Chinese humanoid robot companies are paying ordinary people to wear cameras and repeat everyday chores so the footage can train robot hands, while the costs, privacy rules and actual demand all remain unsettled.
Before a robot can learn to fold a shirt, someone has to film shirt-folding ten thousand times. Around Wudaokou in Beijing, home to China's top universities and robotics labs, that filming is now an industry.
Three generations, each more human
The first was teleoperation: a person stands back and moves a robot arm with a remote. It was slow, exhausting and expensive. The second strapped human hands into an exoskeleton. Generalist AI used a thousand workers to collect 250,000 hours in six months.
The third is egocentric capture. People wear a headset or a light helmet and just do their normal work, washing dishes or chopping vegetables, while cameras record the position of fingers and joints frame by frame. Nvidia's EgoScale project filmed over 20,000 hours with more than a thousand people.
JD's experiment in Suqian
In March 2026, Liu Qiangdong announced a collection centre in his hometown of Suqian, Jiangsu, aiming for 5 million hours by 2027 and 10 million by 2028. More than 100,000 residents signed up to wear JD's headgear and film five categories of daily tasks.
But the cost will not come down. An industry insider puts raw footage at about 100 yuan an hour, and usable data at 500 to 700 yuan. Adverts on Xiaohongshu offer 100 to 200 yuan a day, with no permanent contract and no benefits.
Data nobody wants, privacy nobody guards
The footage is hard to resell: customers want specific lighting and backgrounds, and custom data is often exclusive. So many firms collect only for themselves. About 30% of robot buyers still want second-generation gripper data, which is good enough for warehouses.
Abroad, Human Archive pays workers in India and Southeast Asia about $1 an hour to wear headsets, then sells the footage to AI labs in the US and Europe. Some Indian workers reportedly had no idea their videos were training AI.
A researcher jokes that America crosses the river by feeling the stones, and China crosses it by feeling America. Whether the far bank is the same for everyone, nobody can say.
Why it matters
The bottleneck in robotics is no longer the algorithm but how long people will agree to be filmed and for how much. When training data is bought by the hour from ordinary workers, labour law and privacy rules lag behind, and the first cost lands on the people in front of the camera.
Curated from high-quality sources, with concise summaries and key takeaways.
