Estimated reading time: 7 minutes · Last updated:
Indian workers are strapping cameras to their heads while they cut hair, stitch garments and sweep floors, then selling the video as physical AI training data. Local collection firms gather and process the footage and sell it to robotics companies and research labs. Those buyers are building physical AI, the branch of artificial intelligence that lets robots act in the real world. Rashdan Technologies, one such firm, pays between $5 and $10 for a stretch of eight to 10 hours on camera. Ashim, 22, films his razor work for five or six hours a day, as first reported by the BBC. The demand is real and the pay is small. Nobody yet knows whether the people filming are training their own replacements.
Key takeaways
- How the pay works: For a recorded session lasting eight to 10 hours, Rashdan Technologies hands over between $5 and $10 (£3.74 to £7.47), which works out at roughly 50 cents to $1.25 an hour.
- Scale of collection: Humyn Labs, founded by Manish Agarwal and based in Bangalore and San Francisco, works with 96,000 data collectors across Latin America and Southeast Asia.
- Rising client demands: Shahil Baig of Rashdan Technologies says clients now want four to six cameras around a worker's head, which makes consent harder to secure.
- Jobs question unresolved: Bino Paul of the Tata Institute of Social Sciences says it is too early to tell whether this data will help automate the workers' own jobs.
- India's role: Agarwal argues India must move from supplying raw footage to cleaning, structuring and validating it, or watch the work lose value.
Table of contents
Why Physical AI Needs Footage of Human Hands
Robots meant for kitchens, factories and fields cannot learn everything inside computer simulations. That is the argument of Manish Agarwal, founder of Humyn Labs, a firm that collects and processes data generated by people. Robots need examples of real bodies doing real tasks, down to the angle of a wrist and the pressure of a grip. Agarwal's illustration is the difference between holding a plastic bottle without crushing it and holding a steel glass.
That kind of record barely existed until recently. Agarwal notes that, unlike text or images, physical human movement has never been stored anywhere online in bulk, so it has to be captured from scratch in many different settings. Most of it is video. Some collectors also wear hardware on their hands and fingers to register physical contact.
Humyn Labs, which has bases in Bangalore and San Francisco, currently works with 96,000 data collectors in Latin America and Southeast Asia. It sifts thousands of hours of their video for the most useful clips, labels them so an AI model knows what it is seeing, and packages the result for clients. Location matters too. Agarwal says a robot intended for homes or factories in the US should learn from environments that resemble them, and a residential house in India may not.
What Filming a Working Day Looks Like
Ashim, 22, wears a head camera that captures only his hands and fingers while he works with a razor. His clients' faces are blurred. An operations team trained him to keep his hands centred in the frame, the lighting bright and his head steady. He says the money is good, but the sense of doing something futuristic matters more to him, and he never expected a global tech firm to want his razor skills.
Mohammed Arman has stitched and cut garments in Bangalore's textile trade for 10 years. He sends money to his family in Bihar, and the camera simply pays him extra for work he already does. "For me, it's simple. This is survival," he says.
Jolly Arifur, a science teacher in Guntakal, Anantapur, wanted income she could earn at home around her young son. For seven months she has filmed herself scrubbing, cooking, sweeping tiled floors and folding laundry. She earns two to three dollars an hour and records three to five hours a day, which comes to roughly $6 to $15 daily alongside her teaching. All three are among hundreds of thousands of people worldwide paid to record what they do.
Why More Cameras Make Consent Harder
Rashdan Technologies collects and processes this footage for robotics companies and frontier AI labs. Its co-founder and chief executive, Shahil Baig, says each buyer wants something specific: industrial tasks, farm work, household chores, or small-business crafts such as pottery, embroidery, garment-making and candle-making. For eight to 10 hours of footage, the firm pays between $5 and $10 (£3.74; £7.47). Baig says that sum is meaningful for many of the people filming.
Baig describes worker consent and comfort as a priority. His firm records hands rather than faces, blurs faces when wider rigs catch more of a scene, and strips out names and contact details before handing data over.
The harder conversation is already arriving. Clients now want four to six cameras arranged around a worker's head to capture the whole scene, not only the hands. Baig says persuading someone to wear one small camera is very different from asking them to wear what amounts to a five-camera headset, and Rashdan pays more for that tier. He also expects dozens more collection firms to appear in India this year.
