What it is
A 38B multimodal autoregressive world model that unifies image, multi-view embodied scene, transfer, and video generation for robotics. Instead of post-training a foundation model only on robot data, it co-trains general image/video generation with embodied tasks in one autoregressive model, preserving pretrained visual knowledge while adding multi-view, embodiment-aware scene, transfer, and video synthesis.
Key results
- pi_0.5 OOD success rises to 63.2% from 36.9% with generated data
- Ranks 1st on World Arena for embodied video generation
- Outperforms GPT-Image-2.0 in human evals on embodied scene generation and transfer
Shows foundation world models can be adapted into embodied world models without giving up broad visual generation ability, and can generate useful training data for robot policies.