Is China Using Humanoid Robots in Factories? Here's What the Data Shows
Why read: a four-layer data check that separates real factory deployments from show-floor demonstrations, with the numbers to watch over the next 12 months.
By Embodied AI Frontier
Global humanoid robot production passed 20,000 units for the first time in 2025, roughly ten times the previous year. Interact Analysis, a UK market research firm, estimates that only about 10% went to real-world applications, with the rest split between research, data collection and entertainment.
Chinese manufacturers produced more than 90% of that output. Early demand came mainly from research and education: Unitree's IPO filing shows that 73.6% of its humanoid robot revenue in the first three quarters of 2025 came from research and education customers, while industrial applications accounted for 9.01%.
Deployments on factory floors can be checked case by case. At Longcheer Technology's Nanchang plant, eight AGIBOT G2 humanoids worked 64 cumulative hours over a six-day livestream, completing 64,828 inspection tasks while the line produced 17,625 units, with a 99.99% success rate.
Working efficiency is still climbing. UBTECH reports that its robots currently operate at about 45% of human efficiency, with corporate targets of 60% in 2026 and 80% in 2027. A nut-tightening station test at Xiaomi's auto plant ran autonomously for 3 hours at a 90.2% success rate, against an industry norm above 99.9% for that kind of station.
This article breaks down four layers of numbers (production, shipments, deliveries, deployment), checks individual factory cases against one standard, and lists the figures to watch over the next 6 to 12 months.
So are Chinese humanoid robots actually working in factories?
Article structure
- 1. First, the books: how many were built, sold, and what they do
- 2. On the factory floor: progress that can be checked
- 3. Two more "workshops": training grounds and stages
- 4. Questions still without answers
- 5. Numbers that will talk
Opening: Berlin and Liuzhou
In early September 2026, on the show floor of IFA in Berlin, EngineAI, a humanoid robotics company headquartered in Shenzhen, was demonstrating its flagship model T800. In the middle of the booth, the robot threw a set of boxing moves in front of a crowd, drew back, and stood still. Phone screens lit up across the audience.
At almost the same moment in Liuzhou, a city in southern China's Guangxi region, a newcomer at a robot data-training center was repeating the same grasping motion. Reuters recorded the details during a visit: a novice needs roughly 300 attempts to produce one piece of motion data that a robot can directly use; an experienced trainer cuts that to about 50.
The robot on stage was performing. The people in the training center were teaching a robot how to work. Put the two scenes side by side and one question is hard to avoid: which of them represents China's robotics industry today?
Move that question onto a production line and it becomes more specific: are Chinese humanoid robots actually working in factories?
Answering it requires dealing with a technical problem first. In this industry, "how many were built," "how many shipped," "who received them" and "how many are on the job" are four different sets of numbers, and financial headlines often mix them together. This article takes them apart layer by layer (Chapter 1); then checks individual factory deployments, counting only records with a named customer, unit counts, and operating data (Chapter 2); then turns the camera to training grounds and stages, two "workshops" that are running just as busy (Chapter 3); then lists the questions public data still cannot answer (Chapter 4); and finally, lists the numbers that will verify these threads over the next 6 to 12 months (Chapter 5).
1. First, the books: how many were built, sold, and what they do
When a factory considers buying a new piece of equipment, its procurement department asks for three things first: acceptance records, the number of units on the floor, and operating data. Judged by the same standard, the humanoid robot industry's first step is to layer its accounts.
Layer one: production. Interact Analysis counts global humanoid robot production passing 20,000 units in 2025 for the first time, up from fewer than 2,000 in 2024, roughly a tenfold increase in one year43Humanoid robot production surges tenfold in 2025, but commercial deployments remain limitedView the entry below →. China's Ministry of Industry and Information Technology (MIIT) reported in July 2026, through Vice Minister Ke Jixin, that China produced about 20,000 units in 2025, more than 40,000 in the first half of 2026, and is expected to exceed 100,000 for the year. Production is the largest of the four layers. It measures what assembly lines output, with no distinction for whether the units have been sold.
