LiberAI reports robot tasks completed without task-specific demonstrations

LiberAI introduced Liber-0 on November 3, reporting a 36.0% success rate across 600 trials of 60 robot manipulation tasks without robot demonstrations collected specifically for those tasks. The model had received robot data during pretraining. Liber-0 combines human video with glove-recorded hand poses and trajectories to connect visual observations with physical actions. Its robot platform aligns cameras and movement capabilities with human data collection. In a separate eight-task evaluation, 25% of the robot demonstration pool plus the full glove pool achieved a progress score of 75.6, versus 72.8 with the full robot pool alone. Progress scores measure partial completion; success on challenging and long-horizon tasks was 10.5%.