Robots Building Robots: Inside the 10,000-Unit Humanoid Factory That Just Started Production in China

Introduction

On 12 September 2026, a 14,000-square-metre building in Liuzhou, in the Guangxi Zhuang Autonomous Region of southern China, started doing something that until very recently existed mostly in promotional video. It started building humanoid robots, at a rate its owner says is one every ten minutes.

UBTECH Robotics, the Shenzhen-listed developer behind the Walker S and Cruzr lines, announced the facility as the world's first intelligent manufacturing plant designed around ten-thousand-unit annual capacity for industrial humanoids. Interesting Engineering reported the launch on 14 September, noting that UBTECH shares rose more than three percent in early trading that day. The Global Times covered it the same evening, quoting directly from the company's WeChat account, and humanoid.guide ran a piece on 17 September that was more interested in what the announcement did not say than what it did.

That second instinct is the right one. The Liuzhou plant is a genuinely important datapoint, and it is also a heavily curated one. Separating what has actually been demonstrated from what has merely been claimed is the whole job here, and our robotics and drones coverage has been tracking the gap between humanoid demonstration videos and humanoid production economics for months now.

What the Plant Actually Contains

The physical description is consistent across all three reports, which is a good sign in a field where announcements are frequently single-sourced. The building covers 14,000 square metres and stands 13.8 metres high. Output is nominally the Walker S series of embodied-intelligent industrial humanoids and the Cruzr series of logistics robots. The plant was built jointly by UBTECH and Siemens Digital Industries Software, and that partnership is the part of the story that is genuinely novel.

Siemens' contribution is not a robot. It is the simulation and scheduling layer. According to the accounts, the plant has a 1:1 digital model built in Siemens' Plant Simulation platform, reproducing aisles, workstations, storage areas and AGV routes, so engineers can rehearse material flow from warehouse to dispatch before a single fixture is moved on the real floor. On top of that sits a manufacturing operations management system called Yanshee, built on Siemens' Intelligence Center X agent platform, which UBTECH describes as the factory's "smart brain."

The functions that brain actually performs are unglamorous and specific: work-order scheduling, material distribution, and quality monitoring, closed-loop digitally end to end. Every robot leaving the line is assigned a unique serial number, and through it the component sources, assembly parameters and inspection records for that individual machine remain traceable. For a product whose failure modes are still poorly understood in the field, that traceability record may end up mattering more than the ten-minute cycle claim.

The on-floor automation is more conventional and, for that reason, more believable. Humanoid models from the Cruzr line — the Cruzr Y1 and Cruzr S2 specifically — are deployed for depalletizing, palletizing, loading and general material handling. Final assembly is handled by collaborative robot arms, power-assist manipulators, unmanned logistics vehicles and 360-degree rotating worktables, which together let multiple robot models share one line without a major hardware change between product generations.

The Two Claims That Deserve Scrutiny

The headline figure is the ten-minute interval. It is the number that will travel furthest, and it is the number least supported by evidence.

At one robot every ten minutes, a single line running twenty hours a day would produce 120 units. A 10,000-unit annual target implies roughly 83 hours of output per day — more than three shifts, seven days a week, with essentially no downtime, at a plant that had not previously been demonstrated at anything close to that rate. humanoid.guide noted this directly, pointing out that the company disclosure offers no operating schedule that reconciles the ten-minute interval with the stated annual capacity. Both figures can be quoted as design intent without either being a production forecast.

The second claim is the slogan itself: "robots building robots." It is quotable, memorable, and, on the disclosed detail, broader than the evidence supports. As humanoid.guide put it, the available account establishes that UBTECH's own humanoids handle materials on the floor; it does not establish that they perform the core mechanical or electrical assembly of other humanoids. Collaborative arms and rotating worktables do the assembly. The humanoids move the boxes.

That is not a debunking. A factory where the finished units handle logistics while conventional automation performs assembly is a real and useful configuration, and it is arguably the honest one, because it isolates which tasks humanoids are genuinely good at — unstructured material handling in a human-designed space — rather than claiming they are good at everything. But readers should understand that the robots in this factory are primarily logistics robots that happen to be built here, which is a different and less dramatic claim than the slogan.

