LiberAI Liber-0 model learns robot manipulation from 2.2M hours of human experience
LiberAI, a Beijing startup founded in December 2025, has unveiled Liber-0, a "world action model" it says was pretrained on more than 2.2 million hours of human experience. Most of that is ordinary video, grounded by recordings from the company's LiberSense gloves and by robot demonstrations. It runs on LiberAI's own robot, LEO, which has dexterous hands, articulated wrists and an adjustable waist.
In LiberAI's zero-shot test, 60 tasks proposed by GPT-6 and 600 trials, Liber-0 succeeded 36% of the time overall: 63.5% on pick-and-place and 10.5% on hard, long-horizon tasks, with no robot demonstrations collected for those tasks. The company also says 10,000 hours of glove data beat 2 million hours of web video in its tests.
Why it matters
These are LiberAI's own internal benchmarks, reported by Humanoids Daily. Most results are partial-progress scores, and the gloves don't yet capture touch or force.
More on robot ai.
About the footage: No footage reviewed: Humanoids Daily shows only LiberAI stills of LEO at tabletop tasks, and LiberAI hasn't said whether its demo clips are sped up.
Sources
New to robots? Read our guide: How to tell if a robot video is real: CGI, sped up, or teleoperated?.
Summarized by Servospace from the sources above, drafted with AI help and checked against those sources before publishing. Claims are the companies' or outlets' unless we say otherwise. How we work.