The Women Training Humanoid Robots From Their Kitchens In Rajasthan

Pooja straps an iPhone to her forehead with a buckle band, adjusts the frame until both her hands are clearly visible, and opens an app called Atlas Capture. A list of tasks appears on screen: mop the floor, clean the windows, make a cup of tea, organise the pantry, clean the fridge.

She taps on 'make a cup of tea'.

An instruction on the screen reads: Recordings must be done standing up. Do not sit during your video.

The 30-year-old homemaker in Pathena, a village in Rajasthan's Bharatpur district, is a mother of two. She is part of a growing workforce across India who are being recruited to train humanoid robots, often without formal contracts, for wages that no existing labour law governs.

Pooja was married while still in her first year of graduation. She went on to complete a bachelor's degree in Geography, but work, in the conventional sense, never followed. "I always wanted to work," she says. "But after marriage and kids, stepping out becomes difficult."

Now, her workplace is her home, and the work is this: she moves to the kitchen, reaches for a vessel, fills it with water, places it on the stove — all while making sure every movement stays within the camera's view and both her hands remain visible at all times.

The phone, mounted on her head, records what she does and how she does it.

Each clip she records will travel far beyond her kitchen. It will pass through a local vendor company in Alwar, move to a data aggregation startup, get cleaned and labelled, and eventually reach a robotics company in California or Norway building humanoid machines that cost between $20,000 and $100,000 a unit. The footage will become part of a training dataset that teaches these robots how to see, move, and act in the physical world.

For every hour of usable footage, Pooja earns Rs 250, as per her arrangement. Many workers in this pipeline, she says, also get closer to Rs 50.

There is no official count of how many people in India are doing this kind of work. However, figures reported by individual companies point to a workforce already in the thousands. Awign, one of the data firms in the pipeline, says more than 10,000 workers record data daily. Objectways has around 700 people collecting about 1,500 hours of footage a day. Human Archive says it has more than 1,000 active headsets deployed across India.

Taken together — and accounting for smaller firms in the space — Decode estimates the number could be between 20,000 and 30,000.

 Data supply chain (Made with AI)

Learning to Perform for the Camera

Every movement must be deliberate. Both hands have to remain visible at all times, forcing even the most routine actions to be performed differently. For a clip to qualify as usable, each recording has to run for at least 15 minutes. Tasks must eventually add up to 90 minutes of footage. Workers can break this into multiple clips, typically 15 to 30 minutes each, but the constraints remain the same: stand, stay in frame, keep both hands visible.

"Mopping or cleaning windows can go on for that long," Pooja says. "But tea doesn't take that much time."

So she stretches it. She exaggerates her movements. When she reaches for the jar of tea leaves on a shelf above, she lifts both hands into the frame, though only one is needed. Sweeping the floor, normally a one-handed task, is performed with both hands, because that is what the system demands.

Once uploaded, the videos are reviewed within 24 hours. The app flags what portion is unusable, usually with a brief note. The most common rejection: hands not visible.

In her early days, Pooja failed this test repeatedly. "You forget that the camera is fixed on your head, not on your hands," she says. "When I was cleaning windows, if my hands moved lower, I didn't always tilt my head. The recording would miss it."

 The app flags unusable footage, most often because the hands aren’t visible.

Over time, she has learned to adjust — not just her posture, but her instincts. Her body has learned to perform for a machine.

And when the footage is rejected, she is not paid. The time is gone.

The Supply Chain

Behind Pooja's daily routine sits a layered system that connects her kitchen to a robotics lab. Tracing it shows how AI training data actually gets made — through a chain of intermediaries, each taking a cut, each adding distance between the worker and the company that ultimately uses her labour.

The vendor layer. For Pooja, the work arrived through family. Her brother-in-law is a co-founder of Alwar-based Vega Solutions, one of several small firms operating at the very bottom of this chain. Companies like Vega function as last-mile labour providers: they recruit workers, distribute recording devices, and ensure a steady supply of raw footage.

