Home / Blog / Xiaomi-Robotics-U0

Published: July 15, 2026 Β· Last reviewed: July 20, 2026

Xiaomi-Robotics-U0: 38B Unified Embodied Foundation Model for Robot Training Data

TL;DR

On July 15, 2026, Xiaomi released Xiaomi-Robotics-U0 β€” a 38B-parameter unified multimodal autoregressive model for embodied generation. It's the first model that handles scene generation, embodied transfer, robot interaction video generation, and text-to-image editing in a single architecture. It scores 73.64 (#1 out of 126) on the WorldArena embodied video benchmark, and training real robot policies on its augmented data improves out-of-distribution task completion by +26.3% (from 36.9% to 63.2% on the packaging/box / folding-towel / packaging-phone / packaging-headphone tasks). Generation is 82.9Γ— faster than vanilla AR via the new FlashAR+ accelerator. The model is open-sourced on the Xiaomi Robotics site, GitHub, HuggingFace, and ModelScope.

Why this matters

Robotic policies need lots of training data β€” but real-world data is expensive, slow to collect, and biased toward whatever lighting/background/objects happen to be in your lab. Pre-2026 approaches bolted together separate models: one for scene generation, one for trajectory transfer, one for video synthesis, one for image editing. They disagreed about geometry, dropped arm poses, and couldn't share weights.

Xiaomi-Robotics-U0 puts all four capabilities into one autoregressive model that understands multi-view geometry and physical consistency natively. You can write text like "this scene but with a marble countertop, three extra chairs, and a sunset" and the output keeps the robot's arm pose, the camera angles, and the original action sequence intact. That's the unlock for using generated data to train real robot policies.

By the numbers (July 15, 2026 release)

38B
parameters
4
unified tasks
73.64
WorldArena (1/126)
+26.3%
real-robot OOD gain
82.9Γ—
FlashAR+ speedup
5.44 s
1024Γ—1024 image

Four tasks, one model

U0 unifies what used to be four separate pipelines. Each is trained jointly in the same autoregressive objective, so the model can switch between them without re-encoding.

β‘  Embodied scene generation

From a text description, generate multi-view initial observations for a specified robot body β€” kitchen, warehouse, underwater, cyberpunk city, anywhere. The first frame is multi-view geometrically consistent (so it can be used as a starting state for downstream policy training).

β‘‘ Embodied transfer

Take an existing robot trajectory (multi-view RGB + depth) and a text prompt like "marble countertop, three extra chairs, sunset." U0 produces a new video where the arm pose, scene layout, and original action sequence are preserved β€” but the visual style, lighting, and object set are swapped. The training-data equivalent of an "image filter" for robot videos.

β‘’ Robot interaction video

Given an initial frame and a manipulation instruction ("pick up the cup," "fold the towel"), U0 generates the rest of the video with action continuity and physical consistency. Zero-shot to arbitrary scenes and robot bodies.

β‘£ Text-to-image + image editing

General-purpose visual generation that shares weights with the embodied tasks. Internet-scale visual knowledge transfers into the embodied model, so it doesn't forget how to draw a cat while learning how to fold a towel.

Five-dim decoupled control

Embodied transfer is hard because most models break the original geometry when you ask them to change one thing. U0's "five-dimensional decoupled control" decomposes a scene into five orthogonal axes you can change independently via natural language:

Change any one of those, the others stay locked. The robot's arm pose, the original scene geometry, and the action sequence all survive the edit.

Benchmarks

WorldArena embodied video benchmark (Tsinghua + PKU): 126 model submissions, U0 (anonymous codename "UNIS") scored 73.64, #1 overall, leading on instruction following, interaction realism, and multi-view consistency. In direct head-to-head on 300 test items against closed-source GPT-Image-2.0, U0 won on depth consistency, structural fidelity, and semantic alignment β€” and crucially, GPT-Image-2.0's outputs had cross-view object displacement that made them unusable for robot training, while U0's preserved the original trajectory.

Real-robot OOD gain: in-house Xiaomi team ran a real-robot policy study on four tasks (box packaging, towel folding, phone packaging, earphone packaging) under out-of-distribution conditions (new cloth patterns, new lighting, reflective / colored-light interference). Using U0-augmented training data with a Ο€β‚€.β‚… policy backbone:

The robots also developed self-correction behavior under strong visual interference (reflections, colored lights) instead of stalling.

FlashAR+: 82.9Γ— faster, 5.44 s for 1024Γ—1024

Autoregressive image generation is slow because every pixel token is predicted in sequence. Xiaomi's FlashAR+ accelerator (extension of their earlier FlashAR work for text-to-image) parallelizes the target-image token group generation while keeping the input prefix serial β€” so the model can ingest a complex multi-image / depth / text condition and only the target region runs in parallel.

Combined with vLLM paged-KV-cache batch scheduling and diagonal-parallel decoding:

This is what makes batch generation at industrial scale (think millions of training samples) practical β€” without FlashAR+, generating a real training corpus would take years on commodity GPU clusters.

Open source

The model, training code, inference code, and demo are fully open-sourced:

Robot hardware compatibility out of the box: ζ–ΉθˆŸζ— ι™ (ARX), ζ™Ίε…ƒ G1/G2 (AgiBot), 松灡 PiPER (Husarion). The model is body-agnostic β€” the same weights can be fine-tuned or prompted for new robot bodies.

How U0 fits with MiMo

Xiaomi-Robotics-U0 is a sibling project to the MiMo LLM family, not a replacement. The split:

Production robot stacks will use both: MiMo for high-level planning and language-conditioned tool use, U0 to generate the training data and possibly the world-model rollouts that the policy trains against.

What it doesn't do (limits)

Get started

For embodied AI researchers and robotics teams:

  1. Clone the GitHub repo and pull the weights from HuggingFace or ModelScope.
  2. For inference: the FlashAR+ path requires 4Γ— A100 / H100 80GB for 1024Γ—1024 generation. INT8 / FP8 paths for 512Γ—512 are in the works.
  3. For data augmentation: combine U0's embodied transfer with a VLM (MiMo-V2.5-Omni works) to generate the variation prompts, then mix augmented + original trajectories in your training pipeline.

This page is an unofficial community write-up of Xiaomi's July 15, 2026 announcement. For the canonical technical report and benchmarks, see the Xiaomi Robotics project page.

Related pages