What does Xiaomi-Robotics-U0 actually generate?

Xiaomi-Robotics-U0 handles five kinds of output in a single model: text-to-image, image-to-image, multi-view embodied scene generation, structured embodied transfer for fine-grained editing, and embodied video generation. Every output is produced autoregressively from a shared discrete visual tokenizer, so the same weights cover both foundation image generation and robot-centric world modelling.

How does it compare to prior robotics and image models?

Xiaomi reports Xiaomi-Robotics-U0 outperforms GPT-Image-2.0 in human evaluations of embodied scene generation and transfer, ranks first on World Arena for embodied video generation, and lifts the out-of-distribution success rate of the pi_0.5 policy from 36.9% to 63.2% on real-world manipulation. Those numbers come from Xiaomi's own technical report.

Can I run Xiaomi-Robotics-U0 locally?

Yes — the code, model weights and technical report are all open under Apache-2.0 on GitHub and Hugging Face, with mirrors on ModelScope. Xiaomi ships a FlashAR variant that hits 5.44 seconds per 1024×1024 image on a single H20 GPU with vLLM, which Xiaomi claims is 82.86× faster than a naive autoregressive baseline.

How is Xiaomi-Robotics-U0 different from Xiaomi-Robotics-0 (Feb 2026)?

Xiaomi-Robotics-0 was a 4.7B vision-language-action model focused on real-time robot execution. Xiaomi-Robotics-U0 is a much larger 38B autoregressive foundation model, and its focus is unified embodied synthesis — generating scenes and video for world modelling — rather than driving a robot at execution time. The two models sit on different rungs of Xiaomi's robotics stack.

Xiaomi Robotics · 2026-07-14 · major

Xiaomi-Robotics-U0 — open 38B unified embodied world model

Xiaomi-Robotics-U0 is an open 38B autoregressive world foundation model that unifies text-to-image, image-to-image, embodied scene generation, transfer and video in one architecture. Weights and inference code are released under Apache-2.0.

Xiaomi-Robotics-U0 GitHub repository open graph image

38B open autoregressive foundation model that generates images, embodied scenes and robot video from one shared tokenizer.

Quick facts

Maker	Xiaomi Robotics
Parameters	38B (also a 34B variant)
License	Apache-2.0 (code + weights)
Tasks unified	T2I, X2I, scene generation, embodied transfer, video
Availability	GitHub, Hugging Face, ModelScope
FlashAR speed	5.44s per 1024×1024 image on a single H20 GPU
Paper	arXiv 2607.11643

What is it?

Xiaomi-Robotics-U0 is Xiaomi Robotics' first open unified embodied generative model — a 38B autoregressive network that shares a discrete visual tokenizer across text-to-image, image-to-image, embodied scene generation, structured transfer and video. Code, weights and a technical report shipped together under Apache-2.0.

How does it work?

The architecture treats every output modality as a stream of discrete visual tokens the model predicts autoregressively, which lets it fold foundation image generation and robot-centric world modelling into one set of weights. Xiaomi releases a plain U0, a 34B variant, and a FlashAR distillation that reaches 5.44 seconds per 1024×1024 image on a single H20 GPU — 82.86× faster than an eager autoregressive baseline.

Why does it matter?

Unified embodied generative models are the missing piece between image foundation models and robot policies: the same weights that render a photorealistic scene can also imagine what a robot's next viewpoint looks like. Xiaomi-Robotics-U0 reports beating GPT-Image-2.0 on embodied scene generation and lifting the out-of-distribution success rate of the pi_0.5 policy from 36.9% to 63.2% on real manipulation tasks — a concrete lift for downstream robot learning.

Who is it for?

Robotics researchers, embodied-AI teams, groups building world models for planning.

Frequently asked questions

What does Xiaomi-Robotics-U0 actually generate?: Xiaomi-Robotics-U0 handles five kinds of output in a single model: text-to-image, image-to-image, multi-view embodied scene generation, structured embodied transfer for fine-grained editing, and embodied video generation. Every output is produced autoregressively from a shared discrete visual tokenizer, so the same weights cover both foundation image generation and robot-centric world modelling.
How does it compare to prior robotics and image models?: Xiaomi reports Xiaomi-Robotics-U0 outperforms GPT-Image-2.0 in human evaluations of embodied scene generation and transfer, ranks first on World Arena for embodied video generation, and lifts the out-of-distribution success rate of the pi_0.5 policy from 36.9% to 63.2% on real-world manipulation. Those numbers come from Xiaomi's own technical report.
Can I run Xiaomi-Robotics-U0 locally?: Yes — the code, model weights and technical report are all open under Apache-2.0 on GitHub and Hugging Face, with mirrors on ModelScope. Xiaomi ships a FlashAR variant that hits 5.44 seconds per 1024×1024 image on a single H20 GPU with vLLM, which Xiaomi claims is 82.86× faster than a naive autoregressive baseline.
How is Xiaomi-Robotics-U0 different from Xiaomi-Robotics-0 (Feb 2026)?: Xiaomi-Robotics-0 was a 4.7B vision-language-action model focused on real-time robot execution. Xiaomi-Robotics-U0 is a much larger 38B autoregressive foundation model, and its focus is unified embodied synthesis — generating scenes and video for world modelling — rather than driving a robot at execution time. The two models sit on different rungs of Xiaomi's robotics stack.

Try it

huggingface.co/collections/XiaomiRobotics/xiaomi-robotics-u0