Layer two: shipments. For 2025 alone, at least three organizations have published figures: Omdia, a UK market research firm, counts about 13,000 units (including wheeled models)44Omdia ranks AGIBOT No.1 worldwide in humanoid robot shipments in 2025View the entry below →; a report by China Electronics News and CCID counts about 17,000 (excluding wheeled models)452025 Humanoid Robot Market Research ReportView the entry below →; and IDC, a US market research firm, counts about 18,00048Humanoid Robotics CommercializationView the entry below →. The three numbers differ by 20% to 40%, a gap that comes from how each defines its scope. Unitree, a Hangzhou-based robotics company, addressed this in its IPO filing in March 202631UnitreeView the entry below →: different research firms can disagree by thousands of units on the same company's shipments, and the industry currently lacks a relatively unified, comparable data source.
Fig 1 Global shipments in 2025 under three counting methods (Omdia / China Electronics News and CCID / IDC)
In the first half of 2026, SAG, a market analytics firm, counted 19,100 units shipped globally, up 272% year over year47Global Humanoid Robot Shipments 2026View the entry below →; Counterpoint Research put the figure above 22,000, an increase of nearly 300%46Global Humanoid Robot Shipments Soar Nearly 300% YoY in H1 2026View the entry below →. The Chinese Institute of Electronics and the China Humanoid Robotics and Embodied Intelligence 100-People Committee reported a China-specific figure: more than 40,000 units shipped from China in the first half, about 97% of the global total492026 Humanoid Robot Industry Development ReportView the entry below →. The two sets of numbers point the same way: Chinese manufacturers account for the overwhelming majority, while the disagreement concentrates on how the global total is defined.
Layer three: deliveries and destinations. This step answers "who bought them." Morgan Stanley analysts, cited by Reuters Breakingviews, reported that China sold about 12,000 humanoid robots in 2025, most going to research, education and testing8China's robot quest triggers system overloadView the entry below →. Government procurement is another publicly checkable channel: Reuters counted nearly 1,000 records of humanoid robot purchases by Chinese national, provincial and local governments in the first half of 2026, worth at least $230 million32China can build kung fu-fighting robots. But it can't get them to do factory workView the entry below →.
Company-level revenue structure points the same direction. Unitree's filing shows that in the first three quarters of 2025, 73.6% of its humanoid robot revenue came from research and education, 17.39% from commercial consumption, and 9.01% from industry applications. Within industry applications, guided-tour scenarios dominated, accounting for 50% to 70%31UnitreeView the entry below →.
Layer four: real applications and robots on the job. This is the strictest layer, and the source of the "10%" figure. Interact Analysis divides each year's production into four destinations: real applications, academic research, robot training and data collection, and entertainment43Humanoid robot production surges tenfold in 2025, but commercial deployments remain limitedView the entry below →. Of the more than 20,000 units produced in 2025, about 10% (roughly 2,000 units) entered real applications, covering industrial manufacturing, warehousing and logistics, and public services.
The remaining 90% went to research, data collection and entertainment, serving different customers and constituting real demand of their own. What the 10% raises is a specific question: which workshops did those roughly 2,000 units go to, what work did they do, and what operating records did they leave?
Fig 2 The four-layer ledger: from production to real applications (Source: Interact Analysis)
A different cut makes the distribution of early demand clearer. Counterpoint Research breaks down global shipments in the first half of 2026 as follows: entertainment performances 33.6%, data production and research 27.0%, service and guidance 19.0%, intelligent manufacturing 12.8%, and warehousing and logistics 4.9% (a sub-category breakdown relayed by industry media)46Global Humanoid Robot Shipments Soar Nearly 300% YoY in H1 2026View the entry below →. Entertainment and data together still account for over 60%.
Put this next to the revenue structure above and the outline of an early market emerges. The most active buyers, in order, are government procurement programs, universities and research institutes, data collection centers, and entertainment venues. Factories, in this picture, are still one of the smallest squares.
