Your Next Coworker Might Be Trained By Someone Folding Laundry In Nigeria
So here's a weird one. Companies racing to build humanoid robots have hit a wall that has nothing to do with motors or batteries. They don't have enough footage of humans doing normal stuff. Cooking dinner, folding towels, wiping a counter the boring parts of your day turn out to be exactly what these machines are starving for.
To be fair, this isn't some far-off problem. It's happening right now, in two very different ways: massive government-backed training centers in China, and a scattered, global gig economy of people strapping phones to their heads.
Why Robots Can't Just "Read" Their Way To Competence
Here's the thing about language models like ChatGPT they got smart by scraping text that already existed on the internet in huge quantities. Humanoid robots don't have that shortcut. They need synchronized data on joint angles, gripper forces, camera frames, and task context, recorded while a physical task actually happens. That kind of footage just doesn't exist at internet scale.
Berkeley roboticist Ken Goldberg has a name for this gap he calls it something like the 100,000-year problem, because the amount of physical-world data robots may need dwarfs what even massive text models required, since controlling joints is more complicated than generating words. That's why suddenly there's real money for anyone willing to record themselves loading a dishwasher.
The Gig Economy Angle: Cameras On Foreheads
This is the part that sounds made up but isn't. Companies like Micro1 have recruited around 4,000 workers across 71 countries, paying roughly $15 an hour to strap iPhones to their foreheads and record household chores. The company is pulling in over 160,000 hours of footage every month, and its own VP admits that's nowhere close to enough they're reportedly aiming for billions of hours eventually
DoorDash jumped in too. In March 2026 it launched a standalone app called Tasks, letting its 8 million US Dashers earn extra cash by filming themselves folding laundry or making beds though notably, it skipped states with stricter privacy laws. Scale AI has been building out its own recorder army as well, and by some estimates has already banked around 100,000 hours of robotics-specific footage.
If you've ever done freelance data-labeling work, this feels like the physical-world version of it. Same grind, same "is this actually being used responsibly" energy, except now the task is literally cooking your own dinner on camera.
Pay isn't equal either. US households can earn triple what workers in India or Vietnam make for the same chore, because robotics companies assume American buyers will be first in line for these machines, so American kitchen data is worth more. Some researchers are already calling this "data colonialism" everyday life getting turned into raw material for company training runs.
The Other Approach: Teleoperation Warehouses
Meanwhile, China is going a totally different route building physical training facilities at serious scale. The Sichuan Humanoid Robot Multimodal Data Collection and Testing Center in Zigong opened in January 2026 and covers 6,000 square meters, with rows of Walker S2 robots being trained by human operators wearing VR headsets who mirror their own movements onto the robot in real time.
This is teleoperation basically puppeting a robot remotely so every motion gets logged as training data. China's 15th Five-Year Plan has folded this kind of embodied AI training infrastructure directly into national policy, and there are reportedly dozens more centers like this one planned or already running.
US companies do their own version too. Figure AI and Tesla run their own teleoperation fleets, with operators in VR gear remotely controlling robots so every action gets recorded as a demonstration. It's slower and more expensive per hour than gig-recorded phone footage, but the data tends to be cleaner because it's already robot-shaped.
Why This Actually Matters
Honestly, this is the real bottleneck holding robots back from your kitchen not the hardware. Current imitation-learning methods need 50 to 200 demonstrations per single task, meaning a facility with 20 different tasks needs thousands of demonstrations just to get started. That math is why every robotics company suddenly cares more about video pipelines than servo motors.
There's a real privacy question buried in all this too. Workers are recording their own homes, sometimes with family visible in the background, and have limited visibility into how that footage gets stored or shared once it's sold to robotics firms. Not exactly a solved problem yet.
What's Next
Nobody officially knows how much data is actually enough. Some in the industry think AI-converted YouTube footage could eventually cut into this whole gig market, but that's speculative for now, not confirmed. For the time being, the fastest path to a humanoid robot that can fold your laundry runs through someone else, somewhere else in the world, folding theirs.
Would you record your own chores for $15 an hour if it meant training the robot that might replace that chore someday?
