This week, the Chinese firm AGIBOT said its 15,000th robot rolled off the production line, a milestone that says less about any single machine than about the speed of the ramp behind it. The company took a year to go from 1,000 units to 5,000, then just three months to reach 10,000, and it now claims the top spot in global humanoid shipments. The robotics race, as one report put it, is leaving the demo stage and arriving on the factory floor. But scaling the bodies turns out to be the easy part. The hard part is teaching them anything useful.
That is the quietly radical argument in a sweeping piece from Blaze Media, which maps an unexpected global division of labour. America builds the brains, the foundation models. South Korea builds the bodies, with the highest robot density on earth, roughly one robot for every ten manufacturing workers. China provides the classroom, with more than two million industrial robots generating real-world data at a scale no one else can match. And India, increasingly, supplies the teachers. The piece argues that the bottleneck in humanoid robotics is no longer the body or the brain. It is experience.
That word matters. Unlike a chatbot, which learned from text scraped off the internet, a robot has to learn the physical world by interacting with it: how tightly to grip an egg without crushing it, how to shift its weight across uneven ground, how to find a coffee pot and pour without spilling. These are things a small child does without thinking and the most advanced machines still cannot reliably manage. There is no web archive of that knowledge, so companies are now manufacturing it. In India, the report describes factory workers, masons, delivery drivers and homemakers being asked to wear head-mounted cameras as they work, capturing what is called "egocentric data," a first-person record of skilled human movement.
This is where the story turns uncomfortable. Many of those workers, the piece notes, were not told exactly why they were being filmed; some laughed when the cameras were first strapped on, then grew uneasy as they realised they might be training their own replacements. The questions write themselves. If a lifetime of accumulated skill becomes an AI dataset worth millions, should the worker share in that value? Can consent be meaningful when refusing the camera could cost you your livelihood? One Indian startup estimates it could take 100 million hours of first-person human activity before machines approach human-level dexterity. That is an enormous amount of ordinary work, recorded.
The tension is sharpest where the robots are heading next. In South Korea, Hyundai workers have voted to authorise strike action, with the company's humanoid program a central issue for the first time; the union wants a say before machines start doing jobs people now hold. The deployment side is no longer hypothetical, with robots moving from factories toward retail, construction and even, in startups' pitches, the home. The striking irony, which the Blaze piece leans into, is that the more capable AI becomes, the more it reveals how much quiet skill is bound up in the most ordinary human work. The robots may be mass-produced. The experience that makes them worth anything is still being extracted, one head-camera at a time, from people who mostly do not own a share of it.