The disclosure is also conspicuously silent on staffing. No headcount, no comparison of human and automated labour, no cycle-time data per station. Without those numbers, the claim that this plant reduces commissioning cost remains a company assertion rather than a measured result.

Quality Control and the Warehouse Problem

Two operational details stand out as more revealing than the capacity number.

Every completed robot reportedly undergoes more than four hours of whole-machine testing, followed by a pass through an automotive-grade lighting inspection tunnel for 360-degree, dead-angle-free appearance screening. Four hours is a long time relative to a ten-minute nominal build interval, which is a useful internal cross-check: the bottleneck is not assembly, it is validation. A facility that genuinely runs at scale will find that testing, not robot production, sets the pace. Companies that quote only the assembly interval tend to skip this.

The second is the storage problem, which is the genuinely hard constraint on humanoid scale. A human-sized machine is roughly 170 centimetres tall and does not stack. UBTECH says it independently developed what it claims is the industry's first fully automated stereoscopic warehouse for humanoid robots, occupying 65 square metres and holding 112 units — a claimed improvement in space utilisation of more than 50 percent. Behind that number is a logistics matrix of automated guided vehicles, unmanned forklifts and transport vehicles linking sub-assembly, final assembly, testing and storage.

Solve vertical storage and you have made robots cheap to hold. Fail to solve it, and every unit in production becomes a unit standing on a floor waiting for a customer. For a company whose customers are still running pilots, that is not an academic concern. It is the binding constraint on the entire business model.

The Supply Chain Argument

UBTECH's vice president Pang Jianxin framed mass production as inseparable from the Chinese supply chain, arguing that without it, mass-producing complex components would be extremely difficult unless cost advantages were sacrificed entirely. He pointed to the accumulation of engineers, scientists and application scenarios over the past ten to twenty years as the basis for both manufacturing capability and research strength.

National Bureau of Statistics data supports the broader claim about the industrial robot base, though not specifically about humanoids. China's industrial robot output reached 98,677 units in July 2026, up 30.2 percent year on year, with January through July production totalling 635,056 units, an increase of 28.5 percent. Those are conventional industrial robot numbers — six-axis arms, not humanoids — and it is worth resisting the temptation to read them as a proxy for humanoid readiness. They measure a mature, high-volume market that humanoids are not yet part of.

The competitive context is also worth stating plainly. As Interesting Engineering noted, the plant arrives as competition in humanoid robotics intensifies, with Tesla advancing its Optimus programme and Figure AI pursuing its own manufacturing plans. A 10,000-unit annual capacity figure is a meaningful near-term production base. It is not, by itself, evidence of demand. Until shipment numbers, customer order books and unit economics are published, capacity is the input to a business case rather than the output of one.

Conclusion

The Liuzhou plant is worth taking seriously, and for reasons more modest than the marketing suggests. A working 1:1 digital twin, a jointly built scheduling system, one-item-one-code traceability, four-hour whole-machine validation, and a purpose-built automated storage solution are all engineering achievements. They are the unglamorous layer on which a manufacturing scale-up either holds together or quietly fails, and the fact that UBTECH invested there rather than in a larger stage set is mildly encouraging.

What the plant does not yet demonstrate is that humanoids are cheap to make at volume. The ten-minute interval lacks a supporting schedule. "Robots building robots" describes material handling, not assembly. The staffing comparison is missing, and the price per unit is still unstated. Those are not disqualifying gaps — they are the normal gaps of a company that has built a factory ahead of a proven market — but they are the gaps to watch.

For anyone tracking this category, the numbers to watch next are straightforward: actual units off the line in the first quarter of operation, the named industrial customers, and the repeat order rate once the Rossmann-style pilots convert into production contracts. Until those exist, ten thousand units a year is a plan, and a plant is a plan that has been built.

Images

An industrial palletising cell: an orange articulated robot moves cartons past roller conveyors and a wooden pallet, with overhead gantry rails, cable carriers and safety fencing enclosing the guarded work area.

Illustrative: a conventional automated palletising and carton-handling cell of the kind that performs material movement on a production floor. This is not the Liuzhou plant.

Illustrative: an industrial pick-and-place installation handling components in totes and trays. Not the UBTECH line, and not the Liuzhou facility.

Illustrative: an earlier-generation UBTECH humanoid on display at a 2016 event — not a Walker S, not a Cruzr, and not the Liuzhou factory.