China’s robots rock, box and mix drinks. Can they outperform humans?

At the World Robot Conference in Beijing this week, humanoid machines played in a rock band, boxed and directed traffic in front of swarms of excited onlookers in a carnivalesque atmosphere.

Others stomped around the corridors among the crowds, some partnered with human celebrities, such as Yang Lan, a prominent Chinese broadcaster. When Yang accepted a rose from her humanoid partner, the robot raised its arms to form a heart sign to cheers from the surrounding throng.

Chinese policymakers have identified robotics as a strategic priority as well as a future source of economic growth, and have thrown the vast might of the world’s largest manufacturing sector behind their development.

The country dominates global humanoid robot production, with companies including carmakers, mobile phone manufacturers and newly founded pure-play robotics groups all pouring into the sector.

The five-day conference and exhibition, which started on Wednesday, has drawn more than 300 exhibitors. The second World Humanoid Robot Games open in Beijing on Saturday.

Wang Xingxing, founder of Chinese robot maker Unitree, the value of which quintupled on its first day of stock market trading this week, used the conference to provide a reality check to the hype in the hallways outside.

The hardware advances in humanoids have been spectacular — Unitree has exhibited robots it claims can jump higher than any person and able to reach a higher running speed than Olympic sprinter Usain Bolt.

But Wang said the industry’s “ChatGPT moment” — when humanoid robots will have the intelligence to perform free-ranging tasks with true autonomy — was yet to come.

“Perhaps in two to three years at the fastest, [or] five to 10 years at the slowest, everyone will be able to see a robot taken to any unfamiliar environment . . . capable of performing about 80 per cent of tasks,” Wang told the forum on Thursday.

He said such a moment would be a “tipping point” that could only be reached by gathering more data for physical AI models. “The biggest bottleneck globally is the generalisation of embodied intelligence; its capabilities are still insufficient.”

Others predicted the “ChatGPT moment” could come more quickly. He Wang, founder of Beijing-based Galbot, estimated it could occur within two years, while Gao Jiyang, co-founder of Beijing-based Galaxea, and Wang Xiaogang, chair of Ace Robotics, said next year was possible.

Even those who articulate lofty goals agree they will need to increase the scale of the AI models used to power their robots. To do this, they must first resolve how best to collect the complex data required for robots to understand the real world.

“A key challenge for the industry is enhancing the intelligence of robots, as they currently lack a comprehensive understanding of the physical world due to insufficient training data,” said Sheng Zhong, head of China industrials research at Morgan Stanley.

Corey Chan, an analyst at HSBC Qianhai, said most advanced companies at present ran models with 4bn to 7bn parameters, but it could take hundreds of billions of parameters for androids to be able to perform a wide range of generalised tasks well.

“The software still hasn’t been up to the level that we hoped,” Chan said. Models might only be able to progress to 40bn to 70bn parameters, which would only allow robots to engage in limited tasks competently, he added.

In general, humanoid robot training data is collected by humans wearing sensors, by simulation software or by the machines themselves performing tasks.

“Whenever data collection requires human operators, it entails substantial, unavoidable costs,” said Ju, an engineer at robotics company DexForce who did not want their full name to be published, adding that lower-cost video game simulations were becoming more popular.

Yu Chao, founder of Shenzhen-based Lumos Robotics, said the quality of data harvested could vary “significantly” depending on the method used.

Still, the market for humanoid robots appears to be growing — investment bank Morgan Stanley expects China to ship 50,000 models this year. In 2025, Chinese humanoid shipments numbered in the thousands, some of which were used in government-funded robot training sectors and others for entertainment or research purposes. But commercial deployments, such as in factories or the service sector, have been scarce.

Lumos’s Yu said robots were “absolutely” needed in factories, while other companies at the conference believed they would shine as personal assistants and in service industries.

Recommended

On the conference floor, almost every possible use for humanoids was on display.

On one stand, shiny metallic androids dressed in uniforms performed hand signals like traffic police. Elsewhere, yellow robotic arms helped human workers serve bubble tea and mixed cocktails. At robot producer Ubtech’s stand, humanoids stacked boxes in a staged warehouse.

An android woman with elvish features, including pointed ears, welcomed guests to the “Embodied AI Garden”, which also featured a robot unicorn.

Rinat Mirzaitov, who recently founded Humanoid Analytics to track deployments in the sector, said the industry would make progress incrementally, with robots mastering individual tasks one by one rather than achieving a sudden “magic moment” breakthrough.

Poe Zhao, an independent analyst who visited the conference, said humanoid robots for household use were “still far off”, adding: “The next year will be about paid deployments on factory floors, not about who has the best-looking humanoid.”