Xiaomi opens Robotics-U0: a unified embodied 38B model, open source

Xiaomi publishes an open-source embodied generative model with 38 billion parameters that unifies four robotic tasks into a single architecture. The race for embodied backbones takes a step towards openness.

The fact

Xiaomi publishes Robotics-U0, a 38-billion-parameter embodied generative model, under an open-source license. According to Pandaily (26/07/15), it unifies four tasks previously handled by separate stacks into a single architecture: scene generation, embodied transfer, video generation, and text-to-image.

Our analysis

The technical direction is explicit and consistent with the thread we are following: just as foundation models in NLP, robotics is seeking a single backbone that can be adapted through fine-tuning or prompting, rather than a sum of assembled subsystems. Xiaomi is competing with the approach defended by X-Square Robot (recently covered) and Nvidia's initiatives around GR00T—but with a different strategic lever: open source. 38 billion parameters remain a manageable size for an academic lab, which matters for adoption. Two limitations at this stage of communication: the demonstration focuses on generation and scene, not yet on real-time high-frequency control.

To watch

Three signals: (1) reproducibility by a third-party lab outside China; (2) effective integration into an embedded robotic stack, with published latency figures; (3) an open-source response from Meta, Nvidia, or Hugging Face at the same scale. If the third point occurs within six months, Robotics-U0 will have moved the needle.