Xiaomi Robotics-1 embodied foundation model begins apartment trial operation
Xiaomi demonstrated a 30-plus minute long-horizon run across three rooms and more than 50 tasks, now in trial operation at Xiaomi youth apartments.
Xiaomi technical official account published a video on November 3 introducing applications of the Xiaomi-Robotics-1 embodied foundation model.
Using Xiaomi-Robotics-1 as the technical base, Xiaomi built an agent closed-loop system covering data collection, model inference and task execution, achieving a long-horizon task validation running over 30 consecutive minutes, covering three rooms and containing more than 50 tasks. The system has begun trial operation at Xiaomi youth apartments, able to plan autonomously and complete tasks such as desk tidying, floor sweeping, item sorting and shelving, room inspection and whole-home re-checking.
Xiaomi formally released Xiaomi-Robotics-1 on July 16, 2026. The model was pre-trained on more than 100,000 hours of UMI data and post-trained on more than 10,000 hours of cross-embodiment data. Pre-training used 100,000 hours of real-world manipulation trajectories covering home, office, restaurant, outdoor and industrial settings, containing more than 1,700 scenario types, 2.4 million interaction segments and more than 260 distinct tasks. Xiaomi-Robotics-1 ranked first overall on both the RoboCasa 365 and RoboDojo leaderboards in July 2026, and its automated annotation pipeline can complete high-quality annotation of the full dataset within two weeks.
In industrial settings, Xiaomi robots have achieved stable operation at 94 to 98 percent success rates on work stations including self-tapping nut loading, cover sorting and bin folding at the Xiaomi automobile factory. Another Xiaomi robot VLA model, Xiaomi-Robotics-0, uses a split large-and-small brain architecture and achieves 80 millisecond inference latency on consumer graphics cards.