One thing the books confirm: at its strictest definition, "working in factories" is still small, small enough that there is no authoritative total, only case-by-case verification. The next chapter does exactly that.
2. On the factory floor: progress that can be checked
To check factory deployments, this article applies one uniform standard: a named customer, a unit count, acceptance or continuous operating hours, and operating data. Where only the company itself has reported a figure, it is marked as such. Where an announcement was made and nothing followed, that is recorded too. That is part of the research.
The most complete sample sits in Nanchang.
Longcheer Technology x AGIBOT. Longcheer Technology (603341.SH) is a Shanghai-based ODM manufacturer of consumer electronics. In the tablet quality-inspection section of its Nanchang plant, a fleet of G2 robots from AGIBOT (a Shanghai embodied intelligence company) is on the job. The sequence of this partnership runs as follows:
Table 1 Longcheer Technology's G2 deployment timeline (Oct 2025 to Jun 2026)
| Date | Milestone |
|---|---|
| Oct 2025 | Framework order worth hundreds of millions of yuan announced, with nearly 1,000 units planned (a framework target, not units deployed) |
| Dec 2025 | Robots arrived at the Nanchang plant for handling tests and line integration |
| Apr 2026 | Two units moved into formal posts; after a cumulative 140 hours of continuous operation, "all preset technical targets were fully met" |
| May 2026 | Eight units joined the line, covering the entire tablet quality-inspection section |
| Jun 2026 | Six-day livestream of the eight-unit fleet: 64 cumulative hours of work, 64,828 inspection tasks, 17,625 units produced on the line, 99.99% success rate; the 15,000th G2 was delivered to Longcheer the day it came off the line |
| Q3 2026 | Planned expansion to 100 units (planned basis) |
Three numbers need to be kept apart. The 17,625 units are what the production line produced in six days; the 64,828 is the number of inspection tasks the robots completed; and 99.99% is the task success rate. The three measure different things, and mixing them up produces a false impression. As for the livestream in late June, Yao Maoqing, a partner at AGIBOT, told media that what appeared on camera was the routine daily work these G2 units had performed since March.
The efficiency data is equally specific. Each G2 in this fleet has a cycle time (CT) of 9 to 11 seconds, against the 11.25-second takt time (TT) the line sets for the inspection station. A single unit handles about 310 pieces per hour and covers two process steps. For comparison, the quality inspectors who previously staffed this section worked at 9 seconds per piece, while minding seven inspection machines at once. After the eight G2 units came online, they fully took over the workload of eight inspectors, who were reassigned. Line records show the robots' continuous stability exceeded that of the human crew, and effective output rose.
A reporter from Yicai, a Chinese financial media outlet, recorded another scene at the plant: the robots move smoothly, but a single motion is still slower than a human. Zhang Long, the project lead at Longcheer, described them as "about the level of a seven- or eight-year-old child." On the cost side, Sina Finance reported that the equipment investment roughly matched two years of the workers' labor cost13Sina FinanceView the entry below →. The learning curve has also been shrinking: the first workstation took about four months to deploy, the second was cut to one week, and later workstations took one to two weeks (per AGIBOT).
One clarification belongs here: "nearly 1,000 units" was the phased target written into a framework contract signed in October 2025. The publicly verifiable figure for units actually on the job is eight. The company's stated next step is to expand to 100 units in the third quarter of 2026.
Beyond Longcheer, the same product line is spreading. AGIBOT announced a partnership covering nearly 100 units with Fulin Precision, an auto parts maker, in August 2025, placing robots at multiple key points on its lines. In March 2026, at a SAIC-GM Buick battery production line, a unit named "Nengzai No. 1" (A2-W) took up high-precision work. Joyson Electronics completed initial deliveries to its automotive safety plants under a procurement contract worth over RMB 100 million, with the robots entering commissioning; after scaled training, their task success rate held above 99.9%. From one workstation to multiple customers, this is the first replication signal in the ledger.