Inside a clean, brightly lit plant in Nanhai, Foshan, a humanoid robot reaches the end of the line roughly every half hour. Not as a demo. As a cadence. The line was switched on in March 2026 and is designed to turn out more than 10,000 humanoids a year — the first automated production line in China built for that scale.
The plant is a joint bet by Leju Robotics (乐聚机器人), a Shenzhen humanoid maker, and Guangdong Dongfang Precision (东方精工), a packaging and intelligent-equipment manufacturer with nearly three decades of precision-manufacturing experience.
Key takeaways
- The Foshan line, enabled in March 2026, has an annual capacity above 10,000 humanoids and a takt time of one unit every 30 minutes.
- It integrates 24 assembly steps and 77 full-dimension inspection points, with efficiency reported about 50 percent higher than manual assembly.
- The joint venture, Orient Yuanqi Intelligent Robotics (东方元启), was formed in September 2025 with Dongfang Precision holding 80 percent.
- A companion Foshan data-training center targets more than 1 million high-quality training records in its first phase (2026–2027) across 100-plus real industrial scenarios.
- A 10 billion yuan (≈ US$1.4bn / HK$11bn) AI industry fund backs the local embodied-intelligence ecosystem.
From one prototype to a consistent thousand
Humanoid robotics spent years as a showcase sport: a machine walks, waves, maybe does a backflip, and the video circulates. What Foshan changes is the boring part nobody films — the part where a product has to come off the line identical, thousands of times, with tolerances that hold. A humanoid has hundreds of parts; a loose screw or a mis-tuned joint amplifies through the whole body and shows up as a wobble or a failed grip. That is why the industry's real wall was never the demo. It was consistency.
The line attacks that wall with structure. An AI-driven central controller schedules materials,节拍 (takt) and process steps in real time. Unmanned warehousing and AGV (automated guided vehicle) delivery feed parts just in time. Final assembly uses a flexible, non-fixed line that mixes models. Before a unit ships, it passes 77 inspection points and 41 automated condition tests covering noise, communication and simulated motion. The operators say the line already meets automotive-grade (车规级) delivery standards — the language car suppliers use for parts that cannot fail.
Why a packaging-equipment maker matters here
Dongfang Precision is not a robotics name. It builds carton and corrugated equipment. That is precisely the point. Leju supplies the full stack — body hardware, motion-control algorithms, operating system — while Dongfang contributes the manufacturing discipline: line design, supplier integration, quality systems. The logic mirrors how China's EV boom was won — not by a single genius lab, but by folding a new product into an existing, world-class supply chain.
A telling detail: the same campus hosts the Greater Bay Area (Foshan) Embodied-Intelligence Robotics Data-Training Center, built with the local hi-tech zone and a city investment arm. It recreates real factory environments — home-appliance and equipment manufacturing among them — so robots can collect interaction data and train models. The first phase (2026–2027) aims for more than 1 million high-quality training records across 100-plus scenarios. Data, not just hardware, is treated as the bottleneck.
The commercial question
Scale does not equal demand. A 10,000-unit line only matters if orders follow, and the honest open question for the whole humanoid sector in 2026 is whether deployments convert from pilots to paid, repetitive work. Early adopters are in 3C electronics, automotive and logistics — places with structured tasks and a clear labor gap. Cost per unit is expected to keep falling as volume rises, which in turn widens the set of tasks where a robot beats a wage.
According to Li An, chief scientist at BrainNet (脑机网), the real bottleneck for humanoids is no longer the hardware prototype but consistent, validated mass production — and Chinese contract-manufacturing depth is what most foreign observers still underestimate. That is a judgement, not a figure, and it is testable over the next few quarters as more lines like Foshan's report actual shipment and return-on-investment numbers.
Honest limitations
- The capacity, takt time, step counts and efficiency figures come from company disclosures and Chinese business-media reports (China Securities Journal, China Daily, People's Daily); we did not independently audit the line.
- "Automotive-grade" is the operator's self-description of quality control, not a formal certification we can confirm.
- The 10 billion yuan fund and the 1-million-record data target are planned figures, not realized outcomes.
- Shipment and utilization data for the humanoids produced here are not yet public, so payback claims remain forward-looking.
What readers can do now
- If you buy or deploy humanoids, benchmark vendors on consistency metrics — defect rate per 1,000 units and calibration drift — not on stage demos.
- Track whether Foshan-style lines publish real shipment and on-site uptime numbers in the next two quarters; that is the signal that matters more than capacity headlines.
- For supply-chain watchers, follow the component layer (joints, force sensors, dexterous hands) — that is where margin and bottleneck risk now concentrate.