Xiaomi-Robotics-1 Tests Whether Handheld Data Can Scale Physical AI

Xiaomi-Robotics-1 was pre-trained on trajectories captured with Universal Manipulation Interface devices. A person holds an instrumented gripper while an egocentric camera records the scene. The system tracks the gripper pose and opening state, producing action data that resembles an end effector without requiring a full robot at every collection site.

That distinction matters. A normal video shows appearance and motion but does not automatically provide the action coordinates a robot controller needs. UMI data are closer to robot demonstrations because the gripper hardware records a manipulable action trajectory while people work in real environments.

Robot teleoperation ties every hour of data to an expensive machine, a trained operator and a controlled deployment. A handheld interface is cheaper to move through homes, offices, shops, industrial sites and outdoor spaces. Xiaomi reports more than 1,700 scenarios in the pre-training corpus.

The advantage is breadth, not perfect embodiment matching. A handheld gripper does not share a mobile manipulator's arm geometry, joint limits, base motion, latency or payload. Xiaomi therefore treats UMI collection as pre-training and uses a separate robot-alignment stage before deployment.

At 100,000 hours, manual task labeling would become a second data bottleneck. The team split trajectories into fixed-length clips and used Qwen3.5-27B to describe how the grippers and interacting objects changed state inside each segment. The paper says a producer-consumer pipeline labeled the full corpus in roughly two weeks.

These captions are descriptions such as an object moving from one state or location to another. During post-training, the language is shifted toward imperative instructions that a person would give a robot. The model must learn the connection between a requested result, the current camera observations and an action chunk.

The post-training mixture contains about 10,000 hours of cross-embodiment data. Xiaomi reports more than 7,200 hours of in-house robot trajectories collected with mobile manipulators and dual-arm robots in real homes, more than 1,000 hours of instruction-labeled UMI data and filtered public robot datasets.

This stage maps the broad action knowledge learned from portable grippers onto actual robot bodies. It also aligns descriptive state transitions with direct instructions. The process does not remove the need for robot data; it attempts to use robot data more efficiently after large-scale pre-training.

For out-of-the-box evaluation, the post-trained model handled shoe storage, bag packing, table organization and sofa tidying. The task types had appeared in post-training, but the evaluation rooms and object instances were unseen. In the reported 5B experiment, overall success rose from 26% without action pre-training to 75% when the full 20,000-hour experimental pre-training subset was used.

A separate adaptation test held out four new tasks: phone packing, laundry loading, printer refilling and box packing. With less than 10 hours of demonstrations per task on average, Xiaomi-Robotics-1 reached 75% average success in the company-run evaluation, compared with 40% for the listed pi-0.5 baseline under the same low-data budget.

The complete corpus exceeds 100,000 hours, but the controlled data-scaling experiment trained a 5B model on fractions of about 20,000 hours because of the compute budget. The model-scaling comparison also used the same 20,000-hour set for the 2B, 5B and 10B variants.

This is still useful evidence: validation action error improved as the data subset grew, and post-training success improved with stronger pre-training. It is not proof that every reported model was fully trained across the entire 100,000-hour corpus, and the paper does not establish an unlimited scaling law.

The project shows a plausible division of labor. Portable UMI collection supplies diverse manipulation experience. Real-robot post-training teaches embodiment constraints. Smaller task-specific datasets specialize the policy for difficult new workflows.

The remaining questions are operational: independent reproduction, failure rates over long deployments, safety around people, performance on embedded compute and robustness when rooms, objects or instructions differ more sharply from training. The results are controlled team-reported experiments, not a commercial household deployment.