Xiaomi humanoid robots hit 98% on nut loading, add 90% sorting/folding tasks in EV factory

Xiaomi updated its humanoid CyberOne Gen2's training inside its own EV factory, and the core shift is a move from single-station validation to multi-station, flexible-part work on a real production line. Self-tapping nut/screw tasks have climbed to a 98% success rate, while new long-horizon tasks—center-console/side-panel sorting and bin folding/recycling—first reached around 90%. Several posters read it as Xiaomi pushing humanoids from demos toward sustainable factory automation.

Key details

Multiple posts say that since Q1, Xiaomi's humanoids have been training on dual-sided workstations in the car factory, with the self-tapping nut loading or screw workstation hitting a 98% success rate. @teortaxesTex pegs that figure as rising from 90.2% to 98%, while @Dr_Singularity says it was reached after a quarter of optimization and sits just 1 point below humans. CyberOne Gen2 also autonomously performs two new tasks in the final-assembly logistics zone: sorting large flexible center-console/side-panel parts, and folding and recycling totes or bins. Posts report both new tasks above 90% success and capable of long, continuous operation.

Technology and impact

Several posters stress that the showcase is less about single-point accuracy than the move from single-station to multi-station and from fixed motions to flexible parts. @broodsugar attributes the progress to a control-stack upgrade—whole-body reach, dual-arm regrasp, and global planning—and notes multi-robot collaboration. Per @CyberRobooo, Xiaomi was also said to have quietly released the autoregressive world model Xiaomi-Robotics-U0, whose "Real Robot Improvement" co-uses U0-generated style-transferred demonstrations with real demonstrations for policy post-training, to improve adaptation to unfamiliar tasks.

2026-07-14 ~ 2026-07-15 · 9 related posts

2 near-duplicate retellings: PTrubey · tak3sh8