LiberAI releases world action model Liber-0

LiberAI, a physical AI company founded by Liu Songming, has released its next-generation world action model Liber-0, along with a technical report and an online simulator. The model uses three expert streams for language, vision, and action plus Multimodal Attention to jointly predict future frames and actions. According to LiberAI, the model used in the simulator was trained on 100 hours of simulated teleoperation data and mainly examines object-level zero-shot transfer rather than real-robot generalization on complex long-horizon tasks; the test tasks did not collect targeted robot demonstration training, and each task was run 10 times on real robots, for a total of 600 trials. LiberAI claims that with 25% real-robot demonstrations combined with full glove data, the model outperforms a baseline using 100% real-robot demonstrations.

AI 根据报道生成 · 5 days ago更新