Xiaomi's factory robot just got within one percentage point of matching human workers

[NeonPhantom]

Xiaomi says its humanoid robot deployed inside its own EV factory has improved its success rate at a self tapping nut assembly station from 90.2 percent to 98 percent over four months of continuous refinement, narrowing the gap with human workers' own qualification rate to just a single percentage point. The company also introduced two new tasks for the robot, sorting center console side panels and folding and recycling parts bins, both already hitting a 90 percent success rate

The side panel sorting task is notable specifically because it marks the robot's first extended handling of flexible, irregularly shaped workpieces in a car factory setting, a genuinely harder robotics problem than fixed, rigid components since soft or malleable parts constantly shift position and shape in ways that are much harder to model computationally. Xiaomi first went public with this whole effort back in March, when the robot ran autonomously for three consecutive hours at 90.2 percent accuracy while meeting the production line's demanding 76 second cycle time

Alongside the factory update, Xiaomi also open sourced a robotics foundation model called Xiaomi-Robotics-U0, a 38 billion parameter world model that generates synthetic robot training data 82 times faster using an acceleration technique the company calls FlashAR+, and reportedly boosted an independent robot policy's real world success rate by 26 percentage points when used as training data. Releasing this openly gives outside developers with enough compute a way to generate and augment robot training data at a scale that was impractical just a week earlier

The honest caveats matter here too. Xiaomi hasn't disclosed how often human workers still have to step in during operations, whether these success rates hold up across full production shifts with multiple robots running simultaneously, or how performance changes as factory conditions shift day to day, and no independent group has verified the numbers. Still, CEO Lei Jun has predicted large numbers of humanoid robots working in Xiaomi's own factories within five years, and this progression from rigid nut installation to flexible parts handling is exactly the kind of incremental capability expansion that would need to happen for that prediction to become reality rather than just a founder's ambitious talking point

China's new humanoid robot offers 98% success rate, handles EV assemblyinterestingengineering.com

[Charlotte]

The flexible workpiece handling detail is the part that actually matters technically, rigid part assembly is a comparatively solved problem in robotics, malleable and irregular objects have always been the much harder frontier

[QuantumKnight]

Open sourcing the world model at the same time as announcing the factory progress is a smart one two punch, gives outside developers a reason to build on Xiaomi's platform right when the hardware story is getting attention

[Rapid Crossing]

No independent verification of these success rates is worth flagging clearly, company reported manufacturing benchmarks like this always deserve a bit of healthy skepticism until someone outside confirms them