Are Workers Training the Robots That Replace Them?
It is too early to say, according to Bino Paul of the Tata Institute of Social Sciences (TISS), a senior professor in its School of Management and Labour Studies in Mumbai. AI may serve some workers as a tool while robots automate other jobs. His worry is timing: once machines begin taking over, workers will have little leverage. He warns that the people involved may only grasp what they helped build once it is too late to undo.
Paul wants governments to regulate AI. He calls for what he terms a socio-technical system, one in which human agency keeps ethics in place, rather than treating AI as a purely technical matter.
Agarwal sees a separate risk for India itself. He describes the country today as a supplier of raw material, with footage shipped to labs abroad. He expects the value of simple recording to fall as more firms compete for contracts. His prescription is for India to own the pipeline: cleaning, structuring and validating video, and eventually setting the standards for good training data.
| Worker | Work filmed | Location | Recording time | Pay |
|---|---|---|---|---|
| Ashim, 22 | Razor work with clients | Not stated | Five or six hours a day | Not stated |
| Mohammed Arman | Garment stitching and cutting | Bangalore | Not stated | Extra on top of usual work |
| Jolly Arifur | Cooking, scrubbing, sweeping, laundry | Guntakal, Anantapur | Three to five hours a day | Two to three dollars an hour |
What could push this trade up or down
The case for
- Robotics companies and frontier AI labs need footage of real physical tasks that no existing archive holds, which keeps demand for collectors alive.
- Multi-camera rigs command a higher pay tier at Rashdan Technologies, according to Baig.
- Agarwal argues India could capture more value by cleaning, structuring and validating footage rather than only recording it.
The case against
- Agarwal expects the value of raw recording to erode as more firms compete for collection contracts.
- Baig expects dozens more Indian collection firms this year, adding competition for the same buyers.
- Paul warns that if robots trained on this data automate jobs, workers will have little recourse once it happens.
What to be careful about
- Multi-camera headsets capture more of a worker's body and surroundings, widening the personal data at stake beyond hands alone.
- Pay is already as low as roughly 50 cents an hour at Rashdan's rate, and more competitors could push it lower.
- Workers may be supplying the data that later automates their own trades, a possibility Paul says cannot yet be ruled out.
- Buyers want environments resembling US homes and factories, which could limit demand for footage filmed in Indian settings.
The bottom line
The trade rests on a simple exchange. Labs and robotics firms need footage of skilled hands that the internet never recorded, and workers in India will supply it for a few dollars a day. Rates vary widely, from Rashdan's roughly 50 cents to $1.25 an hour to the two to three dollars an hour Arifur earns. Pressure is building at both ends. Clients want multi-camera rigs that reveal more of a person and their surroundings, while new entrants threaten to push prices down. Paul's question about replacement has no answer yet. Agarwal's push for India to move beyond filming is an argument, not a plan anyone has announced.
What to watch
- Watch whether dozens more collection firms launch in India, as Baig expects; no specific date has been given.
- Watch how many workers agree to the four-to-six camera rigs clients are now requesting, and what Rashdan pays for that tier; no figures or dates have been published.
- Watch for government moves to regulate AI of the kind Paul is calling for; no timetable has been set.
Frequently asked questions
How much do Indian workers earn filming their work for AI?
Rashdan Technologies pays $5 to $10 for eight to 10 hours of recording, roughly 50 cents to $1.25 an hour. Jolly Arifur earns two to three dollars an hour filming domestic chores.
Are workers' faces shown in AI training footage?
Shahil Baig says Rashdan Technologies records hands rather than faces and blurs faces when multi-camera rigs capture more. He says no names or contact details go into the data it hands to clients.
Why does it matter where robot training video is filmed?
Manish Agarwal of Humyn Labs says a robot built for US homes or factories should learn from similar environments. He doubts that an Indian residential house offers a suitable setting for training a machine meant for that use.
What is physical AI?
Physical AI is artificial intelligence that controls robots acting in the real world. It needs recordings of human movement and grip force, which Agarwal says have to be captured from scratch.
Related reading