Xiaomi has released and fully open-sourced Xiaomi-Robotics-U0, a 38-billion-parameter multimodal autoregressive foundation model for embodied AI that unifies four distinct robot task categories within a single framework1,2. The model's code and weights are available on GitHub and HuggingFace.
Xiaomi-Robotics-U0 covers four core capabilities. Embodied scene generation creates multi-view initial scenes for specified robot hardware from text descriptions, spanning environments from tabletops and kitchens to warehouses and open worlds. Embodied transfer migrates existing robot trajectories to new environments — changing lighting, background, surface materials, target objects, or workspace style — while preserving original arm poses and scene layout. Robot interaction video generation produces subsequent video frames based on initial observations and operation instructions, maintaining motion coherence and physical consistency with zero-shot generalization to unseen scenarios. The model also retains general text-to-image and image editing capabilities, allowing internet-scale visual knowledge to transfer to embodied AI tasks.
Xiaomi describes Xiaomi-Robotics-U0 as the first unified generative model in the embodied AI field capable of handling all four task categories simultaneously. The model establishes a complete pipeline for generating and editing robot image and video training data.
On benchmarks, Xiaomi-Robotics-U0 achieved top scores on the WorldArena benchmark across 126 participating global models. In real robot evaluations under out-of-distribution conditions, strategy task completion rates improved an average of 26% when trained on data augmented by the model, according to Xiaomi. The model uses a UNIS inference acceleration architecture that Xiaomi says improves generation efficiency by approximately 83 times compared to the raw autoregressive paradigm.
The practical implications center on synthetic data generation for robotics. The model can enhance existing datasets by altering objects, lighting, backgrounds, or adding clutter without requiring fresh physical data collection. It can also generate entirely new scenes covering hazardous, extreme, or long-tail environments that are inaccessible to physical robots.
ANALYSIS The open-source release of a model at this scale — with weights, code, and a full data-generation pipeline — lowers the barrier for robotics labs that lack the resources to collect diverse real-world training data. ANALYSIS The 26% task-completion improvement claimed under out-of-distribution conditions, if independently replicated, would represent a meaningful contribution to sim-to-real transfer in embodied AI.
The release follows Xiaomi's robot factory deployment. The open-source package includes the full project page, GitHub code repository, HuggingFace model weights, and a ModelScope collection.