Inside the race to train robots on human work

One Friday evening this summer, a personal chef prepared a three-course meal — burrata salad, baked salmon, and tiramisu — for two allies and me.

The meal was free, but there was a catch. A camera mounted on Chef Aliser Demir's ballgame cap was record everything.

Demir came to my flat through Shift, the data-collection arm of the German startup Microagi. The startup offers liberated personal chefs in San Francisco in toggle for first-person footage of their work. Shift offered to pay Demir, who runs a personal chef business, his customary charge of $100 to $150 per person.

"I've been feeding group since I started this job," Demir said. "Now I am additionally feeding AI."

Demir is part of a earth attempt to collect data on how humans engage alongside the bodily world, alongside the goal of construction machines that can personnel factories, spotless homes, and even combat wars. Elon Musk, who is betting Tesla's forthcoming on the humanoid robot Optimus, has promoted a utopian forthcoming of "sustainable abundance," in which AI and robots create individual labor mostly optional.

For now, those ambitions visage a problem: a enormous deficit of training data. "It is extensively accepted in the robotics earth that data is the key bottleneck," stated Kyle Chan, a chap at the Brookings Institution who studies innovation and manufacturing policy.

This is the third installment of the AI Data Grab, Business Insider's sequence on how individual activity is becoming raw matter for AI — and the companies that are cashing in. So far, we've reported on how AI labs are purchasing up employees' emails and Slack messages, and how law firms are turning decades of individual ability into AI systems. Now, the scramble is moving into the bodily world, anywhere companies are paying group to document skills robots have yet to learn.

Contract workers in Mexico and the Philippines remotely control robots in American warehouses as they group attractiveness products, record movements so the machines can eventually do the activity on their own. Other workers across the world strap cameras to their foreheads to grasp first-person footage of themselves folding laundry and making inn beds.

"This is how the awesome cognition transfer happens from humans to robots," stated Josh Gruenstein, CEO of Tutor Intelligence. His business operates a Massachusetts "data factory" anywhere workers remotely authority robots to create training data.

A booming data market

Demand for robot training data is approximately 10 times higher than a twelvemonth ago, Vineeth Velmurugan, who leads robot learning at the AI data startup Encord, said. Encord's customers, including Dyna Robotics and Agility Robotics, formerly requested 1,000 hours of annotated data. Now they desire 10,000 or equal 100,000, he said.

That red-hot petition for real-world data comes as robotics has rotate into one of Silicon Valley's hottest bets. Humanoid robot makers Figure AI and Agility Robotics have fetched multibillion-dollar valuations, as have the companies evolving robot "brains" specified as Skild AI and Physical Intelligence.

Robotics companies raised $33.4 milliard globally in the archetypal fractional of this year, according to PitchBook data, surpassing the $28.7 milliard raised in all of 2025. Cheaper hardware and computing, improved AI models, labor shortages, and competition alongside China are all fueling capitalist interest.

But a gulf remains between robots that execute notable demos and machines dependable adequate to activity an eight-hour factory shift. Robotics companies anticipation additional training data volition help, and suppliers throughout the world, from scrappy startups to established companies akin Scale AI, Instawork, and Uber, are rushing to market it to them.

Custom hardware is key to record the bodily world. Mecka AI, which has raised additional than $120 myriad from investors including Sequoia and Nvidia, started alongside an app that paid group to document mundane activities two years ago. But the business shortly ran into the limits of average video.

Mecka now designs headgear and sensor-equipped gloves to "sensorize humans," CEO Josh Gao said. The equipment records movements and force, specified as the force needed to grip a pane or expert an egg. Contractors tag office footage before Mecka sells it to AI labs and robotics companies.

Microagi, the genitor business of Shift, is expanding its gathering efforts into homes and workplaces throughout additional than 15 countries. It pays additional than 35,000 group to document themselves performing bodily tasks, CEO Bercan Kilic said. Shift stated participants can acquire up to $30 an hr and that it has paid out about $10 myriad in 2026.

