LiberAI has introduced Liber-0, a world Action model designed to transfer human manipulation experience to robots. The Beijing-based startup says the model was pretrained on more than 2.2 million hours of human experience, combining video, glove-recorded hand movements and learning that accounts for differences between human and robot bodies.
LiberAI tested Liber-0 on 60 tasks proposed by GPT-6, spanning pick-and-place, other manipulation and more challenging multi-step activities. Across 600 trials, it recorded an overall success rate of 36%. Success rates were 63.5% for pick-and-place tasks, 34.0% for other manipulation tasks and 10.5% for the challenging tasks.
The evaluation was described as zero-shot because the checkpoint had not been trained on robot demonstrations collected specifically for those target tasks. However, its pretraining did include robot data. Examples in the test suite included putting a pen in a pencil case, closing a drawer and placing a pen in a pouch before storing the pouch in a drawer.
LiberAI says the model learns from ordinary video by predicting future frames alongside actions, while also using glove and robot recordings that contain measured movement. Its LiberSense Glove system captures hand motion, and an embodiment representation gives the model information about body shape and joint movement to help adapt actions to different platforms.
In a separate experiment, robot demonstrations covered pick-and-place actions while glove recordings added actions such as folding, pressing and opening. The non-pick-and-place progress score rose from 2.2 with robot data alone to 47.5 after adding glove data, and to 54.6 when the glove set also included pick-and-place. A policy trained on robot demonstrations for every action scored 56.3. These results came from a separate evaluation, not the 60-task test.
Another experiment examined whether glove recordings could reduce the amount of robot data required. In an eight-task evaluation, a policy using 25% of the robot demonstration pool and all of the glove pool scored 75.6 on the progress measure, compared with 72.8 for a policy using the full robot pool. In that setup, the mixed policy used 75% fewer robot demonstrations, while still using the complete glove dataset.
LiberAI also reported that its pretrained checkpoint achieved progress scores of 81.4 for pick-and-place and 72.3 for other manipulation tasks on internal benchmarks. Robot post-training raised those scores to 83.6 and 73.8, respectively.
The company said the glove system currently records hand poses and trajectories, but not contact or force. It identified tactile and force sensing, as well as data collection across a wider range of homes, as future research directions.
Sources: The Robot Report.