Key takeaways
- China's Ministry of Industry and Information Technology (MIIT) expects full-year 2026 humanoid robot (人形机器人) output to exceed 100,000 units, roughly five times 2025 volume.
- AgiBot (Zhiyuan) and Unitree lead production, but industry data shows a large share of shipped units still goes to research labs and schools, not assembly lines.
- UBTECH's consumer-facing U1 drew 13,361 orders at 169,800 yuan (≈ US$23,900 / HK$187,000) within weeks of its June 30 launch.
- The real constraint has shifted from hardware to software — embodied AI (具身智能) that can handle unstructured, unpredictable tasks.
- Production capacity is no longer the bottleneck; proven, repeatable factory deployment is.
The number that reset the conversation
On July 7, 2026, at the World Artificial Intelligence Conference (WAIC) press briefing in Shanghai, Gan Xiaobin, deputy director of MIIT's Science and Technology Department, said China's full-year output of complete humanoid robot units is expected to surpass 100,000 in 2026. State news agency Xinhua carried the figure, and industry estimates put 2025 output at around 20,000 units — a fivefold jump in a single year.
For a category that, eighteen months earlier, was shipping hand-built prototypes, the statement marked a threshold: humanoid robots (人形机器人) had moved from showcase curiosity to an industrial product with a national production target attached to it.
The projection did not arrive alone. MIIT said it is running special actions such as "model-data resonance" (模数共振) and "humanoid robots and embodied-intelligence real-scene training" (人形机器人与具身智能实景实训) to surface high-value use cases, and noted that industry players have already drafted nearly 200 key standards for the category. The standards work matters because, without shared interfaces and safety benchmarks, a 100,000-unit market risks fragmenting into incompatible, hard-to-service machines.
Who is actually building them
The surge is not evenly distributed. According to MIR DATABANK data reported by Pandaily and 36Kr, AgiBot (Zhiyuan) announced its 15,000th general-purpose embodied robot (具身智能机器人) rolled off the line in late June 2026. Unitree, meanwhile, had produced roughly 11,000 units by March and plans to lift annual capacity toward 30,000. UBTECH, the listed pioneer, targeted a combined 10,000-unit goal across its Walker and industrial lines for 2026, and on June 30 launched the U1, a full-size consumer humanoid, with 13,361 cumulative orders at 169,800 yuan (≈ US$23,900 / HK$187,000).
XPeng's robotics arm and several automakers are also entering through their existing supply chains, treating the humanoid (人形机器人) as a natural extension of car manufacturing rather than a separate business.
The uncomfortable gap: produced ≠ deployed
Here is the part the headline number hides. Through the first three quarters of 2025, more than 70% of Unitree's humanoid revenue came from research and education customers, with industrial applications at just 9%. In plain terms, a large share of the robots being "produced" were not entering factories — they were landing in university labs, training centers, and demonstration showrooms.
UBTECH's own shipment ranking slipped even as its production rose, a pattern analysts read as inventory buildup or slow delivery cadence. The consumer U1's 13,361 orders, while real, are not counted in industrial production statistics because they are a to-consumer product. The gap between building a robot and putting it to work is the industry's central tension in 2026.
Why factories are still waiting
Hardware supply chains have matured fast. Chinese suppliers of harmonic drives, servo motors, and optical sensors now undercut Japanese and German equivalents by 30–50%, which is why a capable humanoid can be priced in the tens of thousands of dollars rather than hundreds of thousands.
The bottleneck is software. Today's humanoid robots handle pre-programmed tasks reliably but struggle with novel situations, unstructured environments, and fine motor judgment. Embodied AI (具身智能) — built on vision-language-action (视觉—语言—动作) models that let a machine perceive, reason, and act in the physical world — is advancing, but the training datasets for most industrial tasks do not yet exist at scale. Until that changes, a robot that can dance at a launch event is not the same robot that can reliably load a production line eight hours a day.
What it means for readers outside China
China's humanoid push matters globally for two reasons. First, its component cost curve is setting the world price for the category; cheaper actuators and sensors in China ripple into every robotics lab on the planet. Second, the deployment gap is a live experiment the entire industry is watching: whoever solves "working reliably in a real factory" first will define the next decade of automation, and right now no one has fully solved it.
Honest limitations
This article relies on MIIT's projection, company disclosures, and MIR DATABANK figures reported by Pandaily and 36Kr; we did not independently audit shipment or capacity numbers. "Output" and "deployments" are reported by different sources and should not be read as the same thing. The 100,000-unit target is a projection, not a confirmed result, and final 2026 figures may differ. We excluded unverified export and overseas-deployment claims, and we did not include specific per-model prices beyond officially disclosed figures.
What readers can do now
- Track the gap, not just the headline: when reading humanoid (人形机器人) news, separate "units produced" from "units working in factories."
- If you run operations, pilot one narrow, repeatable task — sorting, inspection, material transport — before betting on general-purpose labor.
- Watch component suppliers (harmonic drives, servos, vision sensors): their cost curve predicts robot affordability better than launch events do.