Japan’s AI Robot Association (AIRoA) has released AIRoA-0, a robot control model trained without collecting any data on robots. Instead, people recorded 16,151 hours of handling objects with YUBI, a handheld gripper that captures video and movement, and the model learned from those recordings.
AIRoA-0 is a vision-language-action model: it takes in camera images and an instruction and outputs robot movements. It is built on π0.5, a model openly released by US start-up Physical Intelligence. In AIRoA’s own tests on ten two-armed tasks, each tried 15 times across four robot arms (Franka, TX-G2, OpenArm and UR), it succeeded 78% of the time on average. The best of four other models trained on the same data managed 27%.
The group has published the model weights on Hugging Face, and the training, inference and YUBI data-collection code on GitHub, for research and development use. The 16,151-hour dataset itself has not been released yet.
Why it matters: robot training data is usually gathered by people remotely steering real robots, which is slow and costly. If cheap handheld recordings can do much of that job, far more labs could train capable robots. The results are AIRoA’s own and have not been independently checked.
Source: AIRoA