"We don't access the video in any form," said Rajat Chaudhary, a co-founder of Vega Solutions. "Our job is onboarding. The bigger challenge right now is that we are trying to get as many people as we can for a diversity of tasks, because data is never enough."

Vega is not new to labour supply. "We have been in labour vendoring for almost a decade," Rajat said. "We already had a network of freelancers who worked as exam invigilators or in temporary staff roles. Now, we have moved that same workforce into this."

A company that once supplied exam invigilators now supplies training data for humanoid robots. The infrastructure of India's informal labour economy — networks of temporary workers managed by small vendor firms — is being repurposed for the AI industry.

The shift has required capital. Vega purchased 30 iPhones for Rs 6 lakh to distribute among workers. The return, Rajat says, can only be expected after two to three months, given operational costs. "The higher your initial investment, the longer it takes to recover it, especially as the requirements may change within a few months."

Vendor companies typically receive around Rs 250 for an hour of usable video. From that, workers are paid between Rs 50 and Rs 150. The gap is shaped by subcontracting arrangements, device ownership, and negotiation. Workers who do not own iPhones are paid less, with the cost of the device factored out. In factory settings, cuts are shared further with factory owners.

Pooja earns Rs 250 an hour because her brother-in-law co-founded the vendor firm, which pays this rate for household videos. Between June 7 and June 30, she recorded 11.2 hours of usable footage and received Rs 2,800. Most workers sit further down the chain, where margins thin out.

Pooja picks a task from the list, straps the phone to her head and starts recording.

The data aggregation layer. Once recorded and reviewed, the footage moves beyond the worker's control. Only clips that clearly show the task with both hands visible are retained. The videos pass to data-collecting startups — companies like XP Robotics, Objectways, Human Archive, Awign, and EgoData — which aggregate, clean, and label the raw footage, converting it into structured datasets that machines can learn from.

The app Pooja uses, Atlas Capture, is operated by Atlas Capture LLC, a US-based company headquartered in Wyoming. The platform operates in over a hundred cities worldwide; its contributor interface is available in English, Chinese, Vietnamese, Tagalog, Indonesian, and Spanish.

Atlas's privacy policy states that data is primarily processed in Canada, may be transferred to other countries, and that video data may be retained indefinitely for research purposes, subject to applicable deletion rights. The recordings can contain faces, voices, biometric information, and other personal data.

The policy also puts a specific responsibility on contributors: if a video features other people, the person submitting it represents and warrants that they have obtained the necessary consent for their personal information to be collected, used and disclosed.

Pooja says she agreed to do the work but was not asked to sign any separate consent or privacy document before she began recording. She was given an iPhone with the Atlas app already installed and started using it as instructed.

Decode reached out to Atlas Capture for comment on its consent requirements and how it handles footage featuring people other than the contributor. This story will be updated if and when we receive a response.

Alongside these Indian firms, larger global data networks run by companies like Scale AI and Appen also operate in this space. Apart from collecting data directly, these companies rely on digital platforms — Remotasks, Outlier, Crowdgen — where thousands of gig workers log in to annotate, label, and verify footage.

The end users. At the top of the chain sit the embodied AI companies building humanoid robots. XP Robotics lists clients including Figure AI, 1X Technologies, Sanctuary AI, Agility Robotics, Covariant, Apptronik, Machina Labs, and Dexterity — companies racing to build machines that can function in the unpredictability of the real world.

Figure AI's humanoid, Figure 3, powered by its Helix AI system, is being positioned as a general-purpose worker. 1X Technologies is building Neo, a home robot designed to assist with everyday domestic tasks — making tea, cleaning windows, sweeping floors. The same tasks Pooja performs, over and over, with a camera strapped to her head.

These machines are priced in the range of $20,000 to $100,000 per unit. The embodied AI industry has attracted billions in global investment and is projected to reach approximately $23 billion by 2030.

Decode reached out to Figure AI and 1X Technologies for comment on their labour standards and data-collection operations in India. This story will be updated if and when we receive a response.

