LiberAI releases Liber-0 world action model for robots

A world action model has been developed that learns robot manipulation skills from more than 2.2 million hours of human video and recorded hand motion, rather than relying solely on robot-specific demonstrations. The model achieved 36% overall success on 60 diverse manipulation tasks without being trained on robot demonstrations for those specific tasks, though performance varied by task type, with simpler pick-and-place tasks succeeding 63.5% of the time and complex multi-step tasks succeeding 10.5% of the time. This approach matters because collecting robot-specific training data is slow and expensive, while human activity video is abundant and already widely available. The model overcomes the mismatch between human and robot bodies through recorded hand motion data, robot camera positioning designed to match human perspective, and a representation system that accounts for different body shapes and movements.