Upon returning to his studio apartment from a lengthy shift at the hospital, Zeus, a medical student residing in a hilltop city in central Nigeria, activates his ring light, secures his iPhone to his forehead, and commences recording himself. He lifts his hands in front of him like a somnambulist and lays a sheet on his bed. He moves deliberately to ensure his hands remain within the camera’s view.
Zeus serves as a data recorder for Micro1, a United States-based firm headquartered in Palo Alto, California that aggregates real-world data for sale to robotics firms. As corporations like Tesla, Figure AI, and Agility Robotics strive to create humanoids—robots aimed at mimicking human appearance and motion in factories and homes—videos captured by gig economy workers such as Zeus are emerging as a vital new method for training them.
Micro1 has enlisted thousands of freelance workers across more than 50 nations, including India, Nigeria, and Argentina, where numerous tech-savvy youths seek employment. They’re attaching iPhones to their heads and filming themselves performing tasks like folding laundry, washing dishes, and cooking. The compensation is attractive by local standards and is stimulating local economies, yet it prompts complex issues regarding privacy and informed consent. Moreover, the job can be tough and peculiar at times.
Zeus discovered this opportunity in November as conversations about it proliferated on LinkedIn and YouTube. “This seemed like a prime chance to leave a mark and contribute data that would be leveraged for robot training in the future,” he reflected.
Zeus earns $15 per hour, which is a substantial income in Nigeria’s challenging economy characterized by high unemployment rates. However, as an eager student aspiring to be a doctor, he finds the monotony of ironing clothes for hours each day uninteresting.
“I really [do] not like it so much,” he admits. “I’m the kind of individual who requires … a hands-on technical job that necessitates critical thought.”
Zeus, along with all workers interviewed by MIT Technology Review, requested to be identified solely by pseudonyms due to a lack of authorization to discuss their roles.
Constructing humanoid robots is notoriously complex because mastering the manipulation of physical objects is a challenging skill. However, the advent of large language models powering chatbots like ChatGPT has prompted a transformative shift in the field of robotics. Similar to how large language models learned to produce text by being trained on extensive datasets from the internet, many researchers hold that humanoid robots can learn to engage with their environment through training on vast movement datasets.
Editor’s note: In a recent survey, readers of MIT Technology Review identified humanoid robots as the 11th breakthrough for our 2026 compilation of 10 Breakthrough Technologies.
However, robotics demands far more intricate data regarding the physical world, and acquiring this data is considerably more challenging. Virtual simulations can train robots to execute acrobatic maneuvers, but not how to grasp and manipulate objects since simulations struggle to accurately replicate physics. For robots to function in factories and act as domestic helpers, real-world data, despite being labor-intensive and costly to gather, might be essential.
Investors are zealously channeling funds into addressing this challenge, investing over $6 billion into humanoid robots in 2025. Simultaneously, at-home data recording is evolving into a thriving gig economy globally. Data companies like Scale AI and Encord are amassing their own teams of data recorders, while DoorDash compensates delivery drivers for filming themselves completing household tasks. In China, workers at numerous state-operated robot training centers don virtual-reality headsets and exoskeletons to demonstrate to humanoid robots how to open microwaves and wipe tables.
“There is significant demand, and it’s escalating rapidly,” states Ali Ansari, CEO of Micro1. He approximates that robotics firms are currently spending upwards of $100 million yearly to procure real-world data from his company and similar entities.
A day in the life
Workers at Micro1 undergo vetting by an AI agent named Zara, which conducts interviews and assesses chore video samples. Each week, they submit footage of themselves performing chores at home, adhering to a set of guidelines regarding visibility of their hands and natural movement speed. These videos are evaluated by both AI and human reviewers, resulting in acceptance or rejection. Subsequently, they are annotated by AI and a multitude of human labelers who identify the actions portrayed.
“There is significant demand, and it’s escalating rapidly.”
