AI Tech · Robotics · Explainer

The Humans Training AI to Do Tomorrow’s Jobs — and Why Companies Are Paying Them to Do It

Thousands of people across India, Latin America, and Southeast Asia are wearing head-mounted cameras to record themselves folding laundry, slicing vegetables, and arranging objects. The footage is training the next generation of robots. Here’s why this data is so valuable and what it tells us about where physical AI is headed.

When you ask a humanoid robot to fold a towel, it needs to have seen thousands of humans fold thousands of towels first — from a first-person perspective, with detailed information about how hands move, how fabric responds, and how the task looks different depending on the type of towel, the lighting, the surface, and the person doing it. Generating that data at scale is an unsolved infrastructure problem that a growing industry of human data collection companies is now actively building. The UAE has announced plans to have AI agents run half its government by 2028 — but before AI agents can operate physical tasks autonomously, someone has to teach them how the physical world works. That teaching is happening now, in Tamil Nadu, in São Paulo, and in dozens of other locations around the world, one four-minute clip at a time.

The Financial Times’ documentary on how India is becoming the world’s AI data factory — directly relevant to the egocentric data collection industry covered in this article. Independent FT production, not a Wangdoo commission.

Why Robots Need to See Through Human Eyes

AI chatbots and image generators are trained primarily on text and digital images — data that already exists in enormous quantities online. Physical AI — robots that navigate and manipulate real-world environments — needs something different: it needs to understand how the physical world works from the perspective of a body moving through it.

This is why “egocentric data” — first-person video that captures exactly how human hands interact with objects from the viewpoint of someone performing a task — has become one of the most strategically important categories of AI training data. A robot trained on third-person camera footage learns what tasks look like. A robot trained on egocentric footage learns what it feels like to perform them — the visual perspective, the hand positions, the spatial relationships between objects. The difference matters enormously for a robot that needs to operate in an unpredictable household or warehouse environment rather than a controlled factory setting.

The Companies Building This Infrastructure

Objectways

Objectways was founded in 2018, headquartered in Scottsdale, Arizona, with multiple offices in Tamil Nadu, India — including Karur and Coimbatore. The company grew to over 2,200 employees at its peak and had approximately 1,600 employees as of April 2026 according to independent workforce data. It is SOC 2 Type II and ISO 27001 certified, lists Fortune 500 multinationals as clients, and covers data annotation across text, image, audio, video, and LiDAR. Its physical AI work involves contributors — some working from home, some in dedicated studio spaces — wearing head-mounted cameras or smart glasses to record egocentric video of everyday tasks. Some contributors are also outfitted with motion-sensor bands on their wrists, hands, and legs to capture precise kinetic data alongside video.

AFP photographers documented Objectways’ studio operations in Karur in June 2026: workers filming themselves folding towels in fake, fully furnished apartment rooms, recording up to 90 four-minute clips per day. After several thousand hours of filming, the wallpaper is changed to give clients visual variety across the dataset.

Humyn Labs

Humyn Labs, co-founded by Manish Agarwal and Ishank Gupta and based in Bengaluru, committed $20 million in April 2026 to expand its egocentric data collection operations across India, Southeast Asia, Latin America, and the Middle East. The company currently collects 50% of its data from Latin America, 35% from India, and 15% elsewhere in Asia. Alongside video, it is expanding voice data collection across 33 languages, dialects, and accents to support physical AI systems that need to understand spoken instructions in diverse environments. Humyn Labs is also building Robotics Labs — high-fidelity simulation environments that integrate real-world egocentric data with training frameworks.

Why Companies Are Paying Billions for This Data

The business logic is straightforward: humanoid robots are being positioned as one of the largest commercial opportunities of the next decade. Goldman Sachs projects the humanoid robot market could reach $38 billion by 2035. Figure AI, 1X Technologies, Agility Robotics, Boston Dynamics, and Tesla’s Optimus programme are all competing to build general-purpose robots capable of operating in unstructured environments — factories, warehouses, homes. Every one of them faces the same bottleneck: without diverse, high-quality training data showing how humans perform physical tasks across thousands of real-world settings, the robots cannot generalise beyond a narrow set of pre-programmed scenarios.

Collecting that data through robot demonstrations alone is prohibitively expensive and slow — each robot teleoperation session requires expensive hardware, a trained operator, and produces a limited range of scenarios. Human egocentric video collected at scale, across diverse global environments, for $3 an hour, solves that problem at a fraction of the cost. A single Objectways contributor filming 90 four-minute clips per day produces roughly six hours of training footage — footage that would take many hours of robot teleoperation to replicate, at many times the cost. The economics are the reason the industry exists.

Why This Data Is Strategically Valuable — and Hard to Replicate

The appeal of human egocentric video over synthetic or simulated data comes down to one thing: diversity that would be impossible to generate computationally. A real household in Chennai looks different from a real household in São Paulo — different objects, different lighting, different spatial layouts, different cultural practices around the same tasks. A robot trained on a global dataset of real human environments generalises better to new environments than one trained on simulated data built from a limited set of assumptions about what a kitchen looks like.

The scale required is also significant. NVIDIA’s EgoScale research, published in 2026, pretrained a vision-language-action model on over 20,000 hours of action-labelled human egocentric video and found a log-linear scaling law: each doubling of human egocentric data hours produces a measurable, predictable improvement in downstream robot task performance. The average success rate improvement over a no-pretraining baseline was 54% on a 22-degree-of-freedom robotic hand. Objectways’ studio model — where wallpaper and furnishings are changed after sufficient filming — is a direct operational response to this: AI models need to see the same tasks performed in visually distinct environments to learn robust, generalisable behaviour rather than environment-specific shortcuts.

The Human Impact — Jobs, Pay, and What Happens Next

The egocentric data collection industry sits at an unusual intersection: it is creating income for people who need it, using work that directly trains the systems that may eventually displace similar labour. Both things are true simultaneously, and neither cancels the other out.

The pay in India is approximately 250 rupees (~$3) per hour — higher than many informal sector alternatives available to the same workers. AFP’s documentary coverage of Objectways’ Karur facility found workers including engineering graduates, homemakers, and garment factory employees. For a 21-year-old engineering graduate named Rani N., documented at the facility, the work involves recording up to 90 four-minute clips per day of structured tasks across studio settings and home environments. The studio wallpaper and furnishings are periodically changed to provide clients with visual variety across the dataset.

India’s government think-tank NITI Aayog has flagged a structural gap in how AI’s labour impact is being discussed: most analysis focuses on white-collar, professional work, with “little attention, if any” paid to how AI affects India’s 490 million informal workers — the same workers now forming a significant part of the physical AI data supply chain. The tasks being filmed — household chores, object handling, repetitive manual work — overlap with the categories of work that physical AI systems are being built to perform.

Humyn Labs’ co-founder Manish Agarwal has offered a more optimistic framing — that humans and robots will “work together” rather than one replacing the other, with a future where “a welder in India could be managing a welder-robot in Prague.” The CNBC Inside India newsletter reported his additional point that the market for raw data collection work will eventually saturate as AI models improve, and that India needs to “evolve from collector to converter” — building the skills to design, validate, and improve AI systems rather than just supply data to them. That transition, if it happens, is not automatic. It requires investment in training, in education, and in the same kind of industrial policy focus that NITI Aayog says is currently missing.

For the people currently doing this work, the honest answer to “is this stable?” is: not guaranteed. The work pays today, it is scaling today, and the demand is real. Whether it translates into durable, higher-skill employment for the same communities over the next decade depends on decisions that neither the data collection companies nor their clients have yet made publicly.

Frequently Asked Questions