China Leads Humanoid Robot Shipments—but Reliable Robot Labor Is Still Unproven

China’s position in humanoid robotics looked formidable at the 2026 World Robot Conference in Beijing—but the event did not prove that Chinese robots can already replace people in complex, unpredictable work. It showed a country with exceptional manufacturing scale, a large domestic market and a growing pipeline of commercial demonstrations, alongside unresolved problems in reliability, adaptability, safety and cost-effectiveness.

The conference’s clearest signal: manufacturing scale

The five-day event ran from August 19 to 23 and brought together more than 300 exhibitors, over 2,000 exhibits and more than 150 product launches, according to organizers and reporting from the conference. 1 3 The exhibition’s emphasis was not only on acrobatics or staged spectacle. Companies presented robots sorting parcels, packing mobile phones and assisting with household chores—tasks intended to show a shift toward factories, logistics, retail and domestic use. 3

An industry report released during the conference said China shipped more than 40,000 humanoid robots in the first half of 2026, representing 97% of global shipments. 15 16 That figure should not be treated as an uncontested market total: another estimate put worldwide H1 shipments at 19,100 while also assigning Chinese vendors 97% of the market. 4 The disagreement does not erase China’s reported lead, but it does make the exact shipment count uncertain.

China’s advantage appears to extend beyond individual products. Reporting points to a broad domestic supplier base, lower-cost hardware and a home market that absorbed most humanoid-robot demand. 4 Those conditions give Chinese manufacturers more opportunities to test, sell and iterate at scale than companies operating in smaller or more expensive markets.

From viral demonstrations to useful work

The demonstrations at the conference illustrated meaningful progress in several areas: perception, grasping, balance, mobility and the ability to sequence multiple actions in controlled environments. Parcel sorting and mobile-phone packing are especially revealing because they connect humanoid form factors with industrial workflows rather than purely theatrical stunts. 3

But a successful demonstration answers only a narrow question: can the robot complete this task, in this setting, under these conditions? A useful workplace robot must also identify unfamiliar objects, cope with variation, recover from mistakes and repeat the task without constant human intervention. Short demonstrations do not establish that level of generality. 1

The supplied reporting does not independently verify detailed claims about every company display referenced in the conference discussion, including specific demonstrations attributed to Zoomlion, Pudu Robotics or Galbot. Galbot’s ET1 was reported as a new bipedal model being trained to play tennis and scheduled for an appearance at the World Humanoid Robot Games, but that athletic plan should not be confused with proof of general-purpose workplace capability.

Why humans still outperform humanoids in many jobs

The industry’s own executives offered the most important qualification to the conference’s optimistic displays. Unitree founder Wang Xingxing said humanoid robots remain less efficient than people and require substantial time to learn new skills. 8 That matters because commercial automation is judged less by whether a robot can perform a task once than by whether it can do so faster, more cheaply and more consistently than an employee.

Keenon COO Wan Bin drew a similar line between a convincing demo and a dependable product. Completing a task 50% or even 80% of the time is comparatively achievable, he said, while reaching 99.9% completion demands much stronger engineering and training. 8

That reliability threshold changes the economics of deployment. A robot that fails unpredictably may require supervisors, remote operators, redesigned workspaces or frequent maintenance. Those costs can eliminate the benefit of automating the task in the first place. The central commercial challenge is therefore not simply making humanoids walk, run or manipulate objects; it is making them robust enough to work safely and repeatedly in environments their designers cannot fully script.

The sprinting crash captured the remaining gap

A humanoid robot training for a running event lost control, struck a padded barrier and buckled sharply at the waist, with sparks reportedly visible after the impact. The robot’s manufacturer was not confirmed.

The incident was more than an amusing viral failure. It highlighted the difference between achieving dynamic movement and controlling that movement safely: stopping, staying on course, managing momentum and recovering from an error are all essential in the real world. A robot that can sprint but cannot reliably stop is not ready for an uncontrolled workplace.

Athletic demonstrations can still be useful engineering tests. They expose weaknesses in balance, actuators, control software and structural design under demanding conditions. But they should be read as evidence of progress toward robustness—not as evidence that a humanoid has mastered ordinary human work.

What the U.S. restrictions could change

The geopolitical backdrop added another layer to the conference. In July, the U.S. Federal Communications Commission added foreign-made advanced robotic devices, including humanoids, to a list restricting new equipment authorizations. Models previously authorized by the FCC can still be imported and sold, according to reporting cited in the conference coverage. 6

The immediate effect on U.S. humanoid adoption may be limited because relatively few humanoids are currently deployed in the country. The restrictions could matter more for autonomous mobile robots already used in factories and other commercial settings. 6

The longer-term issue is access to experimentation. If inexpensive Chinese platforms become harder to import, American startups, companies and researchers may face higher hardware costs or fewer opportunities to test algorithms on commercially available machines. That is a potential consequence inferred from the restrictions and China’s reported supply advantage, not a measured outcome. 4 6

The real verdict from Beijing

The World Robot Conference showed that China leads the current humanoid-robot race in reported shipment volume, supplier breadth and the scale of public demonstrations. It also showed why shipment counts alone are not enough.

The decisive test will be whether these machines can learn new tasks quickly, operate safely around people, recover from surprises and deliver near-continuous performance at a cost businesses can justify. On that measure, China has demonstrated momentum rather than a finished victory. The gap between a robot that can perform an impressive routine and one that can reliably do useful human work remains the industry’s defining problem. 1 3