Xiaomi-Robotics-1
Points and comments are a snapshot, not live.
Xiaomi scales robot foundation models via 100K hours of embodiment-free pre-training.
Xiaomi-Robotics-1 uses 100,000 hours of embodiment-free (UMI) trajectories for pre-training, covering 1,700+ scenarios. Post-training uses 7,200 hours of real-robot data plus open-source data. The model shows clean scaling: validation action error decreases with more data and larger model size; real-robot success rate rises predictably. After post-training, it adapts to new tasks (e.g., phone packing, laundry loading) with under 10 hours of demonstrations, achieving 75% success rate (vs. 40% for π 0.5). It sets state-of-the-art results on four simulation benchmarks (RoboCasa, RoboCasa365, VLABench, RoboDojo). The approach breaks the data bottleneck for scaling robot policy models.
What commenters are saying
Commenters were impressed by the functional JS-free page and the uncut video showing real robot behavior. Several discussed embodiment generalization, noting the training uses a standardized gripper and camera setup (UMI) while being arm-agnostic. Two camps emerged: those praising the 10B parameter model's efficiency, and those skeptical about real-world reliability, citing the robot's loop when folding shorts. Others debated the practicality of laundry-folding robots versus household chores, with some coining 'slopfold' for imperfect folding. A user noted the cooperation between robots in the video.