Inside the race to train robots on human work

Tyler Le/BI One Friday evening this summer, a private chef prepared a three-course meal — burrata salad, baked salmon, and tiramisu — for two friends and me. The dinner was free, but there was a catch. A camera mounted on Chef Aliser Demir's baseball cap was recording everything. Demir came to my apartment through Shift, the data-collection arm of the German startup Microagi. The startup offers free private chefs in San Francisco in exchange for first-person footage of their work. Shift offered to pay Demir, who runs a private chef business, his usual rate of $100 to $150 per person. "I've been feeding people since I started this job," Demir said. "Now I am also feeding AI." This story is part of a series called The AI Data Grab, an inside look at how tech giants are increasingly turning to workers, users, and the physical world to get data to train AI. Read more stories in this series here. Demir is part of a global effort to collect data on how humans interact with the physical world, with the goal of building machines that can staff factories, clean homes, and even fight wars. Elon Musk, who is betting Tesla's future on the humanoid robot Optimus, has promoted a utopian future of "sustainable abundance," in which AI and robots make human labor largely optional. Shift pays more than 35,000 people to record physical tasks, from cooking meals to repairing engines. Courtesy of Shift For now, those ambitions face a problem: a massive shortage of training data. "It is widely agreed in the robotics world that data is the key bottleneck," said Kyle Chan, a fellow at the Brookings Institution who studies technology and industrial policy. This is the third installment of the AI Data Grab, Business Insider's series on how human work is becoming raw material for AI — and the companies that are cashing in. So far, we've reported on how AI labs are buying up employees' emails and Slack messages, and how law firms are turning decades of human expertise into AI systems. Now, the scramble is moving into the physical world, where companies are paying people to record skills robots have yet to learn. Shift collects data in more than 15 countries to capture a broad range of tasks and environments. Courtesy of Shift Contract workers in Mexico and the Philippines remotely control robots in American warehouses as they pack beauty products, recording movements so the machines can eventually do the work on their own. Other workers across the world strap cameras to their foreheads to capture first-person footage of themselves folding laundry and making hotel beds. "This is how the great knowledge transfer happens from humans to robots," said Josh Gruenstein, CEO of Tutor Intelligence. His company operates a Massachusetts "data factory" where workers remotely control robots to generate training data. A booming data market A teleoperator at Encord’s Bay Area facility guides a robot through arranging flowers in a vase. Rya Jetha/ BI Demand for robot training data is roughly 10 times higher than a year ago, Vineeth Velmurugan, who leads robot learning at the AI data startup Encord, said. Encord's customers, including Dyna Robotics and Agility Robotics, once requested 1,000 hours of annotated data. Now they want 10,000 or even 100,000, he said. That red-hot demand for real-world data comes as robotics has become one of Silicon Valley's hottest bets. Humanoid robot makers Figure AI and Agility Robotics have fetched multibillion-dollar valuations, as have the companies developing robot "brains" such as Skild AI and Physical Intelligence. Robotics companies raised $33.4 billion globally in the first half of this year, according to PitchBook data, surpassing the $28.7 billion raised in all of 2025. Cheaper hardware and computing, better AI models, labor shortages, and competition with China are all fueling investor interest. But a gulf remains between robots that perform impressive demos and machines reliable enough to work an eight-hour factory shift. Robotics companies hope more training data will help, and suppliers across the world, from scrappy startups to established companies like Scale AI, Instawork, and Uber, are rushing to sell it to them. Custom hardware is key to recording the physical world. Mecka AI, which has raised more than $120 million from investors including Sequoia and Nvidia, started with an app that paid people to record everyday activities two years ago. But the company soon ran into the limits of ordinary video. Mecka now designs headgear and sensor-equipped gloves to "sensorize humans," CEO Josh Gao said. The equipment records movements and force, such as the pressure needed to grip a glass or crack an egg. Contractors label workplace footage before Mecka sells it to AI labs and robotics companies. A teleoperator at Encord’s Bay Area facility controls a robot as it pours motor oil into a container. Rya Jetha/ BI Microagi, the parent company of Shift, is expanding its collection efforts into homes and workplaces across more than 15 countries. It pays more than 35,000 people to record themselves performing physical tasks, CEO Bercan Kilic said. Shift said participants can earn up to $30 an hour and that it has paid out around $10 million in 2026. Household chores are the most common, but unusual tasks help expose models to unfamiliar situations. One contributor in Turkey records himself repairing old jet engines in his