The most revealing detail in Bloomberg’s report from an industrial park outside Beijing this week is not the humanoid arm placing a bag of Lay’s on a shelf. It’s the man standing nearby, filming himself grabbing cushions off a sofa and folding sheets on a bed, over and over, so that a machine can eventually do it without him.
That is the actual state of the humanoid industry in July 2026: the robots are not yet good enough to work, so people are being paid to work in front of cameras until they are.
The data problem nobody can buy their way out of
Language models got the internet for free. Robots got nothing. There is no scraped corpus for holding an egg without breaking it, or catching a glass before it slides off a table. The training data for physical work has to be manufactured, one recorded hour at a time.
The gap is not small. Companies broadly agree they need tens of millions of hours of demonstration data to train a usable embodied-intelligence model. Goldman Sachs analyst Jacqueline Du puts the current leaders “at around 500,000.” That’s roughly two orders of magnitude short — the distance between a promise and a product.
China’s answer was to industrialise the shortfall. Local authorities have opened 64 data-collection centres across the country, with 20 more under construction, according to research firm Interact Analysis. One outside Beijing is run by X-Humanoid. Inside, the facilities mimic supermarkets, assembly lines, offices, shops and homes — ordinary rooms rebuilt as sound stages, so that ordinary tasks can be performed into a lens.
Deploy first, learn later
The other half of the strategy is to skip the studio entirely and send the machines into real buildings, where the data is messier and therefore better. Robotera has humanoids in a dozen logistics hubs. Galbot signed a deal to send its machines to lift heavy things at CATL plants. AI² Robotics has installed humanoids in factories making cars, semiconductors and consumer electronics. Agibot says it wants more than a thousand robots in factories this year and more than ten thousand in 2027.
The money is behind it. Investors poured at least 100 billion yuan ($14.8 billion) into Chinese robotics so far this year — more than the previous five years combined, per Beijing tracker ITjuzi. Shenzhen’s X Square Robot alone drew 6.3 billion yuan. Unitree has pledged nearly half the proceeds of its upcoming $610 million listing to developing AI models rather than hardware, which tells you where the company that got famous for kung fu demos now thinks the hard part is.
Scale is the whole argument. The International Federation of Robotics counts about 300,000 robots installed in China in 2024 against 38,000 in the US. American firms are chasing the same models by different means: buying data, running lab simulations, and hiring human workers in India, Vietnam and other low-cost markets to generate demonstrations. Both approaches, note, bottom out in a person doing a task on camera. China just does it where the task actually happens.
That difference produced the best dunk of the quarter. In May, Figure AI ran a 50-hour livestream in which its robots sorted nearly 60,000 packages at roughly human speed with no intervention — a genuinely impressive run. Chinese developers were unimpressed. “Too clean,” they said. “Figure’s demo is still at a lab,” Agibot project director Ai Wen told Bloomberg. “Our deployment is in a real production line.” X Square’s Gan Ruyi went further, arguing the demo would actively hurt Figure’s model, which would learn irrelevant details from a package stream that never varies. “The packages that come your way are always different.”
The job that exists to end the job
Strip away the geopolitics and the $5 trillion market projections and there is a plain fact underneath: every one of these approaches requires a human to demonstrate the work first. That has quietly created a real occupation — data collector — and it is the only job in this story that is definitely growing.
It is also the only job whose success condition is its own deletion. A warehouse sorter who trains well gets a robot that sorts. A man who folds sheets convincingly for the camera gets a machine that folds sheets. Beijing is explicit about the destination: Barclays estimates the government wants humanoids to cover as much as 60% of its projected labour shortfall, and the Ministry of Industry and Information Technology wants 10,000 of them in factories by year end.
The honest caveat is that 500,000 hours against tens of millions is a long way from done, and anyone selling you imminent full automation is selling. But the gap is being closed the only way it can be — by paying humans to close it. Right now, the most valuable asset in that industrial park outside Beijing is not the robot. It’s the man on the bed, and his value is measured entirely in how completely he can be recorded.
Sources
- Insurance Journal / Bloomberg — China Sends Robots Into the World to Learn How to Be Human
- Morgan Stanley — The Humanoid 100: Mapping the Humanoid Robot Value Chain
- International Federation of Robotics — US robot industry returns to double-digit growth
- Barclays — 10 things you need to know about humanoid robots
