The Workers Training the Robots That May Replace Them
Humanoid robots need more than code. They need human movement, judgement and repetition. Around the world, a new data economy is emerging in which workers are filmed, tracked or asked to guide machines so robots can learn how physical labour is done.
What this article covers
This article explains why human labour has become one of the most valuable inputs in robotics, how factory workers and teleoperators are being used to train machines, why “physical AI” needs a different kind of data from chatbots, what ethical questions this raises around consent, compensation and surveillance, and why the next automation wave may depend on the very workers it could eventually displace.
In simple terms
Robots do not learn the physical world the way chatbots learn language. A chatbot can be trained on text, code, images and video found across the internet. A robot that needs to fold a towel, stitch a seam, pick up a tool, sort a package or work safely around people needs something more specific: examples of human bodies doing real tasks in real spaces.
That is why companies are racing to collect what is often called egocentric data: first-person footage of human movements and interactions. The Guardian recently reported on Indian factory workers being asked to wear head-mounted cameras while they stitched garments, folded towels and carried out other manual tasks, with the footage used as part of a growing effort to train AI and robots for industrial work.
The story is not simply that robots may replace workers. The deeper story is that workers are being turned into training data first.
The hidden layer of the robot boom
The public image of robotics is usually polished. A humanoid robot walks across a stage. A robotic arm picks up an object. A warehouse machine moves a box from one place to another. The video looks smooth, inevitable and futuristic. What it rarely shows is the human labour underneath.
Before a machine can act independently, people often have to demonstrate the task, record the task, annotate the task, repeat the task, correct the machine when it fails and guide it through the physical world. The future of robot intelligence is being built from ordinary human actions: hands aligning fabric, workers sorting goods, operators moving robot arms, technicians labelling whether a movement succeeded or failed.
This matters because robotics has a data problem. Large language models had the internet. Robots do not have an equivalent archive of physical experience. There is no simple online dataset containing billions of high-quality examples of hands, tools, bodies, factories, kitchens, warehouses, shops and homes, recorded from the right perspective and labelled in ways machines can use.
That lack of data has become one of the central bottlenecks in “physical AI”, the term increasingly used for AI systems that can operate in the real world. The Guardian reported that industry experts describe data as a major bottleneck in robotics and automation, because robots require first-person recordings of physical work rather than the vast quantities of text used to train chatbots.
The result is a new kind of labour supply chain. Human workers are not only producing garments, sorting boxes or assembling products. In some cases, their movements are being captured, cleaned, annotated and converted into datasets that may help future machines learn to do similar work.
The factory worker as dataset
The Guardian’s reporting from India makes the issue unusually clear. It describes garment workers near Delhi who were asked by management to wear small cameras on their heads during their shifts. One worker, given the name Lalita, said nobody explained why. As she stitched shirts and trousers, the camera recorded the rhythm of her hands, the way she aligned seams, the speed with which she corrected folds and the interactions around her on the factory floor.
At first, according to the report, some workers found the headgear funny. Then the mood changed. Some became more conscious of how they moved. Conversations quietened. Workers worried that every pause, error or distraction might be captured. What they did not know, according to the Guardian, was that their routines were being collected as part of an effort to gather first-hand data from factory floors for the automation of industrial work.
This is where the ethical question sharpens. A worker may be paid for a shift in a garment factory, but if that shift also produces a commercially valuable dataset, what exactly has been sold? The labour? The footage? The bodily skill built up over years? The right to use that skill to train a machine?
The Guardian’s examination of data-collection practices across six factories in five Indian states found that workers wearing devices such as smart glasses and head-mounted cameras received no compensation for generating footage that would later be sold to technology companies. It also reported that none of the seven technology companies it interviewed said they sought consent directly from workers, with some saying permissions were obtained through factory management.
Legally and ethically, those are different things. Permission from an employer is not the same as informed consent from the worker whose body, gestures and working patterns are being recorded. In insecure workplaces, consent can become especially complicated. A worker may technically agree to wear a camera, but if refusing could risk their job or relationship with supervisors, the freedom of that agreement is difficult to assess.
