Robotics

AIRoA Releases AIRoA-0, an Open Robot Foundation Model Trained on 16,151 Hours of Robot-Free Human Demonstrations

· 6 min read · Robotics

AIRoA Releases AIRoA-0, an Open Robot Foundation Model Trained on 16,151 Hours of Robot-Free Human Demonstrations

The AI Robot Association (AIRoA), a Tokyo-based body chaired by Tetsuya Ogata, on November 6 released AIRoA-0, a robot foundation model that learned two-handed tasks from data humans recorded without using robots at all, the association announced via PR TIMES. People captured the demonstrations with YUBI, a hand-held gripper device that logs video and the trajectory of the operator’s hand, and AIRoA used 16,151 hours of that data for pre-training.

AIRoA said it started from the publicly released π0.5 base weights from US startup Physical Intelligence and did not use teleoperated robot data, reinforcement learning or on-robot corrections. Tested in places different from where the training data was gathered, the model sorted more than 12 small parts into numbered compartments from a single instruction, folded towels, swapped parts on a robot vacuum and opened a pencil case to put a pencil inside. Across 10 evaluation tasks, each tried 15 times, it averaged a 78% success rate, against a best of 27% for four leading models post-trained on the same data.

Because AIRoA-0 outputs the position and orientation of the gripper rather than joint angles, the same model can drive any robot fitted with a YUBI gripper; the association has confirmed it on four machines, a dual-arm Franka, TX-G2, OpenArm and a UR arm. Pre-trained and post-trained weights are on Hugging Face, with data-collection, post-training and inference code on GitHub, and AIRoA says post-training on a user’s own data takes under four hours on a single node with eight GPUs. The work was funded under a METI and NEDO programme on data platforms for generative-AI robotics models.

What it means in practice

AIRoA Releases AIRoA-0, an Open Robot Foundation Model Trained on 16,151 Hours of Robot-Free Human Demonstrations — contextual photo

The release adds an open, Japanese-built option to a field dominated by corporate labs, following Panasonic HD’s T-Rex tactile AI for dexterous two-handed work and earlier efforts such as university labs opening shared robotics datasets. Collecting demonstrations with an inexpensive hand-held device rather than costly robot teleoperation could lower the cost of teaching factory and household tasks, a pressing issue for manufacturers short of workers.

What to watch is whether independent researchers reproduce the 78% figure, how many companies adopt YUBI-equipped grippers, and what AIRoA shows at its booth at the Conference on Robot Learning (CoRL) 2026 in Austin from November 9 to 12.

← Back to AI4Japan