Xiaomi Releases Xiaomi-Robotics-U0 38 Billion Parameter Model

Xiaomi released Xiaomi-Robotics-U0.

The model holds 38 billion parameters. It handles multimodal autoregressive generation for robot environments.

Xiaomi treats embodied tasks as extensions of image and video creation. The system optimizes text to image conversion, image editing, scene generation, transfer across robot forms, and video output in one training run.

The approach supports multiple robot body types at once. It adds structured controls for transfer between different robot shapes.

Xiaomi reports state of the art results on single step and sequence generation benchmarks. Human raters ranked the model above GPT-Image-2.0 for embodied scene quality and transfer accuracy.

The model reached first place in the World Arena benchmark for embodied video generation. See the technical report for full benchmark details.

Real world tests showed clear gains. When paired with pi_0.5 on physical robotic arms in warehouse-like settings, the success rate on out of distribution manipulation tasks such as grasping unfamiliar objects amid clutter and variable lighting rose from 36.9 percent to 63.2 percent. Xiaomi open-sources the checkpoints for independent replication on additional hardware.

Code and model checkpoints are now public.

Xiaomi positions the release as an extension of existing image models rather than a separate robot only architecture. The arXiv paper details this joint optimization choice, which places pressure on teams that maintain separate pipelines for vision, video, and robot control.

The open release allows direct comparison with closed systems. Independent labs can now test whether joint optimization delivers consistent gains across robot hardware, as noted in coverage from The Verge.

Future checkpoints will show whether the same base model scales to longer video sequences or more robot platforms. Researchers will track transfer success rates on new robot bodies over the next three months.