Indian garment workers are wearing head-mounted cameras to record their daily tasks, generating "egocentric data" critical for training advanced AI robots. Companies pay these workers between Rs 250 and Rs 350 per hour for the footage, with some workers like Nagireddy Sriramyachandra in Chennai receiving 250 rupees—approximately $2.60—for an hour of video.

This first-person footage captures the rhythm of hands, precision of movements and speed of tasks, offering a direct record of human activity essential for humanoid robots learning to operate in real-world physical environments.

Industry experts identify data as the primary bottleneck in robotics development. Unlike large language models, which train on vast online text, robots require direct recordings of physical work. Companies collecting egocentric footage estimate the field will eventually require hundreds of millions—potentially billions—of hours of human activity drawn from factories, warehouses, shops and homes.

India has emerged as a key center in that global collection effort. An ecosystem of firms—including EgoLab, Humyn AI, FPV Labs, Micro1, Egodata, Neocambrian, XP Robotics, Objectways, Scale AI and CynLr—has developed to build data pipelines for robotics clients.

EgoLab, operating from Gurugram in Haryana, counts Tesla as a significant client. Tesla CEO Elon Musk has said he expects roughly 80 percent of Tesla's future valuation to come not from its electric vehicles but from its humanoid robot division.

The introduction of head-mounted cameras initially brought amusement to workers like Lalita, a 32-year-old garment factory employee outside Delhi. She described the devices as "like a CCTV camera mounted on us." That novelty quickly shifted to concern.

Workers grew more aware of their movements, worried about productivity monitoring. Conversations across sewing lines grew quieter. Some focused more intently on their work, knowing that every mistake, pause or distraction was being captured.

While Sriramyachandra confirmed she was paid for her recordings, at least one examination found workers at other operations received no compensation for footage later sold to technology companies—a discrepancy that points to inconsistent practices across the sector.

The arrangement creates a direct link between human labor and the automation designed to potentially replace it. Companies that secure large, diverse datasets will hold a structural advantage in deploying versatile robotic systems, making access to egocentric data a meaningful competitive moat. What remains unresolved is whether the workers generating that data will share in its value.