Unified Embodied Synthesis

The direct application of powerful foundation image and video generation models to embodied AI scenarios has been hampered by the stringent requirements for multi-view consistency, geometric coherence, and robot embodiment constraints. Existing approaches often necessitate fine-tuning with limited robot-specific data, thereby diluting the extensive visual knowledge gained during large-scale pre-training.

Unified Embodied Synthesis Framework

Xiaomi-Robotics-U0, a 38-billion-parameter multimodal autoregressive model, redefines embodied generation by treating it as a direct extension of foundation image and video generation. This novel approach jointly optimizes text-to-image generation, image editing, embodied scene generation, embodied transfer, and embodied video generation within a single, unified framework. This strategy crucially preserves the generalization capabilities of the pre-trained world foundation model while adeptly adapting it to embodied settings. As detailed in their work, Xiaomi-Robotics-U0 is the pioneering model capable of high-quality multi-view scene generation across diverse robot embodiments. It also introduces structured, controllable embodied transfer for fine-grained editing, all while maintaining multi-view consistency and interaction dynamics.

State-of-the-Art Embodied AI Performance

The model demonstrates significant advancements, achieving state-of-the-art results on both single-step and sequential embodied generation tasks. Human evaluations show it outperforms GPT-Image-2.0 in embodied scene generation and transfer. Furthermore, Xiaomi-Robotics-U0 secured the top rank on the World Arena for embodied video generation. Critically, on challenging real-world manipulation tasks, it boosted the out-of-distribution success rate of pi_0.5 from 36.9% to 63.2%. These outcomes strongly indicate that foundation world models can effectively function as both embodied world models and scalable data engines for advancing embodied intelligence.