On 4 November Lei Jun posted a video that set the Chinese tech press alight. A humanoid stood at the side-panel sequencing station on Xiaomi’s EV final assembly line and worked a full shift of 8 hours and 5 minutes.
The same day brought a second disclosure. A robot on the same base platform had been doing chores in a Xiaomi staff apartment for half an hour, moving through three rooms and clearing more than 50 tasks.
The same week, at the Wuzhen summit, Unitree founder Wang Xingxing poured cold water on the whole exercise. Success rates look good, he said, but put the machine on a real factory floor and the economics still lose to a human, because the problem is stuck in the last few millimetres of manipulation.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://robot.ofweek.com/2026-11/ART-898890-8110-30711032.html.
What the station actually asks for
SPS sequencing sounds technical and is not. Take a flexible side cover out of a crate, square it up, and slot it into a rack in the order the vehicles will arrive. The covers are floppy. Every one comes out of the crate in a different attitude. A change in lighting or a shuffled bin is enough to break the grasp, and a bad placement stops the whole line behind it. Over eight hours the temperature drifts, the lamps shift, supply ebbs and surges, and nobody writes a script for any of it.
The progression is the interesting part. In March the same robot ran three hours, inserting self-tapping nuts at 90.2 per cent two-sided success inside a 76-second takt. Lei Jun looked at the data and said it felt like the industry could finally use this. By July the same station hit 98 per cent, with cover sequencing and empty-box recovery added. November is the first eight-hour attempt.
The shift was not clean. It recharged twice, dropped a cover once, and a wheeled robot came over to pick the part up. It is not a standalone hero. It is one component inside an automation system that was already working.
The video was sped up, and the plant that buys anyway
Viewers worked out that the continuous shot was accelerated, at rates running from 5x through 30x and 60x to 100x. Reverse the maths and even at 5x the motion looks leisurely. The work was real. The presentation was flattered.
The more useful case study sits outside Xiaomi. A listed wheel-hub bearing manufacturer in Hangzhou runs 15 humanoids between narrow production lanes, picking parts with 80mm to 120mm bores, identifying, gripping, transferring and loading in a 40-second cycle at 98 per cent loading success, across more than two months and 1,000-plus hours and over 60,000 parts.
The plant manager wants 15 seconds, which is exactly what a skilled worker takes. The robot needs 40. And the factory is willing to spend a year closing the gap, partly because it owns 80 per cent of the joint venture and the robot company owns 20. Equity explains who can afford to buy. It does not explain who is willing to wait.
Four accounts explain the patience. On speed, 98 per cent is easy in a lab and hard against scratches, oil on the lens, parts stacked carelessly and sunlight arriving at a different angle each morning. On stability, two failures per hundred means an engineer watching four machines for dropped parts and jams. On data, 15 machines return 800 to 1,200 records a day, more than 40,000 real-machine records a month, all classified and fed back, with the grasp dataset donated in batches to the OpenAtom OpenLET foundation. On payback, the estimate on site is about three years.
Xiaomi keeps the problem inside its own factory, where a dropped part goes straight back to its own engineers. Unitree pushes the integration risk to partners. Both are answers to the same question, which is who pays for the year before a robot is worth its wage.