When garment worker Lalita* was first given a head-mounted camera by factory supervisors, her reaction was laughter. “Just like they put a CCTV camera on the wall, they put one on us,” she remarked. The 32-year-old had been employed at a garment factory on the outskirts of Delhi for nearly a year when management instructed the workers on her production line to wear these small cameras as they began their shifts, without any explanation.
As Lalita sewed shirts and trousers, the camera captured everything: the rhythm of her sewing, the precision of her work aligning collars, and even her interactions with co-workers. “At first, we all found it amusing to see each other in that headgear,” she noted.
However, the light-hearted atmosphere soon shifted. Fearing they were being monitored for productivity, workers grew more conscious of their actions. Conversations that would typically fill the workspace dwindled, and many became more focused on their tasks, anxious that any error or distraction would be recorded.
What Lalita and her peers were unaware of was that their daily activities were part of a broader initiative by businesses in India to gather direct data from factory operations—data that is increasingly crucial for the push towards automation in industrial work.
These firsthand recordings, termed egocentric data, are essential for training the robots that may eventually assume human roles on production lines. The rise of humanoid robots signifies the latest advancement in artificial intelligence, with industry experts noting that data scarcity poses a significant obstacle in the field of automation. Unlike expansive language models like ChatGPT, which draw from an enormous pool of online text, robots need meticulously recorded videos of physical tasks.
Companies engaged in data collection emphasize that the future could demand hundreds of millions, if not billions, of hours of human activities recorded from factories, warehouses, retail outlets, and even homes, establishing a reliable pathway for robots to navigate real-world settings.
EgoLab, a data aggregation enterprise operating in Lalita’s Gurugram factory, counts Tesla among its major clients. Tesla's CEO, Elon Musk, anticipates that a staggering 80% of the company’s future value will stem not from electric cars but from its humanoid robotics initiatives.
India is rapidly establishing itself as a central player in the global race to gather egocentric data. In response to the demand, a growing network of companies like Humyn AI, FPV Labs, Micro1, Egodata, Neocambrian, XP Robotics, Objectways, Scale AI, and CynLr is forming to create robust data pipelines for robotics firms.
According to Puneet Jindal, the founder of Labellerr AI, “South Asia is the workshop of the world for many labor-intensive sectors. If you're training a robot to understand human work, few places can match India’s scale, variety, and density of labor. Daily, millions are engaged in sewing, assembly, sorting, and other tasks that robotics companies want to automate.”
Capturing the data is only the initial phase; the recordings must then be cleaned and annotated for clients, ensuring visibility of hands, accurate tracking of movements, and isolating actions from background activities. India currently dominates the data annotation market, accounting for around 35% of it, with about 60% of the revenue sourced from American clients.
Cost also plays a significant role in this equation. As noted by an anonymous technology firm founder, “A company spending $30 per hour for data collection in the U.S. can often obtain similar services in India for less than a sixth of that price.” This often results in firms establishing arrangements with factories to gather extensive footage without compensating workers directly.
Investigations conducted by The Guardian into data collection practices across multiple factories revealed that workers utilizing devices such as smart glasses or head-mounted cameras received no payment for the footage that would eventually be sold to tech companies. “Sometimes they offer us a soft drink,” Lalita commented, referencing her monthly earnings of around $200. “I can't tell if it’s for the footage collection or just because of the heat in Delhi.”
When queried on the lack of compensation for workers generating valuable datasets, several companies asserted that factories receive payment for facilitating the recordings, negating the need for additional payments to individual workers. Critics argue this reasoning obscures the identity of those truly producing the data.
Geeta Thatra, a researcher with the Bengaluru-based Work Fair and Free Foundation, expressed concerns about privacy and surveillance, emphasizing, “Workers may seem to consent to wearing cameras, but can they genuinely refuse without fearing repercussions for their jobs?”
None of the tech companies interviewed indicated they sought consent from the workers directly, with several claiming that permissions were obtained from factory management.
The reach of egocentric data collection is expanding beyond traditional factories. Increasingly, tech companies are engaging informal workers—like construction laborers, delivery personnel, and street vendors—to document their daily activities. In contrast to factory environments, where payments usually go through employers, informal workers often receive direct compensation via local contractors collaborating with these tech firms.
Munazir*, a mason in Bengaluru, recently began recording his work and earns between $30 to $40 per week from this side job, significantly more than his usual daily earnings. He remarked, “The phone feels heavy and uncomfortable to wear. But it’s a new experience; maybe I’ll adjust.”
While Munazir participates voluntarily, he remains largely uninformed about the future use of the footage he records. “It brings in some extra money, but what they do with the data after that is a mystery to me,” he admitted. Companies recognize that workers are often not informed about the ultimate applications of the data they produce.
Madhumita Dutta, a researcher at Ohio State University studying AI and labor, noted, "Traditionally, workers exchange their labor for wages. In this scenario, they're also generating a valuable digital asset." If they lack awareness of how their skills and movements are transformed into sellable datasets, they have little leverage to negotiate compensation or object to its usage.
For Sarayu Natarajan, founder of the Aapti Institute in Bengaluru, the debate transcends consent and compensation. These recordings encapsulate workers' embodied knowledge—the techniques and skills accrued over years—yet once converted to datasets, this knowledge is detached from the original worker.
“It originates from the worker's actions but ceases to be connected to them in the same manner,” Natarajan explained. This raises challenging questions surrounding ownership and fair compensation that existing labor frameworks struggle to address. Typically, workers may only receive payment for their time, not for the enduring worth their data generates. As companies leverage this information to enhance their AI systems, policymakers might need to explore new strategies for value-sharing that recognize worker contributions beyond their immediate paychecks.
Back in the factory, Lalita continues her stitching, unaware of the data generated from her labor now existing in another realm, cleaned and transformed. When asked about whether workers should receive a portion of the value created from the datasets produced by their work, she replied with a laugh, “We’re not even getting our full worth for what we do now. Who will pay us once robots take over our jobs?”
*Names have been changed.
