On July 16, 2026, Xiaomi released Xiaomi-Robotics-1, a robot foundation model that combines 100,000 hours of embodiment-free pre-training trajectories with 7,200 hours of real-robot data collected in actual homes. The project's central contribution is empirical: as model size and pre-training data grow, real-robot task success rates rise steadily and show no signs of saturation.
The pre-training dataset spans over 1,700 scenarios across household, commercial, industrial, and outdoor environments. To manage this scale, Xiaomi built an automatic annotation pipeline powered by a vision-language model that splits long videos into fixed-length clips and describes the state transitions of grippers and interacting objects. This creates a large-scale corpus of manipulation trajectories paired with precise language descriptions — without the prohibitive cost of manual labeling.
Post-training aligns the model with real robot embodiments and natural-language instruction following. The in-house real-robot data covers tasks like tidying a sofa, sorting a shoe cabinet, and putting away kitchenware. After this stage, Xiaomi-Robotics-1 can be used out-of-the-box for mobile manipulation in unseen environments with unseen objects. The model achieves a 75% overall success rate on new tasks like phone packing, printer refilling, and laundry loading from an average of under 10 hours of demonstrations per task — nearly double the performance of existing baselines.
Three knowledge points stand out. First, the scaling behavior is clean: validation action error decreases steadily as data and model size grow, and the gains transfer to real-robot performance after post-training. Second, the embodiment-free pre-training strategy — using UMI (Universal Manipulation Interface) trajectories collected without a physical robot — means that data collection can be decoupled from hardware, dramatically lowering the cost of scaling. Third, the results suggest that robotics may finally be entering the data-driven era that transformed NLP and computer vision, where more data and larger models reliably produce better real-world performance.