Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories

Xiaomi introduces Xiaomi-Robotics-1, a vision-language-action model pretrained on more than 100,000 hours of real-world manipulation trajectories collected with UMI devices. Its two-stage recipe combines broad action pretraining with post-training for robot embodiments and imperative human instructions. An auto-labeling pipeline adds natural-language descriptions of scene-state transitions to trajectory clips. The paper reports scaling gains with both data and model size, transfer of those gains to unseen real-robot tasks, and efficient fine-tuning for dexterous tasks. On simulation benchmarks, it reports 57.6% success on RoboCasa365 versus 46.6% for the previous best, and a 20.07 average score on RoboDojo versus 13.07.