Xiaomi has unveiled a robotics foundation model, Xiaomi-Robotics-1, that aims to close the gap between laboratory skills and real-world generalization. The model is trained to follow spoken or written commands in unfamiliar settings and to adapt to new tasks with minimal additional data. Rather than relying on expensive robot fleets during data collection, Xiaomi recorded over 100,000 hours of manipulation footage using a portable handheld gripper, then used automated labeling and transfer training to deploy the model on physical robots.
Handheld grippers instead of robot fleets
To scale data collection across diverse settings, the team used a compact handheld gripper equipped with cameras. A human operator simply picked up the gripper and performed manipulation tasks in kitchens, offices, stores, factory floors and outdoor spaces. That approach produced more than 100,000 hours of UMI recordings from over 1,700 different environments, enabling coverage far beyond what is practical with teleoperated robot arms.
Labeling such a vast dataset by hand would have been impractical, so Xiaomi applied an auxiliary AI model to describe each motion segment in text. The researchers say this automated labeling process annotated the full dataset in roughly two weeks. After pretraining on the handheld recordings, Xiaomi transferred the learned behaviors to a variety of physical platforms, including wheeled robots and dual-arm systems, accounting for the mechanical differences between a hand-held device and robot manipulators.
Data volume matters more than model size
In controlled tests, increasing model size improved performance, but expanding the training dataset produced substantially larger gains. The team observed that as they increased pretraining data, success rates in unfamiliar environments climbed from about 25 percent to roughly 75 percent. They report that they have not yet seen a plateau where additional data stops improving results.
These scaling patterns mirror findings from other visual-learning studies: when a model must process images and actions rather than only text, adding more varied data can be more impactful than simply enlarging the network. Xiaomi-Robotics-1 reportedly set new top marks on standard robot AI benchmarks and demonstrated complex behaviors in demos, including a hands-off suitcase-packing sequence that took more than ten minutes and required navigating across a room.
Fast adaptation to new tasks with limited examples
A useful foundation model must adapt quickly to novel tasks. Xiaomi tested Robotics-1 on four representative tasks—packaging a phone, loading laundry into a washing machine, feeding paper into a printer, and packing items into a box. With under ten hours of task-specific training data for each task, the model achieved an average success rate of 75 percent. For comparison, a competing model from Physical Intelligence reached about 40 percent on the same tasks.
The researchers highlight particularly strong performance on tasks involving soft, deformable materials such as paper and on tasks that required moving around a room rather than remaining in a fixed workspace. On public leaderboards, Xiaomi-Robotics-1 leads by a wide margin on RoboCasa365 and scores roughly 58 percent higher than the runner-up on RoboDojo, although absolute success rates on some benchmarks remain modest.
Competing approaches and open research
Robotics teams are pursuing different strategies to solve the data bottleneck. Teleoperation yields precise labels but is costly and restricted to a few controlled settings; handheld grippers are intended to broaden coverage at much lower cost. Other projects adopt alternative trade-offs: Nvidia, Carnegie Mellon University and UC Berkeley’s ENPIRE work used AI agents to teach a fleet of eight robots to grasp autonomously, and Nvidia has also explored producing synthetic training data to sidestep collection limits.
China’s BAAI has taken a different path with its Orca world model, which reportedly learns from 125,000 hours of video without action labels and can match a specialized model on several manipulation tasks. Xiaomi’s approach automates labeling and applies it at scale, while Orca attempts to eliminate labels entirely. The debate over generalization surfaced earlier this year when Physical Intelligence drew criticism after claims that one of its models may have recalled training scenes rather than truly generalizing. Xiaomi’s emphasis on larger, varied datasets adds another dimension to that discussion.
As part of its broader robotics roadmap, Xiaomi says its MiMo-V2 agent family could eventually control robots, and the company plans to release Xiaomi-Robotics-1 and its code on GitHub and Hugging Face. The same team recently published Xiaomi-Robotics-0 as an open-source effort focused on fast, real-time operation, signaling a continuing commitment to shared tools and benchmarks in robot learning research.
Source: The Decoder