The humanoid robotics industry has been talking about "entering factories" and "entering homes," but rarely explains how these robots actually learn to perform tasks. The answer is not some magical large model, but part-time workers distributed across dozens of countries, filming themselves washing dishes, folding clothes, and scrambling eggs frame by frame with iPhones strapped to their foreheads.
This is the reality uncovered by a MIT Technology Review report in early April.
iPhone on the head, into the kitchen
Here's how it works: an iPhone is strapped to the forehead, pointed at the hands, and workers follow task instructions to perform various household chores—folding towels, scooping food from a pan into a bowl, tidying kitchen counters. Each task typically lasts tens of minutes. After uploading the video, AI and human reviewers jointly label it, marking the start time, end time, and hand position for each action.
The main company behind this is Micro1, based in Palo Alto, California, which employs contract workers in over 50 countries, primarily in Nigeria, India, and Argentina. The hourly wage is $15, which is considered good relative to local income levels.
On the employer side, companies building humanoid robots—Tesla, Figure AI, Agility Robotics—are buying this data. For robots to learn to work in real home environments, simulators are far from sufficient; real human operation data is irreplaceable raw material.
Micro1 is not alone. Scale AI claims to have collected over 100,000 hours of similar data. DoorDash is even paying delivery drivers to film themselves doing household chores. The entire industry spends over $100 million annually on real-world training data.
Not the first time, but the scale is growing
The use of human-labeled data for AI training is nothing new. Early image recognition relied on Amazon Mechanical Turk workers, and NLP relied on extensive text annotation. But humanoid robot training data has a fundamental difference: it requires not classification of text or images, but three-dimensional temporal data of actions, including force, rhythm, and spatial coordinates.
This cannot be solved with purely synthetic data because robots encounter too many variations in friction, weight, and material properties in the real physical world. Scale AI's data increasingly comes from videos recorded at home by specially recruited workers, not from lab professionals.
What does this mean? It means that the capability boundaries of humanoid robots are, to some extent, determined by home kitchens around the world. Whether a robot can fold towels of different materials depends on whether enough operation videos of those materials have been recorded.
$15 an hour: Is this job stable long-term?
Workers have a largely positive view of the job.
"I feel like I'm doing something completely different from the rest of the world." (Participant Zeus)
Another worker named Dattu said he feels he is "leaving a mark," contributing data to the robot's growth. But researchers have questions about the industry's sustainability. UC Berkeley's Ken Goldberg said directly:
"People underestimate how long it takes to collect enough safe and usable robot training data."
A more practical issue is privacy. The content workers film includes not just their hands, but also their home furnishings, family photos, daily habits, and room layouts. Micro1 says it removes faces and sensitive information, but workers themselves do not know which company ultimately receives the data, where it is stored, or whether it might be resold. Some have requested deletion, but no one knows the outcome.
University of Maryland researcher Yasmine Kotturi believes this information asymmetry is a structural problem: workers bear the information risk without corresponding rights to know.
Another concern: Aaron Prather of ASTM International points out that if workers use improper techniques during recording (such as unsafe cutting methods), these "bad examples" could be packaged into training data, teaching robots dangerous habits.
This supply chain is not going away
| Data provider | Data volume | Primary worker source |
|---|---|---|
| Scale AI | 100,000+ hours | Global, distributed |
| Micro1 | Tens of thousands of hours | Nigeria, India, Argentina |
| DoorDash | Not disclosed | Delivery drivers |
For humanoid robots to truly "enter homes," they need to handle an incredible variety of scenarios: furniture of all shapes, clothing of all materials, kitchens of all layouts, countertops of all heights. This is not a problem that can be solved with thousands of hours of data, but with millions of hours.
This globalized, decentralized data supply chain will persist long-term and will only grow larger. Until robots can learn completely autonomously, humans must first demonstrate the tasks to them.
Only, the person doing the demonstration might be in a rented apartment in Nigeria, with an iPhone strapped to their head, washing dishes while helping future robots learn to walk.
Source: CocoLoop, The gig workers who are training humanoid robots at home (MIT Technology Review)