China’s humanoid robot industry is increasingly relying on human workers to gather the real-world data needed to train machines for household and factory tasks, according to a new report. The effort is part of a broader race to develop so-called physical AI, with Chinese firms betting that local, low-cost data collection can give them an edge.
The challenge is a familiar one in robotics: machines need huge amounts of visual and motion data to learn how to operate in unpredictable environments. For years, companies have used teleoperation, where workers remotely control robots as they repeat chores such as folding clothes or using appliances. But that method is expensive, slow, and limited in the kinds of situations it can capture.
Now Chinese companies are expanding beyond labs and into everyday settings. They are recording people performing household chores, assembly-line work, and other routine tasks in homes, factories, shops, and care facilities. Much of the data is first-person video that shows human hands at work, a format developers see as useful for training robots to imitate actions more accurately.
One example comes from JD.com, the e-commerce giant, which is working with local authorities in Suqian to generate 10 million hours of robotics training data over two years. In one neighborhood built around the effort, residents are paid to film themselves doing chores. Workers at an elderly care center and a kiwi farm have also been recruited to wear cameras that track hand movements.
JD.com has said it eventually wants 100,000 employees and 500,000 outside workers involved in the data collection effort. The company has framed the program as a way to create extra income for local residents while building the data needed for future robots.
The work is also spreading through China’s manufacturing sector. Two data vendors in Guangdong, a major industrial province, told Rest of World they were working with dozens of electronics and packaging factories. Employees there wear head-mounted cameras and wrist sensors that capture movement during assembly-line tasks.
Not every factory is eager to participate. One data supplier in Dongguan said some owners worry that the added process could slow production. To overcome that hesitation, suppliers are pitching a longer-term benefit: robots trained on the data may eventually be able to work in those same factories.
The report argues that China’s advantage lies in the combination of lower labor costs, public interest in robotics, and government support. While U.S. companies are also collecting real-world data, they often outsource work to countries with cheaper labor. Chinese firms, by contrast, can build large datasets at home and tailor them to local homes and workplaces.
That could matter because robots trained on domestic environments may be better suited to Chinese apartments, stores, and factories. Analysts quoted in the report said that the country’s hardware ecosystem and data pipeline may be a major strength, even if the U.S. still leads in top AI talent and robotics model research.
Still, experts say it is not yet clear whether these methods will produce robots that can operate reliably in arbitrary settings. An Oregon State University robotics professor said the strategy is plausible but remains unproven, noting that the industry is applying the same scaling logic that helped large language models improve with more data.
For some workers, the trend has created an unusual source of income. One stay-at-home mother in Shandong said she spends hours each day filming herself cooking and cleaning for pay, a job that lets her remain at home with her teenage son. She said nobody had previously paid her to do household chores.
The robot data race is still taking shape, but in China it is already creating a labor market of its own. Rather than replacing human movement, the next generation of machines is being trained by it.