On October 19, 21st Century Business Herald put a reporter on a factory floor in Hangzhou. The cell it describes was launched jointly on October 15 by Leju Robotics and the auto-parts maker Zhaofeng: 15 Leju humanoids picking, carrying and loading parts into machines. Sina Finance carried it the same day.
The numbers: a 98% feed success rate, about 1,000 hours of operation, more than 60,000 parts handled.
What the cell does
A full feeding cycle takes under 40 seconds, and the robot’s own share is about 35. Zhaofeng chairman Kong Aixiang wants that near 15 seconds within six months to a year, a cut of roughly 57%. The robots can run an 8-hour stretch, but the report says long-term reliability needs about a year of validation.
Staffing is one engineer for every four robots, stepping in when one faults. Early on, the robots could only grab parts laid out in fixed positions. Now they pick up parts whose position shifts and cope with the odd surprise.
The software side has numbers too. The robots feed back 800 to 1,200 task records a day, over 40,000 a month. Leju says it split the feeding skill into 30-odd base modules, cutting development from months to weeks, and that its multi-robot scheduler brings deployment from days to hours. The two companies plan to donate the real-robot dataset to the OpenAtom Foundation’s OpenLET community and open-source it in batches.
Fifteen next to 1,200
Fifteen is hard to size alone. On September 28 we covered Frost & Sullivan’s half-year report: the world bought 1,200 production-work humanoids in H1 2026. Zhaofeng’s 15 are 1.25% of that, in one plant, on one task.
Last week, Shanghai set a target of 8,000 humanoids in factories by 2028. A target like that gets built one bay at a time, and this report gives the first operating numbers from one of them.
Cost is the other yardstick. A Lanjing News piece that 21jingji republished the same day says a humanoid currently costs 600,000-plus yuan to put into a factory, and needs to fall to about 250,000 yuan before it undercuts a worker. The Zhaofeng report discloses neither the purchase price nor how payback is calculated. It says payback might eventually be three years. Kong’s framing is that the line is a proving ground, for validating scenarios and building experience, not something to judge on return on investment yet.
What the report leaves out
A 98% success rate is a good number with no denominator attached. The report does not say whether the other 2% is retried by the robot or handed to a person. It gives no human cycle time for the same station either, so a reader cannot tell whether 35 seconds is fast or slow. We won’t fill that gap.
One division of our own: 60,000 parts over about 1,000 hours is roughly 60 an hour. A 35-second cycle running nonstop would be about 100. The report does not say whether the 1,000 hours is fleet-wide or per robot, or how much of it was spent feeding. With the basis unclear, 98% should not be read as line efficiency.
What it means for jobs
Two job types sit directly in the path. One is repetitive pick-and-load work at the edge of a machine. The other is changeover. The report quotes front-line engineers saying a model switch takes three of them two to three hours to re-confirm positions and re-tune the equipment across several stations. Leju and Zhaofeng want the robots to absorb part of that later.
Zhaofeng’s order book explains why it is willing to try. Some overseas orders run small-batch, many-variety, with minimums of just 250 or 300 sets. That is the pattern that punishes fixed automation, which has to be rebuilt at every changeover, and it is where a re-programmable robot could pencil out. It is also the flip side of the Frost & Sullivan finding that production work stays a small slice of humanoid volume: the buyers who can calculate a return are the ones with frequent changeovers.
On timing, the cell is at the proving-ground stage, with about a year of validation ahead and one engineer to four robots. If that ratio holds at the end of the year and the 15-second target is met, feeding roles start to have a real ledger entry. Before then the work that moves first is the engineers’: who watches four robots, and who maintains 30 skill modules.
Kong said it plainly in the report: a factory needs its own technical and digital base before robots can work in it. Zhaofeng began automating around 2013. Plants without that base are further out.
Two things to watch. Whether Zhaofeng adds robots once the year of validation is done, and whether anyone publishes a measured cycle near 15 seconds within the six months Kong named.
