Xiaomi scales robot training to 100,000 hours

Xiaomi-Robotics-1 uses 100,000 hours of embodiment-free trajectories for pre-training and more than 7,200 hours of real-robot data for post-training. The project reports 75% overall success after fewer than 10 hours per new task and 85% after fewer than 40 hours, with gains carrying into unseen environments and four simulation benchmarks, according to the accompanying paper. For robot builders, the useful fact is a data strategy rather than a humanoid demo: broad UMI pre-training can make scarce embodiment-specific demonstrations more valuable, reinforcing the simulation funnel behind modern physical-AI research.