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

The Core Claim

Foundation models trained on vast internet image and video data are powerful but struggle with embodied AI—generating robot-manipulable scenes that maintain geometric consistency across multiple viewpoints and respect physical constraints. Xiaomi-Robotics-U0 shows that a single 38-billion-parameter model can unify standard vision generation (text-to-image, editing) with embodied tasks (multi-view scene synthesis, robot control transfer), preserving pre-trained knowledge while adapting to robotic constraints. The payoff: the model outperforms specialized competitors on human judgment and achieves a dramatic 26.3 percentage-point boost in real-world robot manipulation success on out-of-distribution tasks (36.9% → 63.2%).

How It Works

The authors treat embodied synthesis as an extension of foundation video and image generation rather than a separate problem. Their key insight is that a single autoregressive (token-by-token prediction) architecture can handle both:

• Standard vision tasks: text-to-image, image editing—the bread and butter of pre-trained models.
• Embodied tasks: generating consistent multi-view scenes (crucial for robots that perceive from different angles), embodied transfer (editing a scene to show a robot performing a new action while preserving geometry), and embodied video generation (predicting future frames that respect both physics and robot morphology).

By training jointly on all these tasks, the model learns to preserve visual priors from its foundation model pre-training while learning robot-specific constraints (e.g., that a gripper must touch an object realistically, or that multiple views must depict the same scene). The unified approach avoids the common trap of fine-tuning on scarce robot data, which typically erodes the general visual knowledge acquired from billions of internet images.

Why It Matters

Generalization and scalability: Foundation models have proven their ability to generalize across diverse visual concepts and scenarios. Xiaomi-Robotics-U0 demonstrates that this generalization extends to embodied intelligence—the model ranks first on World Arena for embodied video generation and beats GPT-Image-2.0 in human evaluations of robot scene generation. This suggests that scaling foundation models, rather than building bespoke robotic models, is a viable path.

Practical impact: The real-world test is the most compelling result. On a challenging manipulation benchmark with out-of-distribution scenarios, the model improves robot success rates from 36.9% to 63.2%. This isn't just a leaderboard win; it's a concrete boost in embodied task performance. The model can also serve as a data engine—synthesizing diverse robotic training scenarios that improve downstream controllers.

Technical novelty: Structured, controllable embodied transfer (fine-grained editing of robot interactions while maintaining multi-view consistency and dynamics) is new. Prior work struggled to edit embodied scenes without breaking geometric or physical coherence.

The authors position this as evidence that foundation world models can play a dual role: as embodied world models for simulation and reasoning, and as data generators for embodied learning. Code and checkpoints are publicly available, making this a potentially high-impact contribution to embodied AI.