On November 3, Xiaomi disclosed that it has started a robot “trial operation” in Xiaomi Youth Apartments. The system is built on Xiaomi-Robotics-1, the company's embodied foundation model, and wraps it in an agent loop that spans data collection, model inference and task execution. For now it is a small-scale pilot.
The one long-horizon validation result Xiaomi offered: continuous operation for more than 30 minutes, across 3 rooms, covering more than 50 tasks.
The run began with tool preparation, then moved through tidying the sofa and tabletops, clearing trash, sorting items and putting them back, a room inspection and a whole-home recheck, before the robot left the site. Xiaomi says the system planned these tasks on its own.
A rough calculation: 50-plus tasks in 30 minutes is about half a minute each. That suggests the “tasks” here are fine-grained, closer to “put one thing back where it belongs,” and Xiaomi has not published a precise definition of a task.
The model and the benchmarks
Xiaomi-Robotics-1 was released and open-sourced in July this year. It was pretrained on 100,000 hours of real manipulation trajectories collected with UMI devices, drawn from homes, commercial spaces, factories, offices and outdoor settings, and then post-trained for different robot bodies. Xiaomi describes it as “ready to use out of the box.”
The published results are below, as reported by Xiaomi and relayed by media outlets:
| Benchmark | Xiaomi-Robotics-1 | Previous best |
|---|---|---|
| RoboCasa365 average success rate | 57.4% | 46.6% |
| RoboDojo average score | 20.07 | 13.07 |
| RoboDojo average success rate | 13.93% | 8.80% |
The gains are sizable, but these are mostly results from simulated environments, and RoboDojo's success rate is still only 13.93%. A lot sits between a leaderboard and a real apartment, and that gap is what this trial is meant to address.
The factory update disclosed the same day
Xiaomi also announced a second setting: a humanoid robot at an automotive plant, working a console side-panel sorting station, attempted its first 8-hour continuous shift with the goal of meeting the station's full-day production demand. The company says the robot corrected abnormal deviations on its own, coordinated with the production line system, and has “preliminarily completed” validation of line-level operating capability.
Looking back one step, what Xiaomi showed at the World Robot Conference in August was a self-tapping nut loading station in a die-casting workshop: 3 hours of continuous autonomous running, a 90.2% installation success rate, and a 76-second line takt time met. That time there was a success rate and a takt time; this time, for the 8-hour run, there is neither.
The two disclosures are framed differently, and reading them side by side makes that clearer. A factory has fixed stations and fixed materials, so metrics are easy to define; in an apartment the sofa, tabletops and trash differ every time, and there is no takt time to compare against. Xiaomi's wording, “trial operation” and “preliminarily completed,” is more restrained than the language typical at product launches.
What it means for Chinese readers
Xiaomi Youth Apartments are housing the company provides for its employees. Public reports put the average monthly rent at 1,999 yuan in both Beijing and Nanjing, with fresh graduates given priority. In other words, the trial is taking place inside Xiaomi's own premises, and residents are most likely its own staff, which leaves far more room for error than a consumer setting would.
That also determines how far this is from consumers. Xiaomi has not announced a price, a commercialization timeline or a rollout plan, so ordinary users will not be able to buy or use it any time soon. For Chinese companies building home robots, the reference value lies in the choice of setting: a staff dormitory offers a real domestic environment, allows continuous data collection, and carries none of the after-sales risk of selling to households.
Xiaomi lists “data collection” as one link in the loop, so the apartments may become a source of data for later training. What would be collected, and how residents would give consent, is not mentioned in any public information.
What has not been disclosed
Not made public: the hardware form and the robot model used, the length of the apartment trial and its success rate, whether humans intervened during the run and how many times, and any failure cases. The success rate, takt time and anomaly frequency of the 8-hour factory task are likewise undisclosed.
All of the figures above come from media accounts of Xiaomi's own disclosures, and there is currently no third-party evaluation to check them against.
Sources: Xiaomi Technology official account, Kuai Ke Ji, CocoLoop, ITHome, Cailian Press; benchmark results and factory station data follow Xiaomi's own disclosures.