explainx / blog
Thousands of Indian workers wear head-mounted cameras for $2.60/hr to record egocentric data that trains humanoid robots. Here is what is actually happening and why it matters.
Jun 20, 2026
Figure AI CEO Brett Adcock confirmed robots now outnumber humans at the company. Here's what the headcount chart, 1/hour production rate, and BMW deployment numbers actually mean.
Jun 25, 2026
Is AI conscious? The honest answer is: we do not know β and the uncertainty is not a gap in AI research, it is a gap in philosophy. Here is the full map: from Chalmers's hard problem to the Chinese Room, from Integrated Information Theory to Anthropic's model welfare team.
Jun 25, 2026
AI bias is not a glitch β it is a systematic pattern of skewed outputs baked into a model through its training data, design choices, or the way outputs are used. It can cause hiring tools to screen out qualified candidates, lending algorithms to deny loans by zip code, and facial recognition to fail on darker skin tones at higher rates. Understanding the types, causes, and mitigation approaches is now a core skill for anyone building or procuring AI systems.
In a garment factory in Tamil Nadu, a worker folds a shirt. Mounted on her head is a small camera recording her every movement β the exact angle of her wrist, the grip pressure of her fingers, the millisecond timing of each fold. Across town, a 25-year-old named Nagireddy Sriramyachandra slices mangoes in her kitchen wearing the same device. The camera occasionally barks "hands not detected" when she drifts out of frame.
Both women are earning βΉ250 an hour β about $2.60. Both are feeding a global pipeline building humanoid robots. And both may be training the systems that eventually replace them.
This is not a thought experiment. It is happening at scale across India right now, bankrolled by Silicon Valley venture capital and serving robotics labs linked to OpenAI, Nvidia, and Google.
The core technical problem is this: humanoid robots cannot learn physical tasks from text or images alone. A language model trained on internet data can describe how to fold a shirt in detail. A robot arm cannot replicate the fold without seeing it from a first-person physical perspective β with depth data, grip force, wrist angle, and timing all captured simultaneously.
The solution the industry has settled on is egocentric data: first-person video footage of humans doing physical tasks, recorded from the perspective of the person doing the action. Workers wear devices fitted with RGB-D cameras β sensors that capture color, depth, and motion data in real time. The recordings feed into Large Behaviour Models (LBMs), the physical-world equivalent of large language models, which learn to map human movements into robot instructions.
The reason India has become the center of this collection effort is straightforward: it has a massive, low-wage informal workforce, loose regulatory oversight of data collection, and deep existing infrastructure for tech outsourcing.
The highest-profile player is Human Archive, a Y Combinator-backed startup that raised an $8.2 million seed round from Wing Venture Capital, NVP Capital, and angels from OpenAI, Nvidia, Google, and Meta. The company has deployed over 1,000 camera headsets across workers in residential homes, restaurants, hotels, construction sites, logistics facilities, and industrial environments globally β with a significant portion of operations in India.
Co-founder Raj Patel has described the company's goal plainly: "Our technology will become foundational infrastructure for automating manual labor, increasing global abundance." Co-founder Rushil Agarwal says cameras are attached to workers across every environment where physical work happens.
Egolab.AI, founded in January 2026, calls itself "India's largest first-person POV Data Aggregator." The startup collects egocentric footage from garment factory workers at suppliers including Pearl Global, using lightweight cameras with minimal setup. Its pitch to clients is high-volume, labor-sourced egocentric data at competitive rates.
Objectways is an AI data company operating in both the US and India, with Fortune 500 companies listed as clients. The company pays household workers and textile workers βΉ250/hour to mount sensory devices and record task footage. Objectways runs structured data collection operations where workers are recruited through local networks and trained briefly on how to keep their hands in frame.
Pronto, a Bengaluru-based home services platform, is running pilot programs with outward-facing body cameras on service workers β positioning data collection as an add-on to existing gig work.
