Japan's AIRoA open-sources AIRoA-0 robot foundation model, trained without robot data collection, hitting 78% success
On November 6, Japan's AI Robot Association (AIRoA) released AIRoA-0, a robot foundation model built on Physical Intelligence's π0.5 and trained on 16,151 hours of manipulation data recorded by people using handheld YUBI grippers. At a site where no training data was collected it averaged 78% success across 10 tasks, versus 27% for the best of four other models trained on the same data; weights and code are public and it has run on four dual-arm platforms including Franka and UR. My take: the expensive part of robotics isn't the arm, it's the teaching data, so if human-held grippers can scale data collection, the bar for factories and warehouses drops a lot. Smaller firms eyeing automation should watch the data pipeline before rushing to buy robots.