Robotics Firms Are Paying Indian Factory Workers to Teach Machines Their Own Trade

Bloomberg reports that robotics companies are competing to collect videos of humans stitching shoes and welding steel to give their machines new skills, with workers wearing iPhones strapped to their heads angled to catch every motion of their hands. Dawn reports rates of around 250 rupees, about 2.50 dollars, for an hour of filmed work. At Metaintro, we cover the jobs that get built on the way to automation as carefully as the ones automation removes, and this is both at once. The work is real, it pays, and its product is a machine that does what the worker does.

What is egocentric data and why do robots need it?

It is first-person footage, filmed from the worker's own point of view, and it exists because robots learn differently from chatbots. Developers feed this footage, called egocentric data, into specialised AI models to help robots copy humans. A language model can be trained on text that already exists. A robot that needs to fold a shirt or strip a label from a plastic bag has no equivalent archive, because nobody was recording hands doing ordinary work from the angle a machine would need.

That gap is what created the job. The tasks being recorded are deliberately mundane, including folding clothes, making coffee, preparing sandwiches, slicing mangoes and washing dishes. In industrial settings the footage covers stitching shoes and welding steel. One company's chief executive listed the requests as folding clothes, coffee making, cooking a very specific thing and sandwich making. The specificity matters, because a robot needs many examples of one narrow action rather than a broad impression of a category.

The equipment reflects the same precision. Contributors use smartphones, GoPro cameras, head-mounted RGB cameras, motion sensors worn on wrists and legs, and smart glasses supplied by the data companies. Some work at home, others in factories or in specialised studios. The result is a recording of not just what a task looks like, but of how a body performs it.

What does the work actually pay?

Around 250 rupees an hour, roughly 2.50 dollars, for the home-based filming. One contributor, a 25-year-old in Chennai, framed it as a straightforward comparison, asking who else would pay 250 rupees an hour for work done at home. That is the honest case for this work, and it deserves to be stated as clearly as the criticism.

The comparison becomes sharper against what the same workers earn otherwise. Bloomberg profiles a 43-year-old who has spent years sorting and cleaning discarded plastic for about 20,000 rupees a month, roughly 211 dollars, in a recycling colony on the western fringe of New Delhi. Against that baseline, an hourly rate of 250 rupees for filming familiar movements is not exploitative on its face. It is better paid, safer and less physically punishing than sorting waste.

Volume is where the picture gets more complicated. A 21-year-old engineering graduate described recording around 90 videos a day, each about four minutes long, and called the work merely tolerable. Ninety short takes of repeated household actions is closer to a production line than to casual side income, and the description of it as tolerable rather than good is worth taking at face value.

How does this compare with other data work globally?

It sits in a well-documented pattern rather than standing alone. The International Labour Organization surveyed 3,500 workers in 75 countries on five microtask platforms and found average earnings of 4.43 dollars an hour when only paid work was counted, falling to 3.31 dollars once unpaid hours were included. Median earnings on the fuller measure were 2.16 dollars an hour. Regionally, workers in Asia and the Pacific earned 2.22 dollars per hour of paid and unpaid work, against 4.70 dollars in Northern America.

Two things follow from putting those numbers beside the 250 rupee rate. The first is that egocentric video work pays at roughly the level established across data work in the region rather than above it. The second is a structural warning the ILO data makes visible. Workers spent an average of 20 minutes on unpaid activities for every hour of paid work, searching for tasks, taking unpaid qualification tests, researching clients and writing reviews. Task scarcity drove that, with 58% reporting that available work was insufficient and 88% wanting more hours. An hourly rate only describes earnings if the hours are actually there.

Academic work on the sector reaches a blunter conclusion about how it is organised. A four-country study of data work covering Venezuela, Brazil, Madagascar and France, drawing on data collected between 2018 and 2023, found that the cross-country supply chains linking data workers to core AI production sites are "reminiscent of colonial relationships" and maintain historical economic dependencies, generating inequalities that compound with those inherited from the past. The value created flows toward the companies commissioning the data, and the work flows toward the countries where labour is cheapest.

Who is actually building this industry in India?

A layered supply chain, which is why the pay looks the way it does. One operator runs the primary operation with offices in India and the United States, claims Fortune 500 multinationals as clients and works with Amazon SageMaker. Beneath that sit subcontractors, including a consulting firm in Andhra Pradesh that supplies recordings to roughly a dozen larger firms and works with around 2,000 contributors. A separate Bengaluru-based firm records both video and conversations.

That structure explains the economics without requiring anyone in it to behave badly. Each layer takes a margin, and the person wearing the camera sits at the bottom of the chain, furthest from the client paying for the dataset. It is the same shape the ILO documented on microtask platforms, where the platform pays the worker the client's price minus its fee.

The country-level ambition is explicit. India has positioned itself as a global middleman for the creation, processing and annotation of AI data. That is a real economic strategy with real employment behind it, and it is also a position in the value chain rather than at the top of it. For Indian workers and for policymakers, the open question is whether middleman status becomes a stepping stone or a settled role.

