When Liu Yun lost her automotive industry job, she found work washing dishes and wiping tables at a Beijing studio. Wearing sensor bracelets and finger-mounted motion-capture chips, the 40-year-old repeated each task 200 times a day, which was already the limit she could sustain.
‘Rather than training robots, it feels more like teaching a two-year-old child,’ Liu told Jizhou Studio.
Her movements were being recorded as training data. Across China, freelance workers known as robot trainers are carrying out domestic tasks including picking up cups, opening bottle caps and washing dishes, so the actions can be turned into training sets for robots. The intended uses include kitchens, factories and supermarkets.
At Liu’s workplace, staff had to produce 160 samples a day. The strongest performers completed more than 260, while large screens ranked workers publicly by productivity and accuracy.
The work pays between 4,000 yuan and 6,000 yuan (US$600 and US$900) a month. Many workers leave after a few weeks, amid high turnover.
For Sun, a 22-year-old student from Shanxi province, the job was summer work. He folded clothes, swept floors and recorded standardised movements at a training facility, earning 120 yuan (US$18) for a day shift and 135 yuan for a night shift.
Guo Hongjuan, from Jiangsu province, works from home for about six hours a day. She arranges books, wipes tables and moves objects while wearing equipment that records her actions, earning around 4,000 yuan a month.
The work is meant to address a practical obstacle for robot makers: gathering enough examples of how people handle ordinary objects. Spirit AI, a Beijing-based robotics startup, uses about 1,000 contractors in homes and on production lines. The company reports that robots still struggle with fine-motor tasks such as unscrewing a bottle cap.
Gao Yang, Spirit AI’s co-founder, told Reuters he expects the next 1-2 years to be the first window for industrial robot use. He said simpler commercial services would follow, while homes remained a longer-term target.
Shi Guangyao, North China regional head of Beijing Yuanke Shijie, said capturing workers’ movements through motion-capture gear produced higher-quality data that could be applied to different kinds of robots.
Maniformer, a Shanghai company, launched a crowdsourcing platform on 23 September for ordinary people to record everyday tasks for robots to learn from and earn money. The platform allows people nationwide to register, receive motion-capture devices, record tasks and withdraw payment within two minutes.
Yao Maoqing, Maniformer’s chief executive, said large volumes of real human motion data were the factor holding robots back from factories, shops and homes.
China has built 70-plus embodied artificial intelligence training facilities, where competing robot manufacturers train side by side. In these centres, robots practise 45 ‘atomic skills’, with each robot repeating them up to 600 times a day.
The Shijingshan Embodied Intelligence Training Center in Beijing covers more than 1,500 square meters at the Sci-Fi Industry Innovation Center in Shougang Park. It has 4D Gaussian light-field capture systems and around 270 functional robots in its first phase.
A 5,000-square-meter facility in Shanghai’s Zhangjiang district brings together more than 100 humanoid robot models from roughly 12 manufacturers in one space. Across the country, the 70-plus facilities generate millions of data points annually as robots repeat their assigned skills.
China’s Ministry of Industry and Information Technology is targeting 100,000 robot units in 2026.
American humanoid startups pay $65–120 per hour for specialised teleoperation operators, while Chinese robot trainers earn a fraction of that wage.
For Liu, Sun and Guo, the immediate work remains physical and repetitive. Their recorded movements are being converted into training sets for robots intended for kitchens, factories, supermarkets and homes.
Comments (0)