Why robots need human bodies
The reason this data is valuable is simple: physical work is hard.
Humans underestimate the intelligence inside ordinary movement. Picking up a soft object, correcting a folded seam, keeping balance while turning, adjusting grip pressure, finding a tool on a crowded bench, judging whether a material has shifted, reacting to a colleague nearby — these are not trivial actions for machines. They require perception, motor control, context and adaptation.
That is why robotics companies often need human examples. A robot can be programmed to perform a rigid movement inside a controlled environment, but factories, warehouses and homes are full of small variations. Objects move. Lighting changes. Materials bend. People get in the way. Tools are not always where they should be. The real world is messy in ways software struggles to predict.
China’s robotics boom shows the same problem from another angle. In a Guardian feature from March 2026, Leju Robotics was described as using human teleoperators to guide humanoid robots through tasks such as wiping tables, organising cutlery, moving glasses of water, sorting boxes and packing items. One worker guided a robot’s hand around a potato and moved it into a basket. Another logged whether each action succeeded or failed, creating data that could later be fed into a vision-language-action model.
That is an important phrase: vision-language-action model. It describes a system that does not only recognise objects or understand language, but connects what it sees and understands to physical action. The ambition is to create robots that can respond to changing environments more like humans do. But to get there, they need examples of action, not just words.
The new data factories
The companies building this future increasingly talk about data collection as an industrial operation in its own right. Business Insider reported that Apptronik, a humanoid robotics startup backed by investors including Google DeepMind and Mercedes, has opened a nearly 90,000-square-foot “Robot Park” in Austin where Apollo robots practise tasks such as loading boxes and sorting toys. Most of the time, the robots are remotely controlled, with data from the sessions used to improve the AI models that act as the robots’ brains.
Apptronik’s cofounder and CEO Jeff Cardenas described the facility as a “data factory” for generating the kind of data the company needs. That phrase captures the shift. The factory of the future may not only manufacture products. It may manufacture training experience for machines.
The same logic is attracting startups whose main business is not building robots, but producing the data and infrastructure that robot builders need. TechCrunch reported in June 2026 that XDOF had raised $70 million to build data pipelines, collection tools and annotation systems for frontier labs and robotics companies. The company said it was already working with 20 customers, including several frontier AI labs, though it could not name them.
Another TechCrunch report described Human Archive as one of the startups betting that India’s gig economy can help train physical AI, noting that companies racing to build physical AI require large amounts of data showing humans at work.
These businesses sit in an increasingly important part of the AI supply chain. They are not always visible to the public, but their work may determine how quickly robots move from scripted demos into real workplaces.
The OpenAI signal
OpenAI’s public hiring also shows how central robotics data has become. The company recently advertised for a Technical Program Manager for Robotics Data Acquisition, describing the role as owning and scaling systems that power robotic data acquisition across development and evaluation environments, with workflows producing high-quality data for model training and evaluation.
Another OpenAI role for a Robotics Software Engineer says the company is expanding its robotics data collection and evaluation programme across diverse robotics hardware.
Job postings are not product launches, and they should not be overread. But they are useful signals. They show where major AI labs believe hard problems remain. In robotics, one of those hard problems is not only the model. It is the acquisition, evaluation and organisation of physical-world data.
That marks a difference from the chatbot era. The public became familiar with arguments about whether AI companies had used copyrighted books, scraped websites or online content to train language models. Robotics raises a more bodily version of the same question. If a person’s hand movements, working habits and accumulated practical skill become training material, who should control that data? Who should be paid for it? And what should companies be required to explain?
Surveillance disguised as training
There is another concern: once workers wear cameras, the footage can be used for more than training robots.
The Guardian reported that in some factories, footage was also used to generate productivity reports, including rankings based on active working time, estimates of idle periods and tracking of how much time workers spent talking to colleagues. It also raised privacy concerns, including accounts of women garment workers forgetting they were wearing cameras when going to the washroom.
