Rohan Paul
@rohanpaul_ai
This looks like a truly brilliant launch for robot training.
Workers wear a thin 131 g sensor glove that records exact hand movements while they do ordinary tasks, so no robot is needed to collect the data.
Chinese robotics startup LiberAI (@liberai_robot) released Liber-0, the AI model that controls a robot’s arms and hands.
With no robot demonstrations of the tested tasks, Liber-0 completed 36.0% of 600 attempts at 60 new tasks. In LiberAI’s tests, 10K glove hours also beat 2M hours of internet video.
Robots normally learn a task after people steer them through it repeatedly, and Liber-0 replaces much of that with 2.2M+ hours of human experience.
Video alone shows what hands do, not the measured motion a robot must copy.
The company’s LiberSense Glove supplies that motion by logging hand poses and trajectories while people work.
Measured motion beat volume: 10K glove hours outperformed 2M hours of web video in matched runs.
Trained on 25% of the robot demonstrations plus glove data, a policy scored 75.6 for partial completion, beating 72.8 from all of them.