Is this work training your own replacement?

In the most literal sense available, yes, and the people doing it know. A 55-year-old flower garland maker in Bengaluru took part despite concerns about job displacement in her own industry. The Bloomberg feature is framed around exactly this, describing thousands of Indian workers helping AI firms train robots to replace them.

The scale being planned is what makes the concern more than theoretical. Morgan Stanley predicts more than one billion humanoid robots could be operational by 2050, primarily for industrial and commercial applications. Whether or not that forecast proves accurate, it describes what the investment is aiming at, and industrial and commercial tasks are precisely the ones being filmed today.

It is still worth resisting a tidy moral. A worker choosing between sorting plastic waste for 211 dollars a month and filming household tasks at 250 rupees an hour is making a rational choice with the options actually in front of her. The problem is not that individuals accept the work. It is that the gains from what they produce accrue somewhere else, which is the finding the four-country study reached about data work generally.

Which workers are most exposed?

Informal workers, and they are largely missing from the debate. India's own government think-tank has acknowledged that discussions about AI and labour focus on white-collar professionals while largely ignoring the impact on the country's 490 million informal workers. That is the single most important number in this story. The population most likely to be affected by physical task automation is the population least represented in the policy conversation about it.

Informal work is also the hardest to protect. There is usually no employment contract to renegotiate, no union to consult, no notice period and no retraining budget. When a task is automated in a formal workplace, there is at least a process. When it disappears from informal work, it simply stops being available. Skilled trades offer one route out of that exposure, as Metaintro sets out in how to get machinist certification and advance a manufacturing career and the electrician career guide covering skills, earnings and outlook. Metaintro has covered the equivalent vulnerability in gig and platform work in what happens when a platform closes and the income goes with it, where the warning period was measured in weeks.

The timing question also matters more than the certainty question. Robots capable of general household or factory manipulation are not deployed at scale today, and the reason this data is being collected is that they cannot yet do these tasks. That gives a window. Windows are only useful to people who know they exist, which is the argument for treating this as urgent information for informal workers rather than as a technology story.

How does this compare with the last wave of data work?

Closely enough to be instructive, because the text annotation era ran the same course and its economics are now well documented. Research on Amazon Mechanical Turk analysed 2,676 workers completing 3.8 million tasks and found a median hourly wage of about 2 dollars, with only 4% of workers earning more than 7.25 dollars an hour. The average requester paid more than 11 dollars an hour, but the lower-paying requesters posted far more work, so the typical worker spent most of their time on the worst-paid listings.

Two features of that earlier wave are worth carrying into this one. The first is the gap between the advertised rate and the realised rate. The MTurk analysis was notable precisely because it accounted for unpaid time spent searching for tasks, working on tasks that were rejected and working on tasks that were never submitted. A per-hour figure quoted by a client is not the same as a per-hour figure earned by a worker, and the difference sits in rejected takes and idle time. Anyone filming egocentric video should ask specifically what happens when footage is rejected for quality, and whether that time is paid.

The second is what happened to the work itself. Text annotation did not disappear when models improved, but its centre of gravity moved from open crowds toward vetted specialists, and the platform that defined the earlier era is now closing. Metaintro covered that ending in Amazon shutting a 21 year old gig work platform. The workers who came through that transition best were those who moved into review and quality roles rather than competing on volume, which is the same advice that applies here and is worth acting on early rather than late.

There is one meaningful difference, and it favours the current contributors. Text annotation could be performed by anyone literate in the target language, which made the labour pool effectively unlimited. Egocentric data requires someone who can actually perform the task being recorded, whether that is stitching shoes, welding steel or making a garland. Genuine skill is part of what is being captured, and that gives these workers slightly more leverage than the previous wave had, provided they recognise it.

What should employers and HR teams take from this?

That the data supply chain behind an automation project is now a workforce question, not only a procurement one. If your organisation is buying training data, the layered structure described here applies to you. A primary vendor with Fortune 500 clients sits above subcontractors who work with around 2,000 contributors, and the conditions at the bottom of that chain are attached to your project whether or not they appear in your contract.

Three questions are worth asking any vendor. What is the effective hourly rate received by the person generating the data, as distinct from the rate invoiced. What happens to rejected work and who absorbs that cost. And what consent and retention terms cover footage recorded in someone's home, particularly where conversations are captured alongside video. These are answerable questions, and a vendor unable to answer them is describing a risk rather than a supply chain.

There is a reputational dimension as well. Academic work on data work across Venezuela, Brazil, Madagascar and France concluded that the supply chains linking data workers to core AI production sites resemble colonial relationships and sustain existing economic dependencies. That framing is already established in the research literature, and organisations buying this data should expect it to surface in scrutiny of their AI programmes. Getting ahead of it is considerably cheaper than responding to it.

What should you do if this work is available to you?

