Source note
Xiaomi Opens a 38B World Model Built to Generate Robot Data
Summary
Xiaomi-Robotics-U0 is a 38-billion-parameter open robot world model used to generate robot training data rather than control robots directly. Xiaomi reports that data generated with U0 increased π0.5's success on unfamiliar real-world manipulation from 36.9% to 63.2%, although the result is vendor-reported and measured on Xiaomi's evaluations.
Problem
- Robot policies need large, diverse training data, but collecting real-world manipulation demonstrations is expensive and limits performance on unfamiliar scenes and objects.
- The paper addresses whether a world model can serve as a data-generation engine for robot policies instead of acting as the controller itself.
Approach
- Xiaomi-Robotics-U0 is a single 38B-parameter model that generates robot scenes and videos using a method described as similar to image and video generation.
- The generated scenes and videos are used as additional training data for π0.5, a robot policy.
- Xiaomi evaluates U0 on embodied video quality through the World Arena leaderboard and compares human preferences for its generated scenes against GPT-Image-2.0.
- The model weights and code are open, allowing external users to reproduce the data-generation pipeline and test its effect on other policies or setups.
Results
- On unfamiliar real-world manipulation, π0.5's reported success increased from 36.9% to 63.2% after training with U0-generated data, a 26.3-percentage-point gain.
- Xiaomi reports that U0 topped the World Arena embodied-video leaderboard.
- Human raters reportedly preferred U0-generated scenes over scenes from GPT-Image-2.0.
- The reported comparisons are Xiaomi's own evaluations; the excerpt provides no independent replication or separate dataset details.