Xiaomi-Robotics-1 - A Robot Foundation Model Trained on 100,000 Hours of Real-World Data

Xiaomi announced Xiaomi-Robotics-1, a model that combines 100,000 hours of embodiment-free (UMI) trajectory data with 7,200 hours of real robot data to address the large-scale data shortage problem.

Training proceeds in two stages:

• Pre-training: General motion generation capability is learned from large-scale UMI data collected across more than 1,700 scenarios, with state transitions learned via an auto-labeling pipeline that leverages a VLM.
• Post-training: The learned capabilities are aligned to real robots, going through an Instruction alignment process so the model can directly follow natural language instructions.

The research confirmed a Scaling Law in which the real robot's task success rate steadily increases in a predictable manner as data and model size grow. This demonstrates that the model has high generality even for new environments and objects.