This is where robot training overlaps with workplace surveillance. A camera worn to capture movement data can also become a management tool. It can record pace, mistakes, pauses, informal conversations and bodily behaviour. Even if the stated purpose is AI training, the presence of recording devices can alter the social atmosphere of a workplace.
For companies, this creates a serious legal and reputational risk. Workers should know what is being recorded, why it is being recorded, who will receive it, how long it will be stored, how it may be used, whether it can be sold, whether it will be anonymised, whether it will be used for performance management, and whether refusing has consequences.
Without that clarity, “data collection” can become a soft phrase for extraction.
What is really being captured?
It is tempting to think of these recordings as simple video. But they can contain something more valuable: tacit knowledge.
Tacit knowledge is the kind of expertise people build through experience but may struggle to explain in words. A skilled garment worker knows how fabric behaves under pressure. A warehouse worker knows how to balance speed and care. A machinist knows when a movement feels wrong. A construction worker knows how material weight shifts in the hand. These skills are learned through repetition, failure, correction and muscle memory.
When those movements are filmed, labelled and fed into machine-learning systems, the worker’s expertise becomes detachable. It can leave the body, leave the factory, leave the country and enter a commercial AI system.
The Guardian quoted Sarayu Natarajan of the Aapti Institute saying that this kind of data originates in a worker’s body and actions, but once extracted it no longer remains attached to them in the same way. The article also noted that workers are typically paid for their time, not for the long-term value generated from the datasets built on their labour.
That is the heart of the issue. A worker may be paid once for the shift. The dataset may be sold many times. The robot trained on that data may eventually compete with the category of labour from which the data came.
The replacement question
The strongest version of this story is not “robots are coming for every job tomorrow”. That is too crude. Many robots remain limited, fragile, expensive and dependent on highly controlled environments. Even in advanced factories, human dexterity and judgement remain difficult to replace.
The better question is narrower and more uncomfortable: are some workers being asked to help build the systems that could reduce demand for their own labour later?
The answer appears to be yes in at least some contexts, though the timeline and scale remain uncertain. Robotics companies and their customers are pursuing machines that can perform factory, warehouse, logistics, service and domestic tasks. Some of those systems may support human workers by taking on dangerous, repetitive or physically demanding work. Others may reduce the need for human labour in certain roles.
The China robotics report makes this tension visible. It described entrepreneurs aiming to automate final assembly in car factories, while also showing how much work remains difficult for robots. One automation founder estimated that 80% of final assembly had yet to be automated, but said this was what he was targeting.
The realistic picture is mixed. Robots may create new jobs in teleoperation, maintenance, supervision and data collection. They may also reduce or reshape existing jobs. The question is whether workers whose skills train these systems will share in the value created, or whether they will be treated as temporary inputs in a technology supply chain.
Why this matters
This story matters because it changes how we think about AI labour.
The first phase of the AI labour debate focused heavily on hidden digital work: content moderation, data labelling, rating chatbot responses and annotating online material. The robotics boom extends that debate into the physical world. The body becomes the dataset. Movement becomes intellectual property. A working shift becomes training material.
It also matters because robotics is likely to be one of the next great frontiers of AI investment. If models become able to connect language, vision and action more reliably, the commercial prize is enormous. Factories, warehouses, care settings, retail spaces, homes and hospitals all become potential deployment environments. But the road to that future runs through human demonstration.
The companies that build robots will talk about autonomy. Investors will talk about productivity. Governments will talk about competitiveness. Consumers will see polished videos. But underneath all of that, thousands or millions of people may be asked to perform, repeat, record, annotate and guide the behaviours that make the machines look intelligent.
The ethical test is not whether robots should exist. The test is whether the people training them are treated as informed participants, fairly compensated contributors and rights-bearing workers, rather than invisible raw material.
What we know and what remains unclear
What we know is that first-person human activity data has become valuable in robotics, that companies are building data pipelines around it, and that some workers have been recorded without receiving direct compensation for the later commercial value of the footage, according to Guardian reporting. We also know that robotics companies and AI labs are investing in systems for data acquisition, teleoperation and robot evaluation.
What remains unclear is how widespread the least transparent practices are, how different companies handle consent, how much workers understand about downstream use, and whether regulators will treat bodily movement data as ordinary workplace data or something more sensitive.
It is also unclear how quickly humanoid robots will become commercially useful at scale. Some companies are moving quickly into pilots and controlled deployments, but physical-world automation remains harder than software automation. Robots can look impressive in clips while still struggling with reliability, cost, safety and generalisation.
The biggest unknown is whether a new compensation model will emerge. If workers generate data that produces long-term commercial value, should they receive a one-off payment, a higher hourly wage, a data royalty, collective bargaining rights over data use, or the ability to opt out entirely? Current labour arrangements were not designed for a world in which a person’s movements can be converted into machine intelligence.
What happens next
The next signal to watch is regulation. Governments already regulate workplace surveillance, privacy, biometric data and labour rights, but egocentric robot-training data may fall between categories. It is not only video. It is not only workplace monitoring. It is not exactly biometric data in the traditional sense. It is embodied skill captured for automation.
The second signal is corporate transparency. AI and robotics companies will face growing pressure to explain where their physical training data comes from, how it was collected, whether workers consented directly, how they were compensated and whether the data can be resold or reused for other models.
The third signal is the rise of robot training jobs. Teleoperators, data collectors, robot supervisors and evaluation workers may become a visible new labour category. Some may be well-paid technical roles. Others may be repetitive, outsourced and poorly protected. The quality of these jobs will depend on how the industry is structured.
The fourth signal is deployment. As robots move from demonstrations into factories, warehouses and service environments, the link between the workers who trained them and the workers affected by them will become harder to ignore.
The robot future may arrive with polished hardware and ambitious claims of autonomy. But before robots can work like humans, humans first have to show them how work is done.
Key takeaways
Robotics has a data problem: robots need examples of physical work, not just text and images from the internet.
First-person recordings of human movement, known as egocentric data, are becoming valuable training material for physical AI.
The Guardian has reported that some Indian factory workers were asked to wear cameras while working, with footage used in data pipelines for robot training.
Some workers reportedly received no direct compensation for generating footage later sold to technology companies.
Robot training can overlap with workplace surveillance if footage is also used to assess productivity, track idle time or monitor behaviour.
Teleoperation and robot-training centres are emerging as a new labour category, with humans guiding robots through repeated tasks to generate training data.
The central ethical question is whether workers are being fairly informed, protected and compensated when their bodily skill becomes commercial AI training data.
FAQs
What is egocentric data in robotics?
Egocentric data is first-person data captured from a person’s point of view, often using head-mounted cameras, smart glasses or body-worn devices. In robotics, it can show how humans use their hands, move through spaces, handle tools and complete physical tasks.
Why do robots need human training data?
Robots need human training data because physical work is complex. A robot must learn how objects move, how hands grip, how tools are used, how materials behave and how people adapt to changing real-world conditions.
Are factory workers being used to train robots?
In some cases, yes. The Guardian reported that workers in Indian factories were asked to wear cameras while performing tasks, with the footage collected as part of robotics and AI data pipelines.
Are workers paid for robot training data?
The answer varies. In some factory settings investigated by the Guardian, workers reportedly received no direct compensation for generating footage later sold to technology companies. Some informal workers outside factories were paid directly for recording their activities, but were not always told exactly how the footage would be used.
What is physical AI?
Physical AI refers to AI systems designed to operate in the physical world, often through robots. Unlike chatbots, which mainly process language or images, physical AI must connect perception, reasoning and action.
What is teleoperation?
Teleoperation is when a human controls a robot remotely or guides its movements. In robotics training, teleoperators can generate examples that help models learn how to perform tasks more autonomously later.
Will robots replace factory workers?
Some robots may replace or reduce certain repetitive, dangerous or physically demanding jobs over time. However, many real-world tasks remain difficult to automate, and robots are likely to reshape work unevenly rather than replace all workers at once.
What are the main ethical concerns?
The main concerns are informed consent, compensation, privacy, surveillance, data ownership and whether workers have any say in how recordings of their skills and movements are used commercially.