Xiaomi-Robotics-1 combines 100,000 hours of embodiment-free UMI pre-training across 1,700+ scenarios with 7,200 hours of in-house real-robot data. An auto-labeling pipeline powered by a vision-language model annotates trajectories with state-transition descriptions, enabling large-scale pre-training that shows clean scaling: validation action error decreases steadily as data and model size grow.

Post-training aligns the model via embodiment alignment (cross-embodiment real-robot data) and instruction alignment (natural-language command following). After post-training, the model performs mobile manipulation tasks in unseen environments, with real-robot success rates rising predictably as pre-training scale increases — showing no saturation.

For new tasks, Xiaomi-Robotics-1 achieves 75% overall success with under 10 hours per task on average, nearly doubling the π0.5 baseline (40%); at under 40 hours, success reaches 85%. It also sets state-of-the-art on four simulation benchmarks (RoboCasa, RoboCasa365, VLABench, RoboDojo), with relative gains up to 58%.

The work demonstrates a practical path for scaling robot foundation models: embodiment-free pre-training breaks the data bottleneck, and alignment transfers general capability to physical robots with measurable, predictable scaling gains.