Indian workers are strapping head-mounted smartphones to their faces and filming their daily routines — cooking, cleaning, walking, shopping — for about 250 rupees ($2.40) an hour. The footage is being used to train artificial intelligence systems, especially for humanoid robots that need to understand how people move and interact with the world.
How the footage is collected
Workers wear a smartphone mounted on a headband or a harness, recording first-person video of ordinary tasks. The pay — roughly $2.40 per hour — is low by global standards but competitive in parts of India where wages for similar gig work range from 150 to 300 rupees an hour. The workers are not named in the reports, but they are part of a growing labor force that supplies the raw material for AI training.
The head-mounted approach captures what is called egocentric footage: video from the wearer's point of view. This is different from third-person video, which shows the whole body. Egocentric data helps AI models learn how a robot's camera would see the world if it were mounted on a humanoid torso or head.
Companies processing the data
Two companies are known to be involved in turning this raw video into usable training data. Objectways, based in the U.S. but with operations in India, specializes in data annotation for AI. Humyn Lab, based in Bangalore, focuses on human-centric data collection and labeling. Both firms take the egocentric footage and add metadata — bounding boxes, action labels, object tags — that machine learning models need to learn from.
The work is painstaking. Each second of video may require several minutes of human annotation to identify what is happening: a hand reaching for a cup, a foot stepping over a threshold, a person opening a door. The resulting datasets are then sold or licensed to AI developers building robots and virtual assistants.
Investor assessments project the humanoid robot market will reach $38 billion by 2035. That growth depends on robots that can navigate human environments — homes, offices, factories — without bumping into furniture or misreading a gesture. Training those robots requires vast amounts of first-person video showing how people actually behave, not just staged actions in a lab.
The Indian workers' footage fills a gap. Most existing egocentric datasets come from researchers or volunteers in wealthy countries. The Indian data adds variety: different homes, different objects, different cultural routines. That diversity helps AI systems generalize better, though it also raises questions about labor conditions and consent.
The work continues. Demand for training data is not slowing down, and the companies involved are likely to keep hiring. For now, the workers film, the annotators label, and the robots learn — one 250-rupee hour at a time.
