Field Investigation of Embodied AI Data Collection in China: Data Sellers Lack Bargaining Power, Market Decides Which Generation of Technology Survives
ChinaTalk Essay Contest Grand Prize winner A. Chen conducted on-the-ground interviews in Wudaokou, Beijing, regarding the embodied AI data collection industry. She found that despite capital influx and high enthusiasm for humanoid robots, data suppliers actually lack bargaining power: approximately 30% of the robot market still only purchases second-generation UMI gripper data, first-person video data is difficult to resell, and cost floors limit the space for scaling. Chinese companies are more cautious than their American counterparts, waiting for the market rather than technological narratives to determine direction.
In October 2026, ChinaTalk Essay Contest Grand Prize winner A. Chen traveled to Wudaokou in Beijing—the location of Tsinghua Science Park and a hub for China's embodied AI research institutions—to conduct field interviews with researchers and entrepreneurs engaged in robot training data collection. Her core finding was that beneath the capital frenzy and social media virality, embodied AI data collection is actually an unpredictable business where data sellers lack true bargaining power.
A. Chen first outlined the three-generation evolution of robot data collection technologies. The first generation was teleoperation, which emerged around 2017; that year, Stanford University Professor Fei-Fei Li pioneered related projects where users controlled robotic arms via remote controls or VR devices, with hardware costs reaching hundreds of thousands of yuan per set, capable only of training simple binary gripper commands. The second generation was isomorphic collection, marked by the publication of the Universal Manipulation Interface (UMI) paper by Stanford University Professors Cheng Chi, Zhenjia Xu, and Shuran Song in 2024. In this method, human subjects connected their hands to robotic exoskeletons using Velcro straps, elastic bands, or metal hinges, allowing for faster and more precise data collection than teleoperation. The US company Generalist AI reportedly collected over 250,000 hours of data using one thousand workers within six months. The third generation, egocentric collection, rose in early 2026; NVIDIA introduced this concept in its EgoScale project on February 18, 2026, fixing cameras on the foreheads and wrists of over 1,000 people to film 20,854 hours of first-person video covering daily tasks such as washing dishes and cutting vegetables.
However, the market did not eliminate older solutions due to generational technological updates. Dr. Jackson Guo, a researcher at the Kunlun Nest Robot Open Laboratory in Beijing, told A. Chen: "Second-generation data collection—homomorphic exoskeletons and UMI—is still very popular among Chinese buyers. Unless egocentric data collectors can harvest, say, 10 billion hours of data, we cannot surpass UMI in the market." An anonymous industry insider estimated that about 30% of the robot market still only seeks to purchase second-generation UMI-based gripper data. Warehouses and commercial companies are mostly satisfied with binary grippers, while only academic research laboratories require dexterous hand data.
The cost structure further limits the bargaining space for data sellers. An anonymous industry insider told A. Chen that the cost of collecting one hour of raw data is approximately 100 yuan, while the cost of generating usable data is about 500 to 700 yuan per hour. He said, "If egocentric data were as cheap as water (5 to 10 yuan per unit), people would buy it without hesitation, but I don't think data collection costs can go lower than they are today."
Dr. Iris Wang, an entrepreneur and researcher at Tsinghua University, explained the power dynamics within the industrial ecosystem to A. Chen: "The market determines the direction of the industry and innovation, not the other way around." In the data collection ecosystem—where data is sold to startups and tech companies, which then sell finished robots to consumers—egocentric data suppliers lack true bargaining power. Customer demand dictates the direction of data collection and research.
Egocentric data is not yet a universal product. A. Chen found that video material is difficult, if not impossible, to resell: customers have specific requirements for the lighting, background, and actions of every video, and customized data is often protected by exclusive intellectual property clauses. Therefore, many companies collect data solely for internal use, creating complex and fluctuating dynamics across the supply chain.
Regarding strategic differences, Chinese companies showed a clear attitude of waiting and seeing. JD.com founder and chairman Liu Qiangdong announced on March 16, 2026, the establishment of a data collection center in his hometown of Suqian, Jiangsu Province, with the goal of collecting over 5 million hours of video by 2027 and reaching 10 million hours by 2028, with over 100,000 citizens from Suqian already signing up. However, A. Chen observed that most Chinese founders and researchers are still waiting for the market to decide the direction of the next generation of data collection methods—contrasting with the large-scale bets made by US companies. Build AI has raised over $30 million, aiming to collect over 1 billion hours of egocentric video by 2028; there are also more than a dozen startups in Silicon Valley sharing this goal.
Dr. Wang also provided a useful layered judgment: startups raising over 1 billion RMB usually pursue egocentric data for pre-training, while those raising less than 1 billion RMB focus on post-training methods that require little to no egocentric data. This means that the scale of capital itself is determining the choice of technological route.
At the end of the interview, Dr. Guo joked to A. Chen: "The US crosses the river by feeling the stones, and China crosses the river by feeling America." In A. Chen's view, this narrative is far more complex than US-China competition—data flows from India to China and then to Silicon Valley, with supply chains spanning oceans and embedded in the existing world order.
For investors, the core mechanism revealed by this field report is: in the embodied AI industry chain, the economic status of the data link depends on whether buyers truly need data of a specific generation, rather than depending on technological narratives or financing scales. Approximately 30% of market demand remains stuck at second-generation UMI data, cost floors limit profit margins, and customized demand hinders the circulation of data as a general commodity—these factors together determine that data sellers currently lack bargaining power. The difference between the wait-and-see strategy of Chinese companies and the aggressive bets of American companies may reflect different capital structures and risk preferences, rather than the superiority or inferiority of technological judgment.