Japan’s AI Robot Association (AIRoA) released AIRoA-0 on 6 November 2026, a vision-language-action model trained only on handheld gripper demonstrations. The group reports a 78% average success rate on ten bimanual tasks at a site where none of the training data was recorded, against 27% for the strongest of four other models post-trained on the same task demos. Those figures come from AIRoA’s own evaluation, not an independent lab.
The pitch for a small robotics team is practical: you do not need a teleoperation fleet to gather data. People record with YUBI, an open-source handheld gripper, and the same tool mounts on the robot later. Pre-trained weights and post-training code are on Hugging Face and GitHub. Weights sit under the Gemma Terms of Use, so check that policy before you ship anything commercial.
What was measured
According to AIRoA’s technical report (arXiv:2611.03815), the pre-training corpus is 16,151 hours of bimanual YUBI demos, about 10.65 million subtask segments. There was no robot teleoperation, no reinforcement learning, and no on-robot corrections in that training path. The backbone starts from Physical Intelligence’s public π0.5 checkpoint.
On a bimanual Franka, each of ten tasks ran for 15 trials with partial credit for partly completed runs. A-0’s reported averages include 97% on color sorting, 93% on a fuel-bracket task, and 30% on earbuds. Towel folding and robot-vacuum pad swaps both landed at 73%.
The four baselines were π0.5, Xiaomi-Robotics-1, NVIDIA GR00T N1.7, and MolmoAct2. Each baseline got a search over action horizons and update counts. A-0 used one post-training recipe on every task. That design favors A-0, and the authors say so in the report.
The same gripper-pose weights also ran on TX-G2, OpenArm, and UR arms without retraining for the arm, as long as YUBI and its wrist cameras were the end effector. Post-training a new task is claimed at under four hours on one node of eight H200 GPUs.
What the release includes
AIRoA published pre-trained weights (about 3.36 billion parameters), example post-trained checkpoints, post-training code, YUBI collection code, and an inference example. The full training dataset is still scheduled for a later drop. Hardware designs for YUBI are already public. The work is funded under Japan’s METI and NEDO project JPNP25015.
That open path sits next to other robotics software releases we have covered, including NVIDIA’s VFlash planning stack on Jetson Thor, which targets plan latency rather than a full manipulation foundation model. For how vendor and third-party scores can diverge, see our note on what the big AI leaderboards actually measure. The open-weights pattern for language models, as with Mistral Large 4, is a reminder to wait for outside tests before treating a lab chart as settled.
Limits that matter on the floor
Success rates are AIRoA’s, and partial credit softens hard failures. The earbuds task at 30% shows the system is not evenly strong. You need YUBI hardware and a robot that can mount it. The model sees only two wrist cameras, so workspace layout and lighting will matter, and Gemma use restrictions apply to the weights.
None of that kills the result. A 78% average on long, precise bimanual work from hand demos alone, at an unseen site, is a concrete number teams can try to reproduce. Until someone outside AIRoA runs the same ten tasks, treat 78% versus 27% as the authors’ claim with a disclosed protocol, not a settled ranking.
What to do this week
If you already run dual-arm cells, skim the project page and the report’s Table 1. Build or order a YUBI pair, pick one of the short tasks (towel folding used 3.3 hours of post-training data), and time a post-train on your own GPUs. Log success with the same 15-trial rule the report uses, then compare against your current teleop or VLA baseline before a wider rollout.
Sources
Get the weekly email
The week's most useful AI news, a tutorial and an interview.
Lars Strother
Lars Strother is the senior editor for models and hardware at AI Observer, covering models and research, chips and hardware, and robotics. He came to journalism from engineering and still reads the benchmark tables before the press release. His coverage is measured and skeptical, and he wants to know what was tested, on what, and against which baseline.
Spotted a mistake? Tell us on the corrections page.