On a July morning in Shanghai, an official from China's Ministry of Industry and Information Technology (工业和信息化部) stood at a podium and said something the industry had been whispering for months: this year, the country's humanoid robot (人形机器人) output is expected to pass 100,000 units.
The number landed like a starting gun. But the quieter signal was in the fine print — most of those machines are not yet doing real shifts on real factory floors.
The headline: 100,000, and a fivefold jump
Gan Xiaobin (甘小斌), deputy director of the MIIT science and technology department, told the 2026 World Artificial Intelligence Conference (WAIC) press conference that full-year humanoid robot production is projected to exceed 100,000 units — roughly a fivefold leap from about 20,000 in 2025. He also noted that AI application penetration among large industrial firms has passed 30%.
It is a striking milestone for an industry that, two years ago, was shipping showcase units by the hundreds. The question is what "production" actually measures.
The duopoly: AgiBot and Unitree
Behind the national figure are two names that now dominate Chinese humanoid manufacturing.
- AgiBot (智元机器人) reported 15,000 units off the production line in early June.
- Unitree (宇树科技) had completed roughly 11,000 units by March, according to industry data cited by Pandaily from MIR DATABANK.
Both companies describe mass production as the foundation for moving "out of the lab and into industry," with deployment said to span production-line loading, industrial handling, logistics sorting, guide services, retail, security patrol, and commercial cleaning.
The UBTech twist: a consumer product that doesn't count
UBTech (优必选) launched the U1 on June 30 — described as the world's first mass-producible full-size hyper-realistic humanoid. It gathered more than 13,361 cumulative orders, with a U1 Pro price of ¥169,800 (≈ US$23,600 at roughly 7.2 RMB/USD).
Here is the catch: because the U1 is a consumer (ToC) product, those orders are not counted in the industry's production statistics. The line between "humanoids built" and "humanoids deployed at work" is blurrier than the 100,000 figure suggests.
The uncomfortable truth: most robots aren't in factories
Unitree's own revenue mix tells the story. Through the third quarter of 2025, more than 70% of its humanoid revenue came from research and education customers, with industrial applications at only about 9%. In other words, most shipped humanoids went to labs, universities, and classrooms — not assembly lines.
Unitree's 2026 shipments are projected to at least double 2025's roughly 6,500 units, with annual capacity targeted at 30,000. But capacity is a manufacturing commitment, not a deployment record.
Why the gap exists
The industry has solved the hard problem of building bodies cheaply. It has not yet solved the harder problem of making those bodies useful enough that a factory manager reorders.
- Real industrial deployment requires reliability across messy, variable tasks — not choreographed demos.
- Return on investment is unproven: a humanoid has to beat a human or a fixed automation cell on cost and uptime.
- High production with slipping shipment rank (as seen with UBTech) often signals inventory buildup rather than demand.
As one industry observer noted, the shift in questions this year moved from "how cool can it walk?" to "can it work 24 hours straight, understand a vague instruction, and scale?"
What actually counts as deployment
The industry uses "deployment" loosely. A robot performing a repetitive, bounded task inside a single company — loading parts, patrolling a floor, sorting parcels — is genuine deployment. A unit unboxed in a university lab for a term paper is not, even if it appears in the same shipment tally.
The most credible early wins are in logistics and automotive. Car factories offer structured, repeatable environments where a humanoid can learn one task and repeat it thousands of times. That is exactly the kind of setting where a 70%-research-and-education revenue mix starts to shift toward industry.
Why Beijing is pushing the number
Humanoids are not just a product category; they are a strategic bet on embodied AI (具身智能) and domestic robotics supply chains. MIIT has paired the output projection with special actions like "humanoid robot real-world training" to surface high-value scenarios. The goal is to turn China's manufacturing base into both the builder and the first customer of these machines.
That policy tailwind explains why capacity is being built ahead of proven demand. The 100,000 figure is as much a signal of intent as a shipment forecast.
Honest limitations
- The 100,000-unit projection is an official forecast, not audited output; final numbers depend on demand through the fourth quarter.
- Company-level figures (AgiBot 15,000, Unitree 11,000, UBTech 13,361 orders) come from corporate statements and a single industry-data cite (MIR DATABANK via Pandaily), not from a unified government tally.
- "Production" and "deployment" are reported interchangeably in much coverage; the evidence shows most humanoids are still in research, education, or showcase roles.
- Per-unit economics and true factory ROI remain largely unpublished.
What readers can do now
- When you see a humanoid "shipment" number, ask who bought it — a factory, a university, or a showroom — before reading it as adoption.
- Watch for repeat orders and named industrial customers; one-off deployments prove capability, recurring purchases prove a business.
- Track MIIT and exchange filings (several humanoid makers are pursuing listings) for audited capacity and revenue splits rather than press-counted units.