Humanoid Robots Are Being Trained by $15/Hour Gig Workers in 50 Countries
Teaching a humanoid robot to fold laundry turns out to require roughly the same ingredients as teaching a language model to write: enormous amounts of data, and someone to generate it. For robots, that someone is a gig worker recording themselves doing chores.
The Data Problem
Ken Goldberg, a roboticist at UC Berkeley, puts the scale in context. LLMs were trained on enough text that a human reading it would need 100,000 years to get through. Humanoid robots, Goldberg suggests, may need even more data. The difference is that robot training data can't be scraped from the internet. Someone has to move through a kitchen with a camera strapped to them.
The reasoning follows directly from how LLMs were built. Large amounts of real-world movement data, fed into large models. Same paradigm, harder data collection problem.
Who's Paying and How Much
Micro1, a Palo Alto-based company, has hired thousands of contract workers across more than 50 countries including India, Nigeria, and Argentina. The rate is $15 per hour to record household chores on video. The company says it has collected tens of thousands of hours of footage.
Scale AI has gathered more than 100,000 hours of real-world chore footage. Scale AI, Encord, and DoorDash are all actively recruiting data recorders for humanoid robot training.
Robotics companies are spending over $100 million per year on real-world training data, according to Micro1 CEO Ali Ansari. Investors put more than $6 billion into humanoid robots in 2025. Tesla, Figure AI, and Agility Robotics are among the companies racing to build working humanoids.
Micro1 uses an AI agent named Zara to conduct initial worker interviews and review sample videos. Automating the intake process for data collection work is a reasonable efficiency choice at this scale.
Why Not Just Use Simulation
Virtual simulations can train robots for acrobatics. They struggle to model physics accurately enough for object grasping and manipulation. A simulated kitchen and a real kitchen behave differently in ways that matter for a robot trying to pick up a cup. Hence the gig workers.
The Chinese Approach
In China, workers in dozens of state-owned robot training centers use VR headsets and exoskeletons to demonstrate tasks to humanoid robots. It's the same underlying problem solved with more infrastructure and fewer freelance contractors.
Where This Sits
MIT Technology Review's readers ranked humanoid robots 11th on the 2026 Breakthrough Technologies list. That ranking reflects where the field actually is: technically advancing, practically limited, expensive to scale. The data collection race is real. Whether the resulting robots justify the investment is still an open question.
What's clear is that the path to a working humanoid runs through a lot of people filming themselves doing dishes for $15 an hour.
Source: Technologyreview