Xiaomi open sources embodiment-generative large model U0, aiming to "mass-produce data" for robots

```

Xiaomi is extending the capabilities of large models further into the field of robotics.

On July 15, Xiaomi officially released and open-sourced the embodied generative model Xiaomi-Robotics-U0 (hereinafter referred to as "U0"). This model contains 3.8 billion parameters, unifying embodied world modeling and general image generation capabilities within a single multi-modal autoregressive architecture, mainly used for generating, transferring, and augmenting robotic training data.

U0's goal is to "create data" for robots.

Robots need to learn actions like grasping, carrying, and organizing, which require repeated training in a variety of different scenarios. However, collecting real-world data is costly and time-consuming, and dangerous, extreme, or low-frequency scenarios are hard to fully cover.

U0 attempts to turn this process into a mass-produced data pipeline. Existing real-world robot data can have objects, lighting, materials, and backgrounds changed directly, without the need to rebuild scenes, redeploy equipment, or repeat data collection; difficult-to-obtain long-tail scenarios in reality can also be generated directly via the model.

U0 also proposes the FlashAR+ accelerated high-speed inference solution. Through diagonal parallel decoding and cache scheduling technology, the time to generate a high-definition training image at 1024×1024 resolution is compressed from 450.77 seconds to 5.44 seconds, boosting efficiency by 82.9 times.

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

U0 unifies four tasks—embodied scene generation, embodied trajectory migration, robotic interaction video generation, as well as general text-to-image and image editing—within one model, basically connecting the data production chain of "scene generation—trajectory migration—environment expansion—interaction process generation".

In fact, similar open-source attempts are already emerging in the embodied intelligence industry. In March 2025, Quhu Technology open-sourced the spatial understanding model SpatialLM. This model can transform video or point clouds into structured 3D scenes containing walls, doors and windows, furniture, and spatial relationships. Enterprises can fine-tune the model for their own scenarios to enhance robots' understanding of physical space.

Currently, the embodied intelligence industry still faces challenges such as insufficient training data, limited scenario coverage, and high R&D costs. Open-source models cannot fully replace real-world robot data, nor can they solve all the complex physical interactions between robots and real environments, but they can help reduce the cost of data augmentation and model training to some extent, accelerating the transition of robots from the laboratory to factories, warehouses, homes, and other real-world settings.

Risk Warning and DisclaimerThe market has risks, and investment requires caution. This article does not constitute individual investment advice, nor does it take into account any specific user's investment objectives, financial situation, or needs. Users should consider whether any opinions, viewpoints, or conclusions in this article are suitable for their particular situation. Invest accordingly at your own risk. ```