LiberAI unveils Liber-0, a robot model trained on 2.2 million hours of human experience
Beijing-based startup LiberAI has introduced Liber-0, a world action model trained beforehand on over 2.2 million hours of human experience, aimed at robot manipulation without task-specific robot demonstrations, according to The Robot Report. The company announced the model on November 4.
LiberAI said Liber-0 combines human video with glove-recorded hand motion and embodiment-conditioned learning, meaning the model is told about the body that performs an action, to transfer human manipulation skills to robots.
Zero-shot test results
LiberAI evaluated the model on a suite of 60 tasks proposed by GPT-6: 20 pick-and-place tasks, 20 other manipulation tasks, and 20 challenging, long-horizon tasks. Examples included placing a pen in a pencil case and pushing a drawer closed.
https://twitter.com/i/status/2117803284039905456
The company reported success rates of 63.5% on pick-and-place, 34.0% on other manipulation and 10.5% on the hardest tasks, for an overall 36.0% across 600 trials. Here, zero-shot means the tested checkpoint was not trained on robot demonstrations collected specifically for those tasks, although pretraining still included robot data.
How the model links human video to robot action
LiberAI said human video cannot be turned directly into robot commands. In its account, camera positions vary, hand poses estimated from footage are imprecise, and human bodies move in ways that differ from robots.
To address this, the company uses a data-collection glove called LiberSense Glove that records hand movements directly. It also uses a robot platform whose camera views and hand, wrist and waist movements more closely resemble those of humans, and a Unified Embodiment Representation that supplies information on body shape and joint movement. Liber-0 jointly predicts future video frames and actions, which lets it learn from ordinary videos without action labels.
“Human video shows what people do, but not exactly how their hands move or how that movement fits a different body,”
Reduced need for robot demonstrations
In a separate eight-task experiment, LiberAI said a policy trained on 25% of the robot demonstration pool plus all of the glove pool reached a progress score of 75.6, versus 72.8 for a policy trained on the full robot pool. That amounts to comparable performance with 75% fewer robot demonstrations, though the mixed policy still used the full glove data.
In another controlled test, a policy trained on robot demonstrations covering only pick-and-place scored 2.2 on non-pick-and-place tasks. Adding glove recordings of those actions raised the score to 47.5.
Limitations
LiberAI noted that the glove captures hand poses and trajectories but does not measure touch or force. Those signals matter for tasks such as insertion, fastening and handling soft materials. The company listed tactile and force sensing, and data collection in a wider range of homes, as future research directions.
