Xiaomi-Robotics-U0 - Xiaomi's Unified Embodied Synthesis Model

What is Xiaomi-Robotics-U0

Xiaomi-Robotics-U0 is a 38-billion-parameter unified embodied synthesis model launched by Xiaomi. Based on continuous training of a world foundation model, it jointly optimizes five major tasks: text-to-image generation, image editing, embodied scene generation, embodied transfer, and embodied video generation. Xiaomi-Robotics-U0 is the first model to support high-quality multi-view scene generation across multiple robot embodiments. It surpasses GPT-Image-2.0 in human evaluations of embodied scene generation and transfer, ranks first on the World Arena embodied video generation leaderboard, and improves π0.5's out-of-distribution success rate on real-world manipulation tasks from 36.9% to 63.2%.

Main Features of Xiaomi-Robotics-U0

Text-to-Image Generation: Supports high-quality text-to-image synthesis, preserving the visual knowledge of the foundation world model.
Image Editing: Supports Any-to-Image editing, including fine-grained operations such as camera control, temporal control, and structure extraction.
Embodied Scene Generation: Generates multi-view consistent initial observation frames based on robot embodiment and scene descriptions.
Embodied Transfer: Achieves structured and controllable transfer across scenes while maintaining multi-view consistency and interaction dynamics.
Embodied Video Generation: Supports zero-shot multi-view embodied video generation, synthesizing static scenes into temporally coherent manipulation videos.

Technical Principles of Xiaomi-Robotics-U0

Unified Autoregressive Framework: Initialized based on EMU3.5, it uses an IBQ Tokenizer for 16×16 spatial compression, unifying all modalities into a discrete vocabulary for next-token prediction.
Joint Continual Training: Conducts continual training with a unified autoregressive objective on general-domain and embodied datasets, avoiding the loss of generalization caused by using limited robot data.
FlashAR+ Inference Acceleration: Introduces a vertical prediction head with anti-diagonal parallel decoding, combined with vLLM optimization, accelerating 1024×1024 resolution image generation by 82.9×.
Structured Control Decomposition: Decouples scenes into five independent control dimensions: workspace, task objects, irrelevant objects, lighting, and background, supporting scalable embodied video enhancement.
Subtask-Subgoal Interleaved Learning: Uses HDBSCAN clustering to decompose robot trajectories into subtasks and generate interleaved image-text sequences, capturing long-horizon task progress and fine-grained interaction dynamics.

Follow WeChat and reply “Open Source” to join theAI Open Source Project Exchange Group

How to Use Xiaomi-Robotics-U0

Environment Setup: Visit the Xiaomi Robotics official website to download model weights and inference code, and configure a high-performance GPU environment that supports CUDA and vLLM.
Model Loading: Initialize based on the EMU3.5 architecture and load the 38-billion-parameter checkpoint of Xiaomi-Robotics-U0, uniformly calling the IBQ Tokenizer to process multimodal input sequences.
Text-to-Image Generation: Input a text description and directly generate high-quality 1024×1024 resolution images through standard autoregressive next-token prediction.
Image Editing: Provide one to three reference images along with text instructions, and the model performs fine-grained editing such as camera control, time control, or structure extraction based on Any-to-Image capabilities.
Embodied Scene Generation: Input a robot morphology description and structured scene text, and the model outputs an initial observation frame that strictly satisfies multi-view consistency and geometric coherence.
Embodied Transfer: Given the current multi-view observations and a target scene description, the model generates transferred multi-view RGB images while preserving interaction dynamics.
Embodied Video Generation: Provide an initial observation frame and task instructions, and the model generates 15 to 25 temporally coherent frames of robot manipulation video through autoregressive rollout.
Inference Acceleration: Enable the FlashAR+ vertical prediction head and vLLM optimization to achieve an 82.9x speedup in image generation at 1024x1024 resolution.
Data Engine Application: The generated embodied scenes or video sequences are directly fed into downstream robot policy networks to improve out-of-distribution generalization for real-world manipulation tasks.

Core Advantages of Xiaomi-Robotics-U0

Unification: The first model to unify base image/video generation and embodied generation within a single framework, avoiding the forgetting of visual knowledge caused by post-training.
Multi-view Consistency: Supports high-quality multi-view scene generation across multiple robot morphologies, strictly satisfying geometric coherence and camera calibration constraints.
Controllable Transfer: Introduces a structured control mechanism that enables fine-grained editing of dimensions such as workspace, background, and lighting while preserving interaction dynamics.
Data Engine Capability: The generated embodied scenes can be directly expanded into temporally coherent manipulation videos, providing scalable synthetic trajectories for downstream policy learning.
SOTA Performance: Surpasses GPT-Image-2.0 on embodied scene generation and transfer, and ranks first on World Arena for embodied video generation.

Project Address of Xiaomi-Robotics-U0

Project Website：https://robotics.xiaomi.com/xiaomi-robotics-u0.html
HuggingFace Model Repository：https://huggingface.co/collections/XiaomiRobotics/xiaomi-robotics-u0
arXiv Technical Paper：https://arxiv.org/pdf/2607.11643

Comparison of Xiaomi-Robotics-U0 with Similar Competitors

Dimension	Xiaomi-Robotics-U0	π0.5
Positioning	Unified embodied synthetic world model, focused on generating controllable data and scenes	End-to-end vision-language-action policy model, focused on directly controlling robots
Parameter Scale	38-billion-parameter autoregressive Transformer	Adopts a VLM backbone + Action Expert hierarchical architecture
Core Capabilities	Multi-view scene generation, embodied transfer, video generation, image editing	End-to-end robot control, household generalization task execution
Architecture	Initialized based on EMU3.5, unified next-token prediction	Pretrained discrete action tokens, post-trained flow matching to generate continuous actions
Data Engine	Can directly generate embodied scenes and videos, improving the out-of-distribution success rate of π0.5 from 36.9% to 63.2%	Does not itself possess data generation capabilities, relies on real-world collection
Multi-view Consistency	Natively supports multi-view scene generation and geometrically coherent transfer across multiple robot morphologies	As a single-view policy model, it does not directly handle multi-view synthesis
Visual knowledge retention	Jointly train general generation and embodied tasks, VLM benchmarks remain high (ERQA 40.8, SEED 78.6)	VLM benchmark scores are low (ERQA 0.0, SEED 21.5), visual understanding is impaired

Application scenarios of Xiaomi-Robotics-U0

Robot data synthesis: Generate large-scale, diverse synthetic manipulation trajectories for real robot training, addressing the issues of high cost and limited scenarios in real data collection.
Cross-embodiment scenario transfer: Transfer the same manipulation task from one robot embodiment (e.g., single-arm) to another (e.g., dual-arm), maintaining scenario consistency.
Simulation environment construction: Rapidly generate high-quality multi-view robot manipulation scenarios for simulation platform training and policy validation.
Manipulation policy enhancement: Use generated out-of-distribution data to improve the generalization ability of robot policies, e.g., increasing the success rate of π0.5 by 26.3%.
Embodied intelligence research: As a world foundation model, support fundamental research such as long-horizon task planning, subtask decomposition, and interactive dynamics prediction.