LiberAI introduces Liber-0 for zero-shot robot manipulation

LiberAI introduced Liber-0, a World Action Model pretrained on more than 2.2 million hours of human experience. It combines human video, glove-recorded hand motion, embodiment-conditioned learning, and a Unified Embodiment Representation to transfer manipulation skills across robot bodies. In LiberAI’s zero-shot evaluation, 60 GPT-6-proposed tasks were tested over 600 trials, producing a 36.0% overall success rate without robot demonstrations collected for those target tasks. The model jointly predicts future video frames and actions, while the company’s LEO platform is designed to reduce visual and kinematic differences between human demonstrations and robot execution. LiberAI says future work will add tactile and force sensing and expand glove-based data collection across more people, tasks, and everyday environments.

Why it matters

The approach could reduce the amount of robot-specific demonstration data needed for new manipulation tasks, although the evidence is from LiberAI’s own benchmarks. In one separate eight-task transfer experiment, using 25% of the robot demonstration pool with the full glove pool reached a progress score of 75.6, compared with 72.8 for robot-only training using the full robot pool; the glove and robot pools contained equal numbers of demonstrations.

Key points

  • Liber-0 was pretrained on more than 2.2 million hours of human experience.
  • It achieved a 36.0% overall success rate across 60 tasks and 600 trials in LiberAI’s zero-shot evaluation.
  • Success rates were 63.5% for pick-and-place, 34.0% for short non-pick-and-place tasks, and 10.5% for challenging multi-step tasks.
  • Using 25% of the robot data plus 100% of the glove data produced a 75.6 progress score in a separate eight-task experiment, exceeding the 72.8 score from full robot-only training.
  • LiberSense Glove records hand poses and trajectories but does not currently provide tactile or force supervision; LiberAI identifies those signals as future additions.

Based on reporting by The original announcement and The Robot Report.