Why China is sending humanoid robots to school

(Edited and refined by Josephine Hong, with the assistance of AI translation.)

The humanoid robot data training centre in Beijing’s Shijingshan district resembles a vocational school’s training facility — except that the students are all robots.

Dozens of humanoid robots, remotely controlled by data trainers, repeatedly pick up bottled water, place meal trays in microwave ovens and scan parcels. Other trainers, wearing headsets and holding controllers, repeatedly fold clothes and clear tableware.

School for robots across China

In operation since September 2025, the data training centre covers more than 10,000 square metres and is currently China’s largest robot school. It is jointly operated by Shijingshan-based state-owned Jing Shi Science and Technology Innovation Group and Leju (Shenzhen) Robotics, which hold 67% and 33% stakes respectively.

Zhu Kai, general manager of the Shijingshan Humanoid Robot Data Training Centre, said during a group interview organised by the Beijing Information Office on 10 August that, as an important infrastructure for Beijing’s embodied intelligence industry, the centre primarily provides data training for robots in scenarios such as home services, everyday life and industrial manufacturing.

Aside from Shijingshan, large-scale humanoid robot training bases are sprouting up in many parts of China.

According to data from technology market research firm Interact Analysis, at least 90 humanoid robot data collection and training centres were either in operation, planned or under construction in China as of the end of April 2026.

Of the 64 centres in operation, Zhejiang, Guangdong and Jiangsu house the most centres, with more than ten centres and at least 100 robots in each province.

Shandong, a major economic province with a comprehensive manufacturing base, is also seeking to keep pace. Chinese media reported that the province last month announced a list of 26 embodied intelligence robot training facilities as the first batch of priority construction projects.

Zhu said the Shijingshan training centre is being developed in five phases, with a total investment of 700 million RMB (US$104 million). The first three phases have so far received 200 million RMB in investment.

Bridging the brain gap

The unprecedented scale and intensity with which regions across China are developing humanoid robot data training centres stem from a national strategic push. Embodied intelligence was included in China’s government work report in 2025 and 2026, and was also identified as a key future industry in the outline of the 15th Five-Year Plan.

But how can China’s vast production capacity secure orders on a corresponding scale? The robots will need to be able to work like humans in more real-world settings, such as factories, logistics parks and eldercare facilities. The availability of high-quality data will determine whether these robots can perform such work and, ultimately, whether manufacturers can secure orders.

Zhu said that besides collecting high-quality data from real-world settings, the Shijingshan centre is building a virtual simulation training system. The aim is to generate large volumes of data, accelerate the evolution of the “embodied brains” of Chinese-made humanoid robots, and enable intelligent robots to serve industries across the economy and enter more homes.

The centre currently has 100 humanoid robots. Its data collection cluster produces 29,000 hours of data annually, covering nearly 700 scenarios, with a pass rate of 99% for individual data collection tasks, he said.

Physical data bottleneck

High-quality data is crucial to the future of the humanoid robot market, yet the industry widely recognises that there is a huge shortage of such data. Compared with generative artificial intelligence (AI), for which technology companies can draw text, images and audio-visual data from across the internet, far less data is available for the physical AI models required by humanoid robots.

During the 2026 World Artificial Intelligence Conference, Yao Maoqing, a partner at robotics company AgiBot and CEO of physical AI data service platform Maniformer, estimated that embodied intelligence requires at least tens of millions of hours and potentially hundreds of millions of hours of data before it can be practically deployed. Yet the total volume of usable, high-quality real-world physical interaction data available globally currently amounts to only several hundred thousand hours.

The shortage of data is a major pain point for the humanoid robot industry, but it is also an area in which China holds an advantage.

Zhu Kai said that China has an edge in data collection because of the relatively large scale of its data training facilities and strong support from local governments. He added, “Like infrastructure construction back then, this is a new form of infrastructure, and much of it is being undertaken by state-owned capital.”

“China also has an abundance of scenarios,” he said. Government agencies and businesses can both provide a rich range of settings for data collection. “Without scenarios, there is no data.”

This article was first published in Lianhe Zaobao as “中国多地兴建机器人学校 教机器人叠衣服做饭”.

Related: Faster than Bolt, but can China’s robots fold clothes? | How did Shenzhen-Hong Kong-Guangzhou become the world’s number one tech hub?