China’s Humanoid Robots Move From Performances to Real-World Work

The second World Humanoid Robot Games offered two very different images of humanoid robotics. Robots competed in sports and set striking performance records, but the event also tested cable connection, warehouse, restaurant, office, emergency, and other scenario-based tasks. At the same time, visits to Beijing’s embodied-AI innovation park revealed companies building components, collecting training data, raising capital, and pursuing customer deployments. 1213

Together, those signals point to an industry entering an early and uneven commercialization phase—not a finished market for general-purpose robot labor. Hardware is becoming easier to scale, but useful work still depends on reliability, standards, clearly defined applications, and enough high-quality data to handle variation.

The Games were becoming a work test

The Games opened at Beijing’s National Speed Skating Oval, known as the Ice Ribbon, on August 22, 2026. Organizers registered 666 teams and 2,056 robots, with the number of teams up 138% from the first edition and the robot count roughly quadrupling. The program expanded to 51 events, including 30 competitive events and 21 scenario-based events. 26

That format matters because athletic performance is only one narrow measure of capability. A robot can run quickly in a controlled competition without being able to perform a repetitive industrial task safely, repeatedly, and at a commercially viable cost. The work-oriented events therefore shift attention toward autonomy, dexterity, and task completion in environments designed to resemble factories, warehouses, restaurants, offices, and emergency settings. 1912

The autonomy rules also raised the bar. Reporting on the Games described a move away from simply finishing a task toward doing it quickly, accurately, steadily, and without remote control. 9

Scale is appearing across the hardware supply chain

The August visits to Beijing’s Zhongguancun Haidian embodied-AI innovation park suggested that the sector is no longer composed only of isolated prototypes. Companies described increasingly similar robot configurations and core components, a convergence that could reduce customization and make component production more scalable. 13

Zhuoyu Technology said the market had moved from prototype batches of roughly 10 to 50 units in 2024 and 2025 toward larger customer deployments. According to the company, between 10 and 20 customers had each ordered more than 1,000 units. It also reported monthly components sufficient for about 3,000 robots, with a target of 50,000 to 60,000 in November, alongside a backlog described as hundreds of thousands of robots. These figures are company or industry claims rather than independently audited shipment data. 13

The broader market is also attracting larger shipment forecasts. Morgan Stanley raised its estimate for China’s 2026 humanoid-robot shipments to 50,000 units from 28,000, citing commercial validation, policy support, and supply-chain momentum. 2728 Forecasts are not the same as delivered robots, but the revision shows how quickly expectations are changing.

Wujie Power shows how commercialization is being financed

Wujie Power, founded in 2025, illustrates the capital and product ambitions behind the sector. The company said it had two body generations in mass production and a third under development. It also reported more than $200 million in angel financing and nearly $100 million in global orders during the first half of 2026. Its cited use cases included industrial manufacturing and coffee-shop service, and it claimed European industrial full-stack CE certification. 13

Those announcements show that customers and investors are willing to fund pilots and early deployments before humanoids meet the reliability expected of mature industrial equipment. But they do not prove that the robots can already perform broad commercial work autonomously over long operating periods.

Zhang Yufeng, Wujie Power’s founder, drew the key distinction: a controlled industrial site can train a robot to perform a narrow, repetitive skill, while commercial service requires generalization across changing environments, users, objects, and failures. The second problem is substantially harder. 13

There is still no settled “killer application”

The industry’s demand is real, but its application map remains unsettled. Factories can offer structured layouts and repeatable tasks. Logistics adds movement, changing objects, and uptime requirements. Retail and hospitality introduce unpredictable people and interactions. Homes combine all of those difficulties with an almost unlimited range of objects and situations.

China’s June 2026 MIIT–SASAC initiative pushed companies toward real-scene training, validation, and deployment by the end of the year, increasing pressure to move beyond exhibitions and demonstrations. 13 Yet the policy push does not resolve the underlying product question: which tasks generate enough value to justify a humanoid’s cost, maintenance, supervision, and safety requirements?

The most credible near-term market is therefore likely to be constrained, repeatable work rather than fully general-purpose labor. The Games’ scenario events are useful precisely because they expose the gap between a scripted demonstration and a reliable workflow. 113

Reliability and standards are bigger barriers than spectacle suggests

Industry participants acknowledged that many humanoid robots still lack industrial-grade reliability and mature international standards. That is a major difference from conventional industrial robot arms, which are expected to operate for tens of thousands of hours. 13

A robot that completes an impressive event once has demonstrated a capability. It has not necessarily demonstrated safe, economical operation across multiple shifts, under changing loads, with predictable maintenance and failure recovery.

For buyers, the relevant questions are consequently changing:

• How often does the robot fail or require human intervention?
• Can it recover from an unexpected object, person, or obstruction?
• How many hours can it operate before maintenance?
• Can the same skill transfer between robot bodies and sites?
• Is there a recognized safety and compliance framework for deployment?

The answers will matter more to adoption than athletic records, even when those records attract the most attention.

Training data is becoming the central bottleneck

The park discussions also highlighted a shift from a hardware problem to a data problem. Reported customer requirements are rising from thousands of training hours to hundreds of thousands, and eventually to tens of millions. Noitom Robotics said it planned hundreds of thousands of hours of production in 2026 and millions in 2027. 13

The scale of that ambition can be seen in Noitom’s public release of HiPHI, a 617.5-hour high-precision dataset of human motion and object interaction. The dataset is substantial, but its size also illustrates the distance between a valuable research resource and the much larger volumes companies say will be needed for robust embodied-AI systems. 4749

The data is not interchangeable. The discussions described three broad layers:

1. High-quality operational data: precise first-person or teleoperated examples that can teach difficult skills, but are expensive to collect.
2. Retargetable human-centric data: human motion and interaction captured in a way that can transfer across different robot bodies.
3. Simulation data: abundant and relatively inexpensive, but limited by the gap between simulated and physical environments.

Physical capture was described as costing thousands of yuan per hour. That cost is driving efforts to collect data at much larger scale, including Noitom’s plan to crowdsource from nearly 20,000 people by the following year, beginning with home scenarios. 13

The underlying lesson is simple: a humanoid does not become general-purpose merely by having a human-like body. It needs enough varied, accurately labeled experience to connect perception, movement, objects, people, and consequences in the physical world.

What the August evidence actually shows

The Games and the innovation-park visits support a measured conclusion:

• The industry is scaling. The Games’ participation, expanding event roster, component convergence, funding, and reported orders all point to a larger commercial ecosystem. 213
• Commercial validation has begun. Companies are pursuing factory, logistics, hospitality, and other real-scene deployments, with policy adding pressure to test them outside exhibition settings. 113
• General-purpose work has not been solved. Reliability, standards, economics, task selection, and data remain open problems. 13

Humanoid robotics has therefore moved beyond a phase defined mainly by dancing, running, and scripted performances. The next benchmark is less cinematic but more consequential: whether a robot can perform useful work safely and consistently for thousands of hours, across enough real-world variation to justify its place in a business.