Robot makers pay up to about $150 for an hour of training video. The people filming it can earn about $1.20: inside China's embodied-AI data supply chain

To teach a robot to fold laundry or pick up a cup, developers feed it huge amounts of footage of people or robots doing those things. This is known as embodied data, and it has become a hot commodity in China, where "embodied AI" (AI that controls a physical body and acts in the real world) is one of the busiest areas of tech investment. Nearly everyone in the industry says high-quality data is scarce, so collecting it has turned into a business of its own.

On Nov. 11, China's Sci-Tech Innovation Board Daily (Kechuangban Ribao) published an investigation based on several days of interviews with data collectors, suppliers, buyers and county-level collection contractors. Tech outlet 36Kr republished it with permission on Nov. 12. The headline finding: when data is priced by the hour, prices for that hour can differ by a factor of more than 100.

According to reference prices from Jin Ge, co-founder and CEO of Lingyu Intelligence, first-person footage recorded by people wearing cameras or glasses (called "embodiment-free" or egocentric data) goes for about 100 to 300 yuan an hour, roughly $15 to $45. Data generated by real robots goes for about 300 to 1,000 yuan an hour or more, about $45 to $150 and up. But another industry insider, quoted under the pseudonym Wang Jian, said some raw single-camera video shot with phones or GoPros sells for just 5 to 10 yuan an hour, under $1.50.

Where does the money go? Several people in the business told reporters that an order paying more than 100 yuan an hour can pass through five or six layers of middlemen. By the time it reaches the people actually recording, the pay is often around 15 yuan an hour and sometimes as low as 8 yuan, about $1.20. One part-time collector, given the pseudonym Xiaoyun, earned 15 yuan an hour folding and shaking out a coat at home over and over. After packaging and resale, footage like hers was quoted to robot companies at up to 300 to 500 yuan an hour. A county-level contractor said his floor for collectors is 30 yuan an hour, because anything lower means either the project can't make money or the data's sourcing is questionable. The report notes that these figures come from different projects and shouldn't be read as fixed cuts of a single order.

A quieter problem is the same data being sold again and again. Jin said it is "very common" for embodiment-free data to be resold three to five times. Some teams keep private backups after delivering an order. One person described collectors strapping several phones together with rubber bands to record the same motion at once: one copy goes to the client and the rest end up on the black market. Jin's team now adds traceable "data fingerprints" to robot-generated data, but he said there is no good fix yet for human-recorded footage.

Quality is another issue. Wang estimated that fewer than 1% of real robots produce data that meets training standards. A representative of embodied-AI company Yuanli Lingji said that training on 5-yuan-an-hour footage "not only wastes compute, it pollutes the model." The report also says that as of June 2026, more than 70 embodied-AI training centers had opened in China. Some centers built around remote-controlled robots have since shut down, and the industry is shifting toward people recording data in real settings. In September, Mech-Mind founder Shao Tianlan publicly argued that some embodied-AI companies were booking "fake and unsustainable revenue" through data centers and related-party deals.

Why it matters: how capable a robot can become depends heavily on what it is trained on. If much of the data on the market has murky origins, gets resold repeatedly or is simply poor, robot companies may spend heavily without getting better models. Investor valuations based on how many hours of data were collected may also be inflated. The people interviewed generally agreed that buyers are shifting from asking "how many hours did you deliver?" to "what can the robot do now that it couldn't before?"

For ordinary people, there are two takeaways. First, a new kind of gig work is emerging: you can take orders to film yourself doing chores at home, but layers of middlemen can push the pay very low, so ask about the rate, how the footage will be used and who owns it. Second, whether future home and store robots turn out to be reliable depends a lot on whether today's training data is real and clean. Note that these prices and practices come from interviews with industry participants, some of them under pseudonyms. They are not official statistics. Dollar figures are approximate conversions.

Original sources published: Sci-Tech Innovation Board Daily, Nov 11, 2026, 10:18 Beijing time / Nov 10, 2026, 9:18 PM EST; 36Kr (republication), Nov 12, 2026, 09:29 Beijing time / Nov 11, 2026, 8:29 PM EST