Xiaomi's factory intern robot hits 98% on nut assembly, 90%+ on two new tasks

In plain terms

Xiaomi has published performance numbers for the "intern" humanoid robot deployed in its EV factory. On the original task - mounting nuts on car bodies - the success rate reached 98%, up from 90.2%. On two new tasks it has since been assigned (sorting interior parts and folding reusable logistics boxes), the robot logs "above 90%".

The story

ETNews reports (18 July) that Xiaomi has broken down the performance of the humanoid robot it introduced into its EV plant earlier this year. The headline number - 98% success on nut-assembly on car bodies, up from an initial 90.2% - comes with Xiaomi's own framing that this is within 1 percentage point of skilled human workers. Since that ramp, the robot has been assigned two additional tasks: sorting interior parts, and folding reusable logistics boxes for internal supply. Both new tasks are logged at above 90% success. Xiaomi says its next work is on force-controlled manipulation.

The embodied-foundation-race framing matters here. LimX Dynamics, X-Square Robot and Nvidia GR00T have all been claiming that a single "backbone" model will unlock robotics the way LLMs unlocked language. The counter - from Figure and Agility Robotics - has been that industrial reliability requires purpose-built stacks, not general backbones. Xiaomi's disclosure tilts the argument in a specific direction: on scoped, repetitive assembly tasks inside a plant Xiaomi controls, the foundation-model path can hit near-human numbers. On new tasks it hits 90%+ from the start - not 98% - which is the honest data point to hold.

Under the hood

• Task 1 - nut assembly on car body: success rate 98%, up from 90.2% initially (Xiaomi self-report via ETNews).
• Task 2 - interior parts sorting: above 90%.
• Task 3 - folding reusable logistics boxes: above 90%.
• Gap to human: Xiaomi frames the 98% as ~1 percentage point below skilled human workers on the same task.
• Deployment site: Xiaomi EV factory, China.
• Related model: Xiaomi Robotics-U0 (38B, open-source, ~2 weeks ago per publication #1139).
• Next stated direction: force-controlled manipulation.

So what

Three implications. For industrial automation buyers, this is the first China-side data point comparable to Figure's Boeing/BMW work - the "does it work in a real plant?" question now has a task-by-task answer, not a single number. For the embodied race, the vertical-integration model (own factory, own robot, own foundation model, own data pipeline) is the one showing the fastest ramp on scoped tasks; that's a headwind for third-party embodied platforms that need to sell into someone else's plant with someone else's data. For rivals - particularly Agility Robotics and Figure - the challenge is now to publish equivalent breakdowns in equivalent environments. Watch for two follow-ups: independent verification of the 98% figure, and how quickly the two 90%+ tasks are pushed toward the same near-human threshold. That ramp curve is the real signal, not the headline.