Robots learn by observing, imitating and practising, but unlike language models they have no ‘physical-world internet’ to scrape, so embodied training needs paired data on what the robot saw, what it did, and how that changed the world around it. Human and real-robot collection is hard to scale, and synthetic data from separate models often breaks scene consistency. Xiaomi’s answer, Xiaomi-Robotics-U0, open-sourced this week, is the first unified embodied-generation model, merging scene generation, embodied transfer, robot-interaction video synthesis and general image editing in one framework.
As the first unified model in the embodied-generation space, U0 has already validated generation quality, data usefulness and engineering efficiency. In the WorldArena benchmark built by Tsinghua, Peking and others, U0 scored first in the world under an anonymous code, and policies trained on its expanded data lifted task completion by 26.3 per cent in out-of-distribution settings with unknown lighting and backgrounds. Its FlashAR+ accelerator cut single-sample generation at 1024 by 1024 from 450.77 seconds to 5.44 seconds, an 82.9-fold gain that brings bulk embodied-data production within practical cost.
One model, controllable data
U0’s core innovation is covering four generation tasks in one model, so a captured arm trajectory can be re-skinned with new object looks, lighting or backgrounds without re-collecting real data, and entirely new workbenches and long-tail edge cases can be synthesised from scratch. A five-axis decoupled control splits generation into workbench layout, foreground object, unrelated clutter, lighting and background, each adjustable by language while preserving structure and trajectory, so modified frames still match the original action labels and stay usable for policy training.
By folding unity, controllability and low cost into one system, U0 targets the real bottleneck in embodied AI, not model cleverness but data at scale. For Western labs, the notable move is that a major Chinese consumer-electronics firm is giving this data infrastructure away, raising the baseline for everyone racing to make robots that work outside the demo reel.
Read the original report (in Chinese)
*Translated and adapted from LeiPhone (https://www.leiphone.com/category/robot/2m3jnMIPtYerJ6UR.html).*