Xiaomi says humanoid robot completed eight-hour factory sorting trial

The report does not provide the trial count or a precise definition for that success rate. Nor does it establish that the figure describes the new sorting task. The March and November operations involve different work, so their durations cannot by themselves show an equivalent improvement in productivity or accuracy.

The accounts leave open how much additional engineering would be needed to reproduce the integration at another manufacturer's site.

Company-reported: Xiaomi’s robotics team reports 57.4% success on RoboCasa365, against a previous best of 46.6%. These simulation results do not measure factory sorting performance. Source: Xiaomi Robotics research paper.

The factory and apartment tests measure different capabilities

The disclosures accompanied the World Internet Conference event in Wuzhen, where Xiaomi received an innovation award and its robotics team's general-purpose manipulation-model technology received a leading technology award, according to the account of the event. The awards provide the occasion for the announcement; they do not supply operating data for the factory trial.

Xiaomi is also testing a system built on Xiaomi-Robotics-1 in its youth apartments. The company described task chains exceeding 30 minutes, spanning three rooms and more than 50 tasks, including preparing tools, tidying tables, clearing rubbish, returning objects and inspecting rooms, according to a report on the pilot.

That is a separate test of sustained activity across spaces and tasks. It does not show that the factory robot can perform all those household jobs, or that the apartment system has been demonstrated in customers' homes.

Xiaomi-Robotics-1 is the company's vision-language-action model: software intended to translate visual information and language instructions into robot actions. In a research paper first submitted in July 2026, Xiaomi's robotics team described pre-training on more than 100,000 hours of real-world manipulation trajectories, followed by post-training to align the model with robot hardware and human instructions. Those are learning data, rather than 100,000 hours of autonomous factory operation.

The paper's authors report a 57.4% success rate on the RoboCasa365 simulation benchmark, against a previous best of 46.6%—a difference of 10.8 percentage points. That comparison concerns simulated tasks, not success at Xiaomi's sorting station. The paper also describes adapting the model to new tasks through fine-tuning.

The November reporting says both Xiaomi-Robotics-1 and its predecessor, Xiaomi-Robotics-0, have been open-sourced. It attributes an 80-millisecond inference latency on a consumer graphics card to Robotics-0. That software measure does not establish the total cost or cycle time of a deployed humanoid system.

Figure’s robot hands and forearms. In its company-reported BMW deployment account, Figure identified the forearm as its leading hardware failure point. Image: Figure. Source: Figure.

What would establish production readiness

An eight-hour shift is a Xiaomi milestone, rather than an industry first. Humanoid developer Figure said in a November 2025 deployment account that its second-generation Figure 02 robots had worked ten-hour shifts Monday through Friday at automaker BMW's Spartanburg plant, accumulating more than 1,250 hours of runtime and loading more than 90,000 parts.

Figure's task was sheet-metal loading into a welding fixture, followed by welding and feeding by six-axis industrial robots. That illustrates humanoids working alongside established automation, but its output cannot be ranked directly against Xiaomi's different sorting task.

Figure specified an 84-second total cycle requirement, including 37 seconds for loading. Its placement-accuracy target was above 99% of cycles with all three parts correctly loaded, and its human-intervention goal was zero pauses or resets per shift. Those were requirements and targets, not disclosed aggregate results.

The deployment also exposed hardware constraints. Figure identified the forearm as its leading hardware failure point, citing packaging, dexterity and thermal challenges. It said that experience informed redesigned wrist electronics in Figure 03 as Figure 02 was being retired. Sustained operation can reveal engineering work that a successful short demonstration misses.

For Xiaomi, the next decision is whether the sorting operation can be reproduced reliably and economically. Parts handled per hour, errors and human interventions per shift, repeated-shift performance, and total costs including maintenance and integration would help establish that case. The eight-hour announcement and apartment pilot supply neither those figures nor a cost comparison with existing labor and automation.