Xiaomi-Robotics-U0 - Xiaomi's unified embodied synthesis model - AiBoss

What is Xiaomi-Robotics-U0?

Xiaomi-Robotics-U0 is a unified embodied synthesis model with 38 billion parameters launched by Xiaomi. It is continuously trained based on a world-based model and jointly optimized for five major tasks: text-to-image generation, image editing, embodied scene generation, embodied transfer, and embodied video generation. Xiaomi-Robotics-U0 is the first high-quality multi-view scene generation model that supports multiple robot forms. In human evaluations of embodied scene generation and transfer, it surpasses GPT-Image-2.0 and ranks first on the World Arena embodied video generation leaderboard. It also improves the success rate of π0.5 in real-world manipulation tasks from 36.9% to 63.2%.

Main functions of Xiaomi-Robotics-U0

Text-based image generationSupports high-quality text-to-image synthesis while preserving the visual knowledge of the underlying world model.
Image editingSupports Any-to-Image editing, including fine-grained operations such as camera control, time control, and structure extraction.
Embodied Scene GenerationBased on the robot's form and scene description, generate an initial observation view that is consistent across multiple perspectives.
Embodied migrationWhile maintaining consistency across multiple perspectives and dynamic interaction, it enables structured and controllable migration across scenarios.
Embossed Video GenerationSupports zero-sample, multi-view embodied video generation, synthesizing static scenes into sequential operation videos.

The technical principles of Xiaomi-Robotics-U0

Unified Autoregressive FrameworkBased on EMU3.5 initialization, IBQ Tokenizer is used for 16×16 space compression, and all modalities are unified into a discrete vocabulary for next-token prediction.
Joint Continuous Training: A unified autoregressive objective is used for continuous training on generalized and embodied datasets to avoid loss of generalization ability caused by using limited robot data.
FlashAR+ Inference AccelerationBy introducing a vertical prediction head with anti-diagonal parallel decoding and combining it with vLLM optimization, the generation speed of 1024×1024 resolution images is improved by 82.9 times.
Structured control decompositionIt decouples the scene into five independent control dimensions: workspace, task object, irrelevant object, lighting, and background, and supports scalable embodied video enhancement.
Subtask-subgoal interleaved learningThe robot trajectory is decomposed into subtasks using HDBSCAN clustering, generating interleaved image-text sequences to capture long-term task progress and fine-grained interaction dynamics.

How to use Xiaomi-Robotics-U0

Environment configurationVisit the Xiaomi Robotics official website to download the model weights and inference code, and configure a high-performance GPU environment that supports CUDA and vLLM.
Model loadingBased on the EMU3.5 architecture, the 38 billion parameter checkpoints of Xiaomi-Robotics-U0 are initialized and loaded, and the IBQ Tokenizer is called in a unified manner to process multimodal input sequences.
Text-based image generationInput text description and generate a high-quality image with a resolution of 1024×1024 directly through standard autoregressive next-token prediction.
Image editingProvide one to three reference images along with text instructions, and the model will perform fine-grained editing such as camera control, time control, or structure extraction based on Any-to-Image capability.
Embodied Scene GenerationInput robot morphology description and structured scene text, and the model outputs an initial observation image that strictly satisfies multi-view consistency and geometric coherence.
Embodied migrationGiven the current multi-view observations and the target scene description, the model generates the transferred multi-view RGB image while maintaining the interactive dynamics.
Embossed Video GenerationProvides initial observation frames and task instructions. The model generates 15 to 25 frames of temporally coherent robot operation video through autoregressive expansion.
Inference accelerationEnabling the FlashAR+ vertical prediction head and vLLM optimization results in an 82.9x speedup for image generation at 1024×1024 resolution.
Data Engine ApplicationsThe generated embodied scenes or video sequences are directly input into the downstream robot policy network to improve the out-of-distribution generalization ability of real-world manipulation tasks.

The core advantages of Xiaomi-Robotics-U0

UnityThe first model to unify basic image/video generation and embodied generation within a single framework, avoiding the loss of visual knowledge caused by post-training.
Multi-perspective consistencySupports high-quality multi-view scene generation across various robot forms, strictly meeting geometric coherence and camera calibration constraints.
Controlled migrationThe introduction of a structured control mechanism allows for fine-grained editing of dimensions such as workspace, background, and lighting while maintaining interactive dynamics.
Data engine capabilitiesThe generated embodied scenes can be directly expanded into temporally coherent operation videos, providing scalable synthetic trajectories for downstream policy learning.
SOTA performanceIt surpasses GPT-Image-2.0 in embodied scene generation and transfer, and ranks first in embodied video generation on World Arena.

Xiaomi-Robotics-U0 project address

Project official website: https://robotics.xiaomi.com/xiaomi-robotics-u0.html
HuggingFace model library:https://huggingface.co/collections/XiaomiRobotics/xiaomi-robotics-u0
arXiv technical paper: https://arxiv.org/pdf/2607.11643

Xiaomi Robotics U0 Competitive Product Comparison

Dimension	Xiaomi-Robotics-U0	π0.5
position	Unified embodied synthesis world model, focusing on generating controllable data and scenarios.	End-to-end vision-language-action strategy model, focusing on direct robot control.
Parameter size	38 billion parameters autoregressive Transformer	Adopting a VLM backbone + Action Expert layered architecture
Core Competencies	Multi-view scene generation, embodied transfer, video generation, image editing	End-to-end robot control, generalized home task execution
Architecture	Based on EMU3.5 initialization, unified next-token prediction.	Pre-train discrete action tokens, then train stream matching to generate continuous actions.
Data Engine	It can directly generate embodied scenes and videos, increasing the success rate of out-of-distribution mapping from 36.9% to 63.2%.	It does not have the ability to generate data itself and relies on real-world data collection.
Multi-perspective consistency	Native support for multi-view scene generation and geometrically coherent transfer across various robot forms	As a single-view strategy model, it does not directly handle multi-view synthesis.
Visual knowledge retention	Jointly trained general generation and embodied tasks, maintaining high scores on the VLM benchmark (ERQA 40.8, SEED 78.6).	Low VLM benchmark scores (ERQA 0.0, SEED 21.5) indicate impaired visual understanding.

Application scenarios of Xiaomi-Robotics-U0

Robot Data SynthesisIt generates large-scale, diverse synthetic operation trajectories for training real robots, solving the problems of high cost and limited scenarios for collecting real data.
Cross-form scene migrationTo transfer the same operation task from one robot form (such as a single arm) to another form (such as a dual arm) while maintaining scene consistency.
Simulation environment constructionIt can quickly generate high-quality multi-view robot operation scenarios for simulation platform training and strategy verification.
Enhanced operational strategies: Improve the generalization ability of robot strategies by utilizing generated out-of-distribution data, such as increasing the success rate of π0.5 by 26.3%.
Embodied Intelligence ResearchAs a fundamental model of the world, it supports basic research such as long-term task planning, subtask decomposition, and interactive dynamic prediction.