The recorded tasks cluster around physical activities that are economically significant to automate:
| Category | Example Tasks |
|---|---|
| Household | Folding clothes, mopping floors, washing dishes, coffee preparation |
| Food preparation | Slicing produce, sandwich assembly, cooking specific items |
| Garment factory | Stitching, folding, stacking, quality inspection |
| General logistics | Arranging blocks, picking and placing objects |
| Artisan | Making flower garlands, handcraft assembly |
These tasks are not chosen randomly. They map directly to the environments where humanoid robots from companies like Figure AI, Agility Robotics, and 1X Technologies are currently being piloted: factories, warehouses, and domestic service settings.
The picture on informed consent is mixed and, in some cases, troubling.
Some workers understand the context. Nagireddy Sriramyachandra told Al Jazeera: "Who else will give you 250 rupees an hour just for doing housework?" She acknowledged she might get a robot herself someday. But a 55-year-old worker named Ponni put it differently: "The next generation... who might have to do work similar to mine, they will face a problem."
India's IT Ministry has reportedly raised questions about whether workers are being adequately informed that recordings will be used to train automation systems targeting their own job categories. Critics note that a brief onboarding session about "keeping hands in frame" is not the same as informed consent about the downstream use of the data.
The situation is further complicated by the gig economy structure: workers are often recruited through platforms and local agents, several layers removed from the AI company ultimately purchasing the data.
India has been a global hub for data labeling β the task of tagging images, transcribing audio, and classifying text to train AI models β for over a decade. This work is often criticized for low pay and poor conditions, but it operates in a relatively understood ethical framework: workers annotate existing data.
The camera-wearing arrangement is structurally different:
Aditi Surie, a digital labor researcher who tracks India's informal AI workforce, has noted that data collection services targeting physical workers will likely expand significantly before any regulatory framework catches up.
Projections for the humanoid robot market estimate over 1 billion robots deployed by 2050. The primary deployment targets are factories, logistics, and domestic service β sectors dominated by informal workers in countries like India.
India's 490 million informal workers, identified by government think tank NITI Aayog, represent the population most exposed to physical automation. NITI Aayog has explicitly noted that most public discussion of AI and employment focuses on white-collar knowledge work, without examining how automation affects the informal sector that forms the backbone of the Indian economy.
The dynamic is cyclical in the worst way: informal workers need income, so they take camera-wearing gig work. That gig work creates training data. The training data improves robot models. Improved robot models reduce demand for the tasks those same workers do for their primary income. And eventually, once the models are reliable, demand for the camera-wearing gig work itself disappears too.
The AI training pipeline consumes itself.
Companies in this space have largely avoided the "replacement" framing. Human Archive positions its work as building "foundational infrastructure for increasing global abundance." Egolab.AI frames its model as empowering workers with new income streams. Objectways emphasizes its Fortune 500 client relationships and the supplemental income it provides.
The implicit argument is that technology always creates new job categories to replace what it eliminates β a claim with some historical backing but significant uncertainty at the scale and speed of current AI development.
The counter-argument from labor researchers is that previous waves of automation replaced specific physical or cognitive tasks while leaving a large residual of human work. Humanoid robots with reliable LBMs are designed to replace workers, not just tasks β a qualitatively different disruption.
Several practical interventions would reduce the harm without stopping useful research:
Informed consent that actually informs. Workers should receive clear, translated explanations that footage will train automation systems targeting their job categories β not just instructions on camera operation.
Residual compensation structures. A small royalty or one-time payment when a trained model is commercially deployed would align worker interests with the value they created.
Regulatory framework from India's IT Ministry. The Ministry's existing questions about the practice should translate into disclosure requirements and data use restrictions analogous to GDPR's rules on biometric and behavioral data.
Industry transparency. Companies raising venture capital on the strength of Indian labor datasets should disclose collection methods and worker compensation in investor materials.
The AI job displacement debate has largely played out in the context of software engineers and knowledge workers. The workers wearing cameras in Tamil Nadu and Hyderabad represent a different and more immediate form of exposure β one that deserves the same level of attention and policy urgency as the 97,000 AI-related job cuts already recorded in tech in 2026 alone.
Understanding what roles will survive automation β and building the skills that make that survival more likely β is the practical challenge. explainx.ai's forward-deployed engineer guide and the AI survival framework both address this directly.
External sources:
Details on company funding, worker compensation, and deployment scale are accurate as of June 2026. The humanoid robotics and data collection landscape is evolving rapidly; specific company operations may have changed since publication.
