LiberAI trains its Liber-0 robot model on 2.2 million hours of human experience
Beijing startup LiberAI says its Liber-0 model, pretrained on human video and glove data, reached 36.0% success on 60 unseen manipulation tasks.
LiberAI, a Beijing startup founded in December 2025, has presented Liber-0, a world action model for robot manipulation. It was pretrained on more than 2.2 million hours of human experience, mostly video, and it predicts both future frames and the actions that produce them.
The company says glove data matters most. Its LiberSense Glove records hand poses and trajectories directly, and is 0.3 mm thick and 131 g. In controlled tests, 10,000 hours of glove recordings beat 2 million hours of web video on both pick-and-place and other manipulation. After full training, Liber-0 scored 83.6 on pick-and-place and 73.8 on other tasks, on a 0 to 100 progress scale.
On 60 new tasks proposed by GPT-6, it succeeded in 36.0% of 600 trials: 63.5% for pick-and-place, 34.0% for other manipulation and 10.5% for hard, long-horizon jobs. Its checkpoint had no robot demonstrations made for those tasks.
Why it matters
Collecting human data with gloves needs no robot on site, so it may scale more cheaply than teleoperation. The low 10.5% on long tasks shows the limits, though. Watch for independent tests, a public release of the model or data, and results on robots other than LiberAI's own LEO platform. The company has not said when Liber-0 will be available.
