This is the untold story behind the humanoid robot boom—a vast, low-cost data pipeline built on the labor of thousands of people who often have no idea who will ultimately watch their recordings.

How It Works: Humans as Data Collectors

Unlike AI models that learn from text, robots need to see how human hands actually move, grip, and interact with objects in the real world. Computer simulations alone cannot teach a robot to fold laundry without destroying it or to navigate a cluttered kitchen . To bridge that gap, companies are deploying a new kind of gig:

• Micro1, a Palo Alto-based startup, has recruited roughly 4,000 "robotics generalists" across 71 countries. Workers mount iPhones to their heads using a strap and film choreographed tasks like cooking, cleaning, gardening, and pet care. Micro1 pays about $15/hour in some markets (Nigeria, the U.S.) and collectively its workers submit over 160,000 hours of video per month .
• Objectways has also recruited thousands of workers in more than 50 countries, using both wearable cameras and motion-tracking gloves .
• Instawork launched a system called Instacore in June 2026—a wearable device with five body-mounted cameras connected to a compute backpack—turning its existing gig workers into robot-training data harvesters on job sites like commercial kitchens and warehouses .
• Human Archive, a Y Combinator-backed startup, pays Indian workers to wear camera-equipped caps and tactile gloves, with over 1,000 active sensor headsets deployed in home services, hotels, and restaurants .
• World Context, a Pakistani startup founded in 2025, records factory workers using proprietary sensors and head-mounted cameras, selling industrial training datasets to U.S. and Chinese robotics companies .
• DoorDash launched DoorDash Tasks in March 2026, an app that pays 8 million U.S.-based Dashers to record daily chores for robotics training data .

Who Buys the Data: Tesla, Figure AI, and Agility Robotics

The first-person footage collected by gig workers is used to train humanoid robots from Tesla (Optimus), Figure AI, and Agility Robotics, among others . Figure AI in particular has ramped up production dramatically, with CEO Brett Adcock sharing that the company manufactured approximately 150 units in April 2026—a surge made possible by real-world human demonstration data .

The Numbers Are Staggering

The scale of the data pipeline is enormous and growing:

• Micro1 alone: ~4,000 contributors, 71 countries, 160,000+ hours of video per month .
• JD.com (China): Working with the city of Suqian to generate 10 million hours of robotics training data over two years .
• Global market: The humanoid robot foundation model training data infrastructure market was valued at $1.5 billion in 2025 and is projected to reach $62.8 billion by 2034 (a CAGR of 56%) .
• Investment: Over $6 billion flowed into humanoid robotics in 2025 (Tesla Optimus, Figure AI, 1X, Agility Robotics), creating almost insatiable demand for real-world training data .
• Venture estimates: Bessemer Venture Partners projects aggregate robotic data costs will exceed $3 billion by 2030 .

The Pay Divide: $15/Hour vs. $2.60/Hour

A stark pay disparity exists within this global workforce:

• Nigeria / U.S. market: Workers are paid roughly $15/hour (e.g., Micro1 contributors recording with iPhones) .
• India: Workers are paid approximately 250 rupees/hour, or about $2.60/hour—roughly one-sixth of what counterparts in Nigeria or the U.S. make for the same type of work .

Privacy and Labor Concerns

The rapid growth of this industry has raised serious ethical questions:

• Lack of transparency: Many workers are not told which companies will ultimately use their footage. Micro1 sells datasets to humanoid robotics developers, but individual workers often do not know the end buyer .
• Capturing family members: Workers film in their own homes, meaning spouses, children, and private living spaces are frequently recorded. The footage is then stored and used for AI training with limited clarity on retention or deletion policies .
• Informed consent gaps: Critics argue that gig workers in lower-income countries may not fully understand how their movement data will be used—or that they are training robots that could eventually displace similar human labor .

The Global Race: China's State-Backed Data Centers

The demand for embodied AI data is not limited to private companies. China has made it a national priority, building state-backed infrastructure at scale:

• The Beijing Innovation Center of Humanoid Robotics operates an embodied AI robot data and training base with dozens of simulated real-world scenarios, capable of generating more than 500 hours of data per day .
• In May 2026, the National Pilot Base for Embodied AI Applications was inaugurated, training over 140 robots in 40+ job-oriented scenarios including kitchens, warehouses, and farms .
• A dedicated Humanoid Robot Data Training Center in Beijing's Shijingshan district, with a budget of $30 million, began training 100 humanoid robots in October 2025 .
• Nearly 100 embodied AI data collection centers have been built across Chinese cities, backed by local state capital and tech conglomerates, making them some of the most active robot buyers on the market today .

The Bottom Line

The practice of using low-cost global gig workers as "human sensors" has become the backbone of humanoid robot training in 2026. While it generates enormous datasets at a fraction of the cost of traditional teleoperation, it raises uncomfortable questions about labor exploitation, informed consent, and privacy—particularly as workers in India earn a fraction of what their counterparts in other countries earn for identical work, often without knowing who will watch their recordings.

This is the untold story behind the humanoid robot boom—a vast, low-cost data pipeline built on the labor of thousands of people who often have no idea who will ultimately watch their recordings.

How It Works: Humans as Data Collectors

Unlike AI models that learn from text, robots need to see how human hands actually move, grip, and interact with objects in the real world. Computer simulations alone cannot teach a robot to fold laundry without destroying it or to navigate a cluttered kitchen . To bridge that gap, companies are deploying a new kind of gig:

• Micro1, a Palo Alto-based startup, has recruited roughly 4,000 "robotics generalists" across 71 countries. Workers mount iPhones to their heads using a strap and film choreographed tasks like cooking, cleaning, gardening, and pet care. Micro1 pays about $15/hour in some markets (Nigeria, the U.S.) and collectively its workers submit over 160,000 hours of video per month .
• Objectways has also recruited thousands of workers in more than 50 countries, using both wearable cameras and motion-tracking gloves .
• Instawork launched a system called Instacore in June 2026—a wearable device with five body-mounted cameras connected to a compute backpack—turning its existing gig workers into robot-training data harvesters on job sites like commercial kitchens and warehouses .
• Human Archive, a Y Combinator-backed startup, pays Indian workers to wear camera-equipped caps and tactile gloves, with over 1,000 active sensor headsets deployed in home services, hotels, and restaurants .
• