LiberAI’s Liber-0 report presents a robot-learning approach built around more than 2.2 million hours of human experience. Its most revealing result is the uneven success across tasks, rather than the training-volume headline.
The company reports 36 percent overall success across 600 trials on 60 tasks for which the evaluated checkpoint had no task-specific robot demonstrations. Success was 63.5 percent for pick-and-place, 34 percent for short manipulation tasks and 10.5 percent for challenging multistep tasks. These are company-run tests; NAOAT has not independently reproduced them.
Liber-0 combines video learning with measured motion from LiberSense gloves and robot recordings. The company also designed its LEO robot to reduce differences between human and robot movement and camera viewpoints. The model predicts future observations and actions together.
The report distinguishes complete-task success in this evaluation from partial-progress scores used elsewhere in its benchmarks. Those measures answer different questions and should not be substituted for each other when describing performance.
NAOAT analysis
The distribution matters for automation planning. A sequence can fail even when individual motions look convincing, because the final result depends on all the intermediate interactions. The reported category gap makes that problem visible; a single overall score would conceal how much harder sustained manipulation remains.
The hardware alignment also complicates the idea that a large video collection automatically transfers to any robot. Here, data capture, body geometry and observation viewpoints are deliberately connected. For manufacturers evaluating this direction, the open commercial question is how much of that arrangement must be reproduced in a different workcell. The report supports continued experimentation, while leaving dependable production throughput and intervention costs unresolved.