DigiNews

Summary

Xiaomi-Robotics-1 introduces a foundation model for robot manipulation trained with over 100K hours of embodiment-free trajectories. The article details a two-stage training approach (pre-training on large UMI data, post-training for embodiment and instruction alignment) and shows scaling benefits in real-world tasks and four simulation benchmarks. It also highlights open-source links to the project code and model, signaling a practical, data-driven path for robotics AI.