Xiaomi Robotics

“Xiaomi-Robotics-0 is benchmarked as an external baseline, scoring 54.3% on UniVTAC and 23.4% on NeoSim (Tables 1–2). Xiaomi-Robotics-1 is referenced as pretraining 'on more than 100,000 hours of real-world trajectories'”

“The central contribution is demonstrating that massive-scale pre-training on UMI (handheld gripper) data — not robot teleoperation data — produces manipulation capabilities that transfer to actual robots after post-training. The model is pre-trained on over 100,000 hours of real-world manipulation trajectories collected via UMI devices... achieves 75% out-of-the-box success rate across four tasks in unseen environments, compared to 26% for a baseline without action pre-training”