Fifteen humanoid robots in Hangzhou have crossed the line every robotics company talks about and few reach. They have stopped being a pilot. On October 15, Leju Robot and Zhaofeng Electromechanical launched their embodied intelligence plus auto-parts feeding solution, and 15 Kuafu humanoids entered normal operation on Zhaofeng’s parts line – regular shift work, on a live production line, feeding real parts into real manufacturing.
Gasgoo Embodied Intelligence reported the launch on October 16, and the operating numbers are what make it news. This is not a robot waving at a trade fair. It is a fleet with a success rate, a cycle time and a production count attached to its name.
The job, in numbers
The 15 Kuafu units work in a narrow, constantly changing line environment. Each performs more than ten actions in sequence to pick small parts with inner diameters of 80 to 120 millimetres, linking up fetching, moving and feeding across 15 stations and more than ten part types.
Held steady through operation, the feeding success rate is 98 per cent, with each cycle completed in under 40 seconds. The fleet has logged about 1,000 hours and helped produce more than 60,000 parts. Sit with those figures for a moment. A 98 per cent success rate across 60,000 parts is a manufacturing statistic, not a demonstration anecdote. It tells a plant manager exactly what a demo video cannot: how often a human still has to step in, and what the line costs when they do.
That distinction runs through this whole autumn in robotics. Trade floors are full of firsts – this week alone brought Mobo, a family robot preparing its global debut and pre-orders, aimed at the home. Hangzhou is the opposite kind of headline: no unveiling theatrics, just a fleet that clocked in and stayed.
The system underneath the robots
Leju describes a reusable stack beneath the fleet, spanning development, deployment, use, data and safety. Modular skill tools cut development work from months to weeks. A multi-robot dispatch platform syncs skills across the fleet with one click and shortens deployment to hours. Operators get a one-click call for parts, job data flows back in a full loop, and safety modules watch for dropped parts, collisions and overspeed.
Then there is the data flywheel, which may be the most valuable output of all. Fifteen robots running in parallel collect more than 40,000 real job records a month. Those records pass through quality checks and go back into training, so the models sharpen with use. Every shift makes the next shift slightly better, and every new site the system enters starts from everything the Hangzhou fleet has already learned. That is how a single deployment becomes a product that can be copied.
The companies are sharing the homework, too. Alongside the launch, an embodied intelligence plus auto-industry live training base was unveiled, and Zhaofeng donated its real-world parts-feeding dataset to the OpenAtom OpenLET community. Giving away hard-won production data is a confident move. It lowers the barrier for the whole sector, speeds up everyone else’s training, and positions Leju and Zhaofeng as the reference implementation others build on.
Why car parts are the perfect test
Auto-parts feeding sounds modest. That is exactly why it is a serious test. The parts are small, the tolerances are tight, the line does not stop for a robot that fumbles, and the environment changes constantly as stations, batches and part types rotate. A humanoid form factor earns its place here because the line was built for human hands and human reach. Retooling an entire plant around a fixed robot arm is expensive. A machine that walks up to the existing line and feeds it is a different economic proposition.
Compare the shape of this deployment with the service robots arriving in public spaces, such as the outdoor parking robot STAN debuting in Gangneung. Public-facing robots have to win over strangers. Factory robots have to satisfy one audience only: the production ledger. On the evidence from Hangzhou – 98 per cent, under 40 seconds, 60,000 parts – the Kuafu fleet is passing that audit.
What it means for factories watching
For plant operators, the takeaway is practical. Ask vendors for fleet numbers, not unit demos: success rate, cycle time, hours logged, parts produced. Leju says the solution can now copy across stations and scenarios, and that it will keep open-sourcing real-robot data and toolchains with partners. If that replication holds at a second and third site, the Hangzhou line will be remembered as the moment humanoids in car manufacturing moved from trial validation to scaled deployment – the companies’ own phrase, and for once, one the data supports.
Fifteen robots will not transform an industry by themselves. But industries rarely transform in leaps. They transform when a working shift becomes a template, and the template starts travelling. Hangzhou has produced the template. The next deployment announcements will show how fast it travels.
Be First to Comment