China's answer to the robot data problem: 64 data factories

China has stood up 64 robot data-collection centers, with roughly 20 more under construction, to feed its embodied-AI models the one input they cannot download: real human demonstrations. That figure, reported by Bloomberg on July 15, 2026, is the clearest sign yet that China is treating training data — not actuators or algorithms — as the decisive front in the humanoid race.

The centers are built to look like the places robots will work: supermarkets, assembly lines, offices, shops and homes. Inside, paid workers perform ordinary tasks over and over — folding sheets, grabbing cushions, moving stock — while sensors and robots record every motion. The output is demonstration data, the fuel for the vision-language-action models that increasingly run robot control.

A large language model can read the internet. A robot cannot fold a towel it has never seen folded. Someone has to do the folding, on camera, millions of times. — EW analysis

Why data is the bottleneck

Physical skills don't live in text, so the internet-scale corpora that trained chatbots are almost useless for manipulation. Robots need examples of hands doing things in the messy physical world, and there is no free reservoir of them. Bloomberg reports that leading companies hold on the order of 500,000 hours of such data today and estimate they need tens of millions of hours to reach reliable competence. That gap is the real moat — and it is a data-logistics problem as much as a science problem.

This is the axis on which China is trying to compete. Where much U.S. work leans on physics simulation and offshore data labeling, China is mobilizing organized labor and physical floor space to capture demonstrations directly, and deploying robots into factories to harvest still more. "This is where China has an edge: organizing labor and deploying data collection at scale," Gan Ruyi of X Square Robot told Bloomberg. The company frames the home as the ultimate proving ground precisely because it is unscripted: "Households are the ultimate testing ground for models. There is no script."

The numbers behind the push

The data build-out sits inside a broader industrial mobilization. China invested about 100 billion yuan ($14.8 billion) in the sector in 2026 — more than the previous five years combined, per Bloomberg — and the government is targeting 10,000 humanoids deployed to factories by year-end. Those deployments are themselves data-collection engines: Robotera has placed humanoids in a dozen logistics hubs, Galbot has paired with battery giant CATL for heavy lifting, and Ai2 Robotics has installed units at automotive, semiconductor and consumer-electronics plants. For scale context, Bloomberg notes roughly 300,000 robots were installed in China in 2024 against about 38,000 in the U.S.

Why it matters

If the humanoid race is decided by who accumulates demonstration data fastest and cheapest, then labor organization and floor space become strategic assets, and China is building both deliberately. The counter-case is that scraped human demonstrations may not transfer cleanly to robot bodies with different kinematics, and that simulation or better learning methods could shrink the data appetite — routes that would blunt a brute-force data edge. Either way, the contest has moved: the question is no longer only who builds the best robot, but who can most cheaply teach one.