Take it if the rate beats your alternative, and treat it as income rather than as a career. On the evidence, it pays comparably to or better than a good deal of informal work, it is flexible, and it can be done from home. Those are genuine advantages, particularly for workers whose other options involve physical risk. There is no good reason to refuse it on principle while better-paid alternatives remain unavailable.

Three cautions apply. First, confirm what the hours actually are before you plan around the rate, because data work across the sector is defined by insufficient task availability rather than by low headline pay, with 58% of surveyed workers reporting not enough tasks. Second, understand what you are agreeing to, since the footage is of you, in your home or workplace, and one firm records conversations as well as video. Ask what is retained, for how long and whether it can be resold. Those terms are part of what you are being paid for, and like most terms they are easier to raise before you start than after, a point Metaintro makes more generally in what you can negotiate when the rate will not move. Third, treat the work as temporary by design. The dataset your footage joins exists to remove the need for the footage.

The more durable move is to convert proximity into position. Contributors in this chain are close to a growing industry, and the roles above them, in quality review, annotation supervision, studio coordination and client delivery, require exactly the understanding of task quality that a high-volume contributor develops. Someone who has recorded ninety takes a day knows what a usable take looks like, and that judgment is the scarce input as datasets grow. Metaintro has tracked the same climb from task work into oversight in the hours workers now spend checking machine output and how managing AI agents became its own job.

What does this mean for your career?

The broad lesson travels well beyond India, and it is not the one about robots. The pattern here is that automating any physical task requires a period in which humans are paid to demonstrate that task, and that period is a hiring window. It happened with text annotation before language models became fluent, and it is happening now with movement. Anyone whose work involves a specific manual skill should expect a version of this offer eventually, and should recognise it as a signal about their own timeline rather than as a curiosity.

The practical response is to ask what part of your work is not being filmed. Egocentric footage captures motion, sequence and technique. It does not capture judgment about when a task should be done differently, responsibility when something goes wrong, or the relationships that get work through a real organisation. Those are the components that survive the transfer, and they are the ones worth documenting and building on. Metaintro has covered what that looks like when workers are asked to prove their own contribution in what front line workers want from AI at work. Metaintro has covered how that division is already reshaping technical fields in what happens to broker and dispatcher jobs as AI takes over freight quoting and how AI compresses wages without eliminating roles.

For readers in India specifically, there is a policy point worth carrying into any conversation about this. With 490 million informal workers largely absent from the national AI and labour debate, the gap is not primarily technological. It is one of representation. The workers filming these datasets are the earliest and clearest evidence of what physical automation will touch, and they are currently visible to the companies buying their footage and to very few others.

Related Articles

• Samsung Is Building a Humanoid Robot Division and What It Means for Factory Jobs
• AI Is Not Taking These 5.8 Million Jobs, It Is Shrinking Them
• Front-Line Workers Are Fine With AI, They Just Want to Know How It Is Used
• Only 19 Percent of Junior Staff Say They Have the AI Tools They Need
• AI Is Hiding a Learning Debt That Could Stall Your Career
• Game Studios Are Shrinking to One-Person Teams as AI Takes the Work
• 74 Percent of Workers Have Skipped Work Over Unsafe Conditions
• Jobs That Do Not Follow the 9 to 5 and How to Qualify for Them
• Even AI Research Jobs Are Not Safe After Amazon's Latest Round
• AI Labs Are Hiring Away the Professors and Closing the Research
• What to Ask Before Taking a Job With an AI Title
• Managing AI Agents Is Becoming Its Own Job and Its Own Burnout
• How to Get Machinist Certification and Advance Your Manufacturing Career

People Also Asked

Q: How much do AI robot training data workers in India earn?

A: Reported rates are around 250 rupees an hour, roughly 2.50 dollars, for home-based filming. For comparison, the ILO found median earnings of 2.16 dollars an hour across five microtask platforms once unpaid time was counted, and 2.22 dollars an hour for workers in Asia and the Pacific. One worker profiled had previously earned about 20,000 rupees a month sorting plastic waste.

Q: What is egocentric data?

A: It is first-person video shot from the worker's own viewpoint, fed into specialised AI models so robots can copy human movement. It is captured using smartphones, GoPro cameras, head-mounted RGB cameras, wrist and leg motion sensors, and smart glasses. Typical recorded tasks include folding clothes, making coffee, slicing mangoes, washing dishes, stitching shoes and welding steel.

Q: Will this work last?

A: Probably not in its current form, because the data exists to remove the need for it. Morgan Stanley predicts more than one billion humanoid robots could be operational by 2050, mainly in industrial and commercial settings. Treat the work as income now and use the exposure to move toward review, supervision and quality roles, which need the judgment that a high-volume contributor develops.

Work that trains your replacement is still worth understanding from the inside, and the skills it builds are transferable if you aim them at the right roles. Metaintro surfaces verified openings from employers hiring now, including the quality, annotation and oversight roles that sit one level above task work. Create a free profile to see what your experience already qualifies you for.