Ali Ansari, CEO of Micro1
Given that this approach to robot training is still in its nascent stages, it remains uncertain what constitutes effective training data. Nonetheless, “you need to provide numerous variations for the robot to generalize effectively for basic navigation and manipulation in the environment,” asserts Ansari.
Many workers, however, express that producing a variety of “chore content” within their limited living spaces is a struggle. Zeus, a resourceful student in a modest studio, finds it difficult to record tasks beyond the daily ironing of his clothes. Arjun, a tutor in Delhi, India, requires an hour to produce a 15-minute video due to the extensive time spent brainstorming new chores.
“How much content [can be created] at home? How much content?” he queries.
Additionally, there’s the delicate issue of privacy. Micro1 instructs workers to keep their faces out of view of the camera and to refrain from disclosing personal details such as names, phone numbers, and birth dates. They then utilize AI and human reviewers to eliminate any unintended disclosures.
Even in the absence of faces, the videos reveal an intimate glimpse into the lives of the workers: the interiors of their residences, their belongings, their daily routines. Recognizing what personal data might be inadvertently recorded while they engage in household chores on camera can be complex. Reviews of such footage may not effectively screen out sensitive information beyond the most evident identifiers.
For workers with families, keeping their private lives off-camera necessitates constant negotiation. Arjun, a father of two daughters, grapples with his energetic two-year-old trying to stay out of frame. “Sometimes it’s quite challenging to work because my daughter is so little,” he remarks.
Sasha, a former banker turned data recorder in Nigeria, tiptoes while hanging her laundry outdoors in a communal residential area to avoid capturing her neighbors, who watch her in confusion.
“It’s going to take longer than people think.”
Ken Goldberg, UC Berkeley
While the workers interviewed by MIT Technology Review grasp that their data is utilized to train robots, none are aware of the precise ways their information will be used, stored, or shared with third parties, including the robotics firms purchasing the data from Micro1. For reasons of confidentiality, Micro1 doesn’t disclose its clients or the specific nature of the projects to the workers.
“It is crucial that if workers are partaking in this, they are informed by the companies themselves about the purpose … where this technology might lead and how it could impact them in the long run,” asserts Yasmine Kotturi, a professor specializing in human-centered computing at the University of Maryland.
Occasionally, some workers mention having seen others inquire on the company Slack channel about the possibility of deleting their data. Micro1 declined to confirm whether such data is removed.
“People are choosing to engage in this,” asserts Ansari. “They can halt work at any moment.”
Hungry for data
With thousands of workers performing chores in varied households, some roboticists question the reliability of the data collected for safe robot training.
“How we conduct our lives in our homes does not always adhere to safety standards,” remarks Aaron Prather, a roboticist at ASTM International. “If those individuals are imparting bad habits, it could lead to incidents, rendering the data ineffective.” The overwhelming quantity of data collected complicates quality control reviews. Nonetheless, Ansari mentions that the company rejects videos demonstrating unsafe task execution, while clumsy actions can inform robots what to avoid.
Then, there’s the issue of how much data is necessary. Micro1 claims to have accrued tens of thousands of hours of footage, while Scale AI reported it has compiled over 100,000 hours.
“It’s going to require a considerable amount of time to achieve this,” indicates Ken Goldberg, a roboticist at the University of California, Berkeley. Training large language models on text and images would require a human 100,000 years of reading, and humanoid robots may demand even more data, as controlling robotic joints is substantially more intricate than generating text. “It’s going to take longer than people anticipate,” he affirms.
When Dattu, an engineering student residing in a busy tech center in India, arrives home after a full day of university classes, he opts to skip dinner and rushes to his small balcony, cluttered with potted plants and dumbbells. He secures his iPhone to his forehead and films himself repeatedly folding the same set of clothes.
His family watches him with curiosity. “To them, it resembles some kind of space technology,” he remarks. When he discusses his job with friends, “they are often astonished by the thought of being paid for recording chores.”
Balancing university studies along with data recording and other data annotation tasks weighs heavily on him. Still, “it feels as if you’re engaged in something unique compared to the rest of the world,” he states.