The Arbitrage

The footage Pooja produces is called egocentric video in the industry — first-person recordings that capture how hands move, how tasks unfold, and the small adjustments that make routine work possible. Unlike chatbots trained on text, the AI systems behind humanoid robots need to learn from the physical world. They need to watch humans do things.

That demand has turned India into a key supplier. The reason, according to Astha Kapoor, co-founder of the Aapti Institute, which researches the data economy, is an enormous pool of cheap labour that is unable to negotiate.

"We have fewer jobs. The way this work is packaged is attractive to people, particularly women, who would like to work and earn from home," she says.

Kapoor draws a comparison to India's older BPO economy — but with an important difference. The outsourcing model is familiar: do the lower-value jobs here, while the gains are captured elsewhere. But unlike BPO jobs, which were often full-time positions with some degree of structure, much of today's AI data work is gig-based, bringing with it greater job insecurity and poorer quality of work.

The result, she argues, is a form of global arbitrage.

"The labour is being performed in the Global South, while the value is being captured in the Global North."

India sits at both ends of this chain: as labour and as customer. "We do the labour, that labour adds value to the data. Then the data goes elsewhere" — before those products are eventually sold back to Indian consumers. "Our large ecosystem and our large market are both being exploited," Kapoor says.

Even at the bottom of the chain, the work is not without cost to the worker. Pooja says the effort is underestimated. "It's not like I can just go about my household chores and record at the same time. I have to set aside dedicated time for it."

 The footage must show her hands throughout to qualify for payment.

And when footage is rejected — which happens often — the worker absorbs the loss entirely. Labour and public policy researcher Namrata Raju calls this a structural problem.

"When a worker spends 15 to 20 minutes recording a video that is later rejected and goes unpaid, labour has already been contributed, regardless of whether the output is used," she says. "Breaking work into such small, discardable units puts immense pressure on workers and devalues their time."

The work also defies easy classification. It resembles gig work because it is task-based; it is digital labour because the output is data; and because it is done from home, it meets the definition of home-based work. "When work spans categories, workers often slip through the cracks," Raju says.

"They are rarely recognised as formal workers and therefore excluded from basic protections like minimum wage, insurance, or social security."

Much of it, she adds, is being done by women.

The Mismatch

Pooja has nearly exhausted the list of tasks she can perform inside her home.

"The ones that are left, I can't do," she says. "Dishwashers, restocking toilet paper, gardening — we don't have these things here."

The task list on Atlas Capture is designed for the environments in which these robots will operate: suburban homes with dishwashers and garden beds. The workers producing the data live in a different world.

Pooja is trying to close the gap herself. "I am thinking of buying pots and gardening tools so I can record gardening tasks," she says. "It will cost a few thousand rupees, but then I can earn more."

She would be spending her own money to simulate a foreign domestic environment so that a robot in another country can learn to replicate it.

That gap may matter for the data itself. Research on robot learning shows that training data needs variety for robots to generalise to new environments. A study, for instance, found that training across different environments and objects helped robots transfer skills to unfamiliar settings. But if much of the footage is recorded in environments that differ from where the robots will eventually operate, the data may not capture the full range of conditions the robots will encounter.

But the question is not only what these recordings teach the machines. It is also what the people producing them get in return.

Raju says she is concerned about whether such work can offer consistent pay, or evolve to include basic forms of social security. "Without that, the emphasis on job creation risks obscuring the more fundamental issue of whether the work itself is stable or sustainable."

When asked if she knew how her recordings would be used, Pooja says, "To train robots. I saw that on Instagram."

She adds, with a laugh: "If we had a robot to do this work, it would be great. Though I don't know if I would ever be able to afford one."

 A faint red mark remains where the device rested on her forehead.

She unstraps the buckle from her head. A faint red mark runs across her forehead where the device had been resting.

"The phone is a bit heavy," Pooja says, tending to the spot lightly. "But I am used to it now."