Liber-0 hits 36% zero-shot success on 60 robot tasks using 2.2M hours of human data

CyberRobooo · x · 2026-11-04

Liber AI introduced Liber-0, a World Action Model for zero-shot robot manipulation that requires no task-specific robot demonstrations.

Across 60 GPT-6-proposed tasks and 600 trials, it achieved 36% overall success: 63.5% on pick-and-place, 34.0% on other manipulation, and 10.5% on challenging multi-step tasks.
The model is pretrained on 2.2M+ hours of human experience, combining motion captured with the LiberSense Glove and embodiment alignment to translate human skills into robot actions, jointly predicting future observations and actions.
The technical blog identifies three gaps in learning from human video—embodiment, action-supervision, and visual gaps—and describes how the approach bridges them.

The approach is framed as a promising route toward learning from people rather than costly task-specific demo collection.

More from Embodied