Summary
Xiaomi-Robotics-U0 is a 38B-parameter multimodal autoregressive world foundation model for unified embodied synthesis. It supports multi-view scene generation across multiple robot embodiments and introduces structured, controllable embodied transfer — moving robot trajectories to new environments while preserving arm pose and motion. Jointly optimized across text-to-image, image editing, embodied scene generation, and embodied video generation tasks, it achieves an 82× speedup in data generation (from 450s to 5.4s per image) via parallel decoding.
Key Contributions
- 38B multimodal autoregressive model jointly trained across five embodied and image generation objectives
- Embodied scene generation: creates multi-view initial scenes from text descriptions across diverse environments
- Embodied transfer: moves existing trajectories to new environments with configurable lighting, background, surface, and object changes
- 82× generation speedup via parallel decoding optimization
- +26 percentage-point improvement on real-world manipulation OOD success rate (π₀.₅: 36.9% → 63.2%) using synthesized data
Significance
Xiaomi’s open-sourced world foundation model demonstrates that embodied data synthesis can dramatically boost policy generalization — the 26-point OOD improvement with synthetic augmentation is among the strongest reported results for data-engine-driven VLA improvement.