

AIRoA started from the π0.5 base weights that Physical Intelligence released, then trained on 16,151 hours of YUBI recordings. The association says it did not use robot teleoperation, reinforcement learning or on-robot corrections for this manipulation training. The project was funded under a program run by Japan’s Ministry of Economy, Trade and Industry and the research agency NEDO.

What the robot actually did

The test site was a place where none of the training data had been recorded, which is the honest way to run this kind of evaluation. AIRoA ran ten task types, 15 trials each, including folding towels, swapping parts on a robot vacuum cleaner, and stacking cups or oranges. A-0 averaged a 78% success rate. Four other vision-language-action models, post-trained on the same data, topped out at 27%.

A few demos stand out. From a single instruction the robot sorted more than a dozen small parts into numbered cells. It unzipped a pencil case, filled it with pencils and zipped it shut again. In a color-sorting test, a person moved the sorting box partway through, and the robot adjusted and dropped the parts into the box’s new spot.

One model, four different arms

The design choice that matters most for builders is what A-0 outputs. Instead of joint angles for one specific robot, it outputs gripper poses, meaning position and orientation. Each robot’s own controller turns those poses into arm motion. As long as the YUBI gripper is mounted, the same model can drive different hardware. AIRoA says it has confirmed the model running on four dual-arm platforms: Franka, TX-G2, OpenArm and UR.

That is a practical answer to a problem every lab hits. Robot bodies keep changing, and retraining for each new arm is expensive. Tying the model to the hand, not the arm, is a cleaner contract.

Open weights, with a dataset still to come

AIRoA published both pre-trained and post-trained weights, post-training code it says runs in under four hours on a single eight-GPU node, inference code, and the YUBI collection code. The training dataset itself is not out yet. AIRoA says it plans to release it and will announce the timing and scope later. The release is meant for research and development use.

Where A-0 fits

The race for robot brains on our company directory mostly splits into two camps. Well-funded startups like Physical Intelligence, Skild AI and FieldAI build general models as products. Hardware makers like AGIBOT train their own models on fleets of their own robots. A-0 sits in a third lane: a government-backed association building on an open model, publishing the results, and betting that human-held data collection is the cheapest way to scale.

It also lands right after Panasonic’s T-Rex fingertip research, another two-handed manipulation result headed to the same conference. Vision tells a robot what to do, and better data and better touch decide whether it finishes. Both groups are pushing on the same weak spot from different sides.

What to watch

The 78% figure comes from AIRoA’s own evaluation across its own task list, and the details are in its technical report rather than an independent benchmark. The real test is whether outside labs can post-train A-0 on their own YUBI data and get similar results, and whether the promised dataset ships. AIRoA will show the work at booth 210 at CoRL 2026 in Austin, November 9 to 12.

Our take

A-0 is not a humanoid and it does not pretend to be one. It is a bet on the plumbing of robot learning: cheap, portable data capture and a model that does not care which arm it is bolted to. If that holds up, it could matter more to the humanoid makers we track than another flashy demo. Teams hiring for robot learning and data collection roles are on our jobs board.