Xiaomi Open-Sources Embodied Generative Large Model U0 to Mass-Produce Data for Robots

Xiaomi is further extending large model capabilities into the robotics sector.

On July 15, Xiaomi officially released and open-sourced the embodied generative model Xiaomi-Robotics-U0 (hereinafter referred to as “U0”). With 38 billion parameters, the model integrates embodied world modeling with general-purpose image generation capabilities within a unified multimodal autoregressive architecture. It is primarily designed to generate, transfer, and augment robotic training data.

The goal of U0 is to “create data” for robots.

For robots to learn actions such as grasping, carrying, and organizing, they require repeated training across a vast array of different scenarios. However, data collection in the real world is costly and time-consuming, making it difficult to adequately cover dangerous, extreme, or low-frequency scenarios.

U0 aims to transform this process into a data production line capable of batch operations. Existing real-world robotic data can be directly modified by swapping objects, lighting, materials, and backgrounds, eliminating the need to rebuild scenes, deploy equipment, and repeat data collection. Long-tail scenarios that are difficult to capture in reality can also be generated directly by the model.

U0 also introduces FlashAR+, a high-speed inference acceleration solution. By leveraging diagonal parallel decoding and cache scheduling technologies, it reduces the generation time for a single high-resolution training image (1024×1024) from 450.77 seconds to just 5.44 seconds, boosting efficiency by 82.9 times.

This means robotics companies can mass-produce training data covering diverse backgrounds, lighting conditions, and objects in a relatively short period.

U0 unifies four types of tasks within a single model: embodied scene generation, embodied trajectory transfer, robotic interaction video generation, and general text-to-image and image editing. This essentially connects the entire data production pipeline of “generating scenes – transferring trajectories – expanding environments – generating interaction processes.”

In fact, similar open-source initiatives have already begun to emerge in the embodied intelligence industry. In March 2025, Qunhe Technology open-sourced SpatialLM, a spatial understanding model. This model can convert videos or point clouds into structured 3D scenes containing walls, doors, windows, furniture, and spatial relationships. Enterprises can fine-tune the model for their specific scenarios to enhance robots' understanding of physical spaces.

Currently, the embodied intelligence industry still faces challenges such as insufficient training data, limited scenario coverage, and high R&D costs. While open-source models cannot completely replace real-world data nor resolve all complex physical interactions between robots and real environments, they can help reduce the costs of data augmentation and model training to some extent, accelerating the deployment of robots from laboratories to real-world settings such as factories, warehouses, and homes.