Egocentric Data Is the New $2.62-An-Hour AI Training Category

The cheapest raw material in humanoid robotics is not steel, not rare earths, not compute. It is first-person video of skilled hands doing real work, captured on phones strapped to the workers who own them, then shipped into a training pipeline priced in informal wages. The category has a name the industry is now using in print: egocentric data, the new physical layer of the AI training supply chain.

The mechanism is the same one screen-based data annotation ran on for a decade, only it has crossed the body. Bloomberg's August 12 investigation and AFP's May reporting from a Tamil Nadu textile plant document the same workflow in two rooms: a low-single-dollars hourly wage, a head-mounted camera, an opaque buyer, and clips filed under "humanoid training data." Sunita Rathore, the New Delhi waste picker at the center of Bloomberg's reporting, earns roughly ₹20,000 a month (about $210) for the footage, and the confession inside that wage is what it does not buy: joint torques, contact forces, grip pressure. Pixels teach a hand's path through a space. They do not teach the hand.

The pattern repeats whenever a new layer of the AI stack needs human labor: name the category, point a camera at the cheapest supplier, pay the local rate, ship. Whoever sits at the bottom of that chain funds the robots that will price their labor out. The category exists whether disclosure improves or not; the only choice is whether the reader learns to recognize it before the next headline.