Household chores are the most common, but different tasks assistance disclose models to unfamiliar situations. One contributor in Turkey records himself repairing old jet engines in his garage, Kilic said.

Kilic's goal is to collect adequate data to build robots that can activity in factories and location 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 informed group quit this decade, and their replacements were never born. Reshoring lone plant if the robots do."

Income today, automation tomorrow

For the workers supplying that data, though, the trade-off is earnings today in toggle for assisting build machines that could eventually automate their work.

Demir, the personal chef, stated he believes his activity is safe. "Machines can fair watch, listen, and prosecute instructions," he said. "They cannot awareness or smell."

The economics are distinct in countries anywhere labor is cheaper. Human Archive, a data gathering business based in San Francisco, began construction operations in India final twelvemonth through partnerships alongside gig platforms, hotels, and plumbing and electric businesses. It started alongside an app that let group document audio and first-person video. The startup now additionally makes tradition hardware in China, including camera-equipped headsets and wrist trackers that grasp movement.

Today, it operates in additional 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 acquire "almost as much, or additional [than their regular jobs], to fair deteriorate these devices."

Mecka and Human Archive did not say how they justify workers' consent whenever collecting data through partner businesses or local contractors. Shift stated involvement is voluntary, and several workforce at partner businesses choose not to participate. The business stated it obtains definitive consent, explaining what data is collected and how it is used, including that it may be licensed to another labs for AI and robotics training.

The earth competition for data activity can arrive alongside a downside. It drives downward wages and leaves workers alongside small bargaining power, stated Mark Graham, a prof at the Oxford Internet Institute and author of Feeding the Machine, a publish concerning the labor rearward AI. "Often the workers don't equal cognize what they're doing," he said. "They're a extremely available appendage of a much bigger machine."

The activity can additionally be physically taxing. One controller who remotely controls robots in US warehouses from a Texas facility previously told Business Insider that eight-hour shifts logic arm and high arm cramps. "We equal have a massage gun in the office," they said.

Will the data surge last?

The companies collecting all this data are betting that the quest to train intelligent robots volition energy petition for years to come. Research resolute Mordor Intelligence projects that the broader AI training dataset market volition develop from $11.9 milliard this twelvemonth to nearly $49.8 milliard by 2031.

For robotics data suppliers, the masculine livestock case is that startups volition expend a important portion of their backing on training material, stated Siddharth Ramakrishnan, a partner at Scale Venture Partners.

"If scaling laws continue to work, the pie volition fair keep growing," he said.

But nobody knows how much data robots volition need to activity reliably, he added. If they need small than expected, data suppliers could endure business. If they need much more, investors could tire of paying before the robots rotate into commercially viable.

Some are betting that petition 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 activity in homes — among the hardest environments since it is so unstructured — could necessitate a dataset five times the size of YouTube. At the end of 2025, YouTube held an estimated 29 milliard videos, the equal of 280,000 years of footage, according to research resolute Omdia.

"It's difficult to say how this volition perform out," Ken Goldberg, a UC Berkeley robotics prof and cofounder of storage robotics startup Ambi Robotics, said. "It power obtain five years or 50 years. And equal if we had unlimited data now, I don't cognize if this method volition work," he said.

Training robots in virtual environments, known as simulation, could additionally alter the equation. World Labs, a startup slated for an $8.2 milliard acquisition by AMD, says it can rotate a small set of recordings into simulated environments and modify them to create additional training data for robots. If the innovation delivers, robotics companies could need far small data from the bodily world.

But equal sustained petition doesn't justify a profitable business.

Earlier this year, Sam Padilla close downward his robotics data startup, Eidon AI, following struggling alongside logistics, upfront costs, and lean margins. He compared the endeavor to dropshipping, anywhere retailers market products shipped immediately by suppliers and pocket the markup — until competitors offering the identical products for small compression their margins.

Most popular