garage, Kilic said. Kilic's goal is to gather enough data to build robots that can work in factories and address labor shortages in Europe and the US. "If you run factories, the math is already on your desk," Kilic previously told Business Insider. "Your most experienced people retire this decade, and their replacements were never born. Reshoring only works if the robots do." Income today, automation tomorrow For the workers supplying that data, though, the trade-off is income today in exchange for helping build machines that could eventually automate their work. Demir, the private chef, said he believes his work is safe. "Machines can just watch, listen, and follow instructions," he said. "They cannot feel or smell." The economics are different in countries where labor is cheaper. Human Archive, a data collection company based in San Francisco, began building operations in India last year through partnerships with gig platforms, hotels, and plumbing and electrical businesses. It started with an app that let people record audio and first-person video. The startup now also makes custom hardware in China, including camera-equipped headsets and wrist trackers that capture movement. Today, it operates in more than 30 countries and pays workers between $1 to $10 and "even much more," per hour, depending on how disruptive the equipment is, CEO Raj Patel said. These workers, Patel said, could earn "almost as much, or more [than their regular jobs], to just wear these devices." Mecka and Human Archive did not say how they ensure workers' consent when collecting data through partner businesses or local contractors. Shift said participation is voluntary, and some employees at partner businesses choose not to participate. The company said it obtains explicit consent, explaining what data is collected and how it is used, including that it may be licensed to other labs for AI and robotics training. The global competition for data work can come with a downside. It drives down wages and leaves workers with little bargaining power, said Mark Graham, a professor at the Oxford Internet Institute and author of Feeding the Machine, a book about the labor behind AI. "Often the workers don't even know what they're doing," he said. "They're a very disposable appendage of a much bigger machine." The work can also be physically taxing. One operator who remotely controls robots in US warehouses from a Texas facility previously told Business Insider that eight-hour shifts cause arm and shoulder cramps. "We even have a massage gun in the office," they said. Will the data boom last? The companies collecting all this data are betting that the quest to train intelligent robots will fuel demand for years to come. Research firm Mordor Intelligence projects that the broader AI training dataset market will grow from $11.9 billion this year to nearly $49.8 billion by 2031. For robotics data suppliers, the bull case is that startups will spend a significant share of their funding on training material, said Siddharth Ramakrishnan, a partner at Scale Venture Partners. "If scaling laws continue to work, the pie will just keep growing," he said. But nobody knows how much data robots will need to work reliably, he added. If they need less than expected, data suppliers could lose business. If they need much more, investors could tire of paying before the robots become commercially viable. Some are betting that demand for robot training data is nowhere near its peak. Encord's Velmurugan likens today's robot brain to that of a two-year-old, and estimates that training a robot to work in homes — one of the hardest environments because it is so unstructured — could require a dataset five times the size of YouTube. At the end of 2025, YouTube held an estimated 29 billion videos, the equivalent of 280,000 years of footage, according to research firm Omdia. UC Berkeley roboticist Ken Goldberg. "It's hard to say how this will play out," Ken Goldberg, a UC Berkeley robotics professor and cofounder of warehouse robotics startup Ambi Robotics, said. "It might take five years or 50 years. And even if we had unlimited data now, I don't know if this approach will work," he said. Training robots in virtual environments, known as simulation, could also change the equation. World Labs, a startup slated for an $8.2 billion acquisition by AMD, says it can turn a small set of recordings into simulated environments and modify them to generate more training data for robots. If the technology delivers, robotics companies could need far less data from the physical world. But even sustained demand doesn't guarantee a profitable business. Earlier this year, Sam Padilla shut down his robotics data startup, Eidon AI, after struggling with logistics, upfront costs, and thin margins. He compared the business to dropshipping, where retailers sell goods shipped directly by suppliers and pocket the markup — until competitors offering the same products for less squeeze their margins. "There is a clear arbitrage opportunity to make money if you're willing to race and race to the bottom," he said. "But a technical bottleneck does not necessarily imply a good business." Have a tip? Contact Rya Jetha via email at rjetha@businessinsider.com or Signal at rjetha.07. Use a personal email address and a nonwork device; here's our guide to sharing information securely. Read the original article on Business Insider

Read it at Business Insider ↗

← Front page