Xiaomi-Robotics-U0: Unified Embodied Synthesis with World Foundation Models

Xiaomi Robotics

Xiaomi-Robotics-U0 exposes five public task types through one autoregressive framework:

News

• [July 2026] 🎉 Released the Technical Report.
• [July 2026] 🔥 Released Xiaomi-Robotics-U0 and Xiaomi-Robotics-U0-FlashAR weights.
• [July 2026] 💻 Inference code and scripts are now live!

Table of Contents

1. Model & Weights
2. Quick Start
3. Gradio Demo
4. Citation

1. Model & Weights

The currently released model weights support Scene Gen, Transfer, T2I, and X2I.

The Xiaomi-Robotics-U0-Video checkpoint is coming soon.

2. Quick Start

Environment Setup

Create one conda environment for the backend you plan to run:

git clone https://github.com/XiaomiRobotics/Xiaomi-Robotics-U0.git
cd Xiaomi-Robotics-U0

For eager inference:

conda create -n xr-u0-eager python=3.10 -y
conda activate xr-u0-eager
pip install -U pip
pip install -r requirements-ar.txt
pip install -e .

For AR vLLM inference:

conda create -n xr-u0-ar-vllm python=3.12 -y
conda activate xr-u0-ar-vllm
pip install -U pip
pip install -r requirements-vllm.txt
pip install -e .
python -m xr_u0_ar.apply_vllm_patches

For Xiaomi-Robotics-U0-FlashAR vLLM inference:

conda create -n xr-u0-flashar-vllm python=3.12 -y
conda activate xr-u0-flashar-vllm
pip install -U pip
pip install -r requirements-vllm.txt
pip install -e .
python -m xr_u0_flashar.apply_vllm_patches

Keep the AR and FlashAR vLLM patch sets in separate conda environments. The patch scripts check that the installed vLLM version is exactly 0.11.0.

RGB Transfer Depth Setup (Optional)

RGB Transfer uses Depth Anything 3 (DA3) to convert RGB reference images into inverse-depth maps before Xiaomi-Robotics-U0 inference. Install the optional DA3 dependencies in the same conda environment that will run Xiaomi-Robotics-U0:

pip install -e ".[depth]"
python -m pip install --no-deps "depth-anything-3 @ git+https://github.com/ByteDance-Seed/Depth-Anything-3.git"

The default DA3 model is depth-anything/DA3-LARGE-1.1. You can pass the Hub ID directly, or override it with a local DA3-LARGE-1.1 directory:

python scripts/inference.py \
  --engine flashar --backend vllm --task transfer \
  --input-image-type rgb \
  --da3-model-path depth-anything/DA3-LARGE-1.1

You can also download it first and pass the local directory:

huggingface-cli download depth-anything/DA3-LARGE-1.1 \
  --local-dir <local-da3-model-dir>

Configuration

Xiaomi-Robotics-U0 uses composable Python config files. Values such as model_path, tokenizer_path, and vq_path can be local directories or HuggingFace Hub IDs for automatic download.

CLI arguments override the config files, which is useful for quick tests:

python scripts/inference.py \
  --engine ar --backend eager --task t2i \
  --model-path <Xiaomi-Robotics-U0-HF-ID-or-local-path> \
  --tokenizer-path <Xiaomi-Robotics-U0-HF-ID-or-local-path> \
  --vq-path <VisionTokenizer-HF-ID-or-local-path> \
  --dry-run

Use --dry-run whenever you want to inspect the final composed config without loading a model.

Inference

All tasks use the same entry point:

python scripts/inference.py \
  --engine <ar|flashar> \
  --backend <eager|vllm> \
  --task <t2i|x2i|scene-gen|transfer>

Use --engine ar with Xiaomi-Robotics-U0 for T2I, X2I, Scene Gen, and Transfer. Use --engine flashar with Xiaomi-Robotics-U0-FlashAR to speed up those tasks. The Video Gen code remains available through --engine ar --task video-gen, while the required Xiaomi-Robotics-U0-Video checkpoint is coming soon.

Task examples come from configs/tasks/*.py. Override them from the CLI with --prompt and --reference-image when needed. Transfer uses depth-map references by default; RGB Transfer needs --input-image-type rgb; see RGB Transfer Depth Setup.

Xiaomi-Robotics-U0-FlashAR vLLM keeps enable_prefix_caching = False by default.

Distributed Inference

Set visible GPUs and select the multi-GPU profile:

CUDA_VISIBLE_DEVICES=0,1 \
python scripts/inference.py --engine flashar --backend vllm --task t2i --profile multi-gpu

3. Gradio Demo

The demo runs Xiaomi-Robotics-U0-FlashAR through a FastAPI model server and a Gradio UI. It covers T2I, X2I, Scene Gen, and Transfer.

Use the xr-u0-flashar-vllm environment and install the UI dependencies:

conda activate xr-u0-flashar-vllm
pip install -e ".[demo]"

Set paths with environment variables or the matching server CLI arguments:

export XR_U0_FLASHAR_MODEL_DIR=<flashar-model-or-local-path>
export XR_U0_FLASHAR_TOKENIZER_DIR=<flashar-tokenizer-or-local-path>
export XR_U0_VISION_TOKENIZER_DIR=<VisionTokenizer-HF-ID-or-local-path>
# Optional; only needed when using RGB images for Transfer.
export XR_U0_DA3_MODEL_PATH=depth-anything/DA3-LARGE-1.1

XR_U0_DA3_MODEL_PATH is optional. Depth-map Transfer examples do not use it. For RGB Transfer, first install the optional DA3 dependencies as described in RGB Transfer Depth Setup.

Start the API server and UI in two terminals:

CUDA_VISIBLE_DEVICES=0 python demo/flashar_api_server.py --load-on-startup
# CUDA_VISIBLE_DEVICES=0,1 \
# python demo/flashar_api_server.py --load-on-startup --tensor-parallel-size 2
# export XR_U0_FLASHAR_TP=2

python demo/flashar_gradio_app.py --api-url http://127.0.0.1:8000

Open http://127.0.0.1:7860. Outputs are saved under outputs/gradio_flashar/ with neighboring audit JSON files.

4. Citation

If you find this work useful, please cite:

@misc{li2026xiaomiroboticsu0,
  title         = {{Xiaomi-Robotics-U0}: Unified Embodied Synthesis with World Foundation Model},
  author        = {Xinghang Li and Jun Guo and Qiwei Li and Long Qian and Hang Lai and Yueze Wang and Hongyu Yan and Jiahang Cao and Xi Chen and Jingen Qu and Jiaxi Song and Nan Sun and Hanye Zhao and Futeng Liu and Wanli Peng and Heyun Wang and Yunhong Wang and Caoyu Xia and Jack Zhao and Diyun Xiang and Hangjun Ye and Heng Qu and Huaping Liu and Jason Li},
  year          = {2026},
  eprint        = {2607.11643},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2607.11643}
}