China scales up embodied AI's reach

The industry's next milestone is no longer whether a humanoid can walk. It is whether thousands of robots can repeatedly execute mundane jobs — loading, sorting, inspecting and transporting — cheaply enough for factory managers to keep ordering them. That shift was also visible at the World Robot Conference.

Previous robotics shows were dominated by spectacular demonstrations: humanoids doing somersaults, kung fu or synchronized dances. This year's booths increasingly resembled miniature factories, pharmacies and stores, with robots sorting goods, folding clothes, cleaning floors and manipulating industrial components.

The question has changed from "Can it move?" to "Can it work?"

Wang Tianmiao, honorary director of the Robotics Institute at Beihang University, said: "Commercialization will require much more than improving a robot's AI brain. Reliability, manufacturing consistency, supply-chain standards, economics and the choice of suitable applications must advance together."

The World Humanoid Robot Games put those limitations under even harsher lighting. A total of 666 teams registered 2,056 robots for 51 events. Alongside running and martial arts were 21 scenario-based competitions designed to test capabilities closer to those required in real workplaces.

Robots fell. They collided. Some struggled with objects that humans would handle without thinking.

A sprint exposes balance, actuation and control problems. Manipulation tests expose weaknesses in perception, force control and dexterity. A robot's ability to recover from a mistake may ultimately matter more commercially than its best-case performance.

ChatGPT could learn from trillions of words already sitting on the internet. There is no equivalent internet-scale repository showing robots how hard to squeeze every object, how to recover every dropped tool or how to open every drawer.

Ken Goldberg, a robotics professor at the University of California, Berkeley, has identified the shortage of robot-training data as one of the industry's central constraints.

China's huge manufacturing economy could therefore become more than a market for robots. It could become a data engine.

Ubtech said its humanoids have already accumulated more than 100 million pieces of data from real industrial manufacturing scenarios. Agibot has built dedicated data-collection facilities and has said that accumulating high-quality real-world behavior data is critical to improving robot performance.

It creates a potential physical-AI flywheel: more factories enable more robot deployment; more robots generate more data; more data improves models; better models make more robots economically useful.

Wang Peng, a researcher of Beijing Academy of Social Sciences, said: "Rather than a simple world of US brains and Chinese bodies, embodied AI is developing through deeply intertwined software, hardware and manufacturing ecosystems."