Xiaomi-Robotics-1

Xiaomi-Robotics-1 is a VLA-style robot foundation model developed by Xiaomi Robotics team. It is trained on over 100,000 hours of real-world manipulation trajectories using large-scale embodiment-free (UMI) pre-training across more than 1,700 scenarios spanning household, commercial, industrial, and outdoor environments. Post-training uses cross-embodiment real-robot data including over 7,200 hours of in-house home data for embodiment and instruction alignment. The model demonstrates clean scaling behavior where more pre-training data and larger model size improve real-robot success in unseen environments, achieving approximately 75% success rate with fewer than 10 hours of demonstrations per task on average for packing, refilling, and laundry tasks. It achieves state-of-the-art results on simulation benchmarks including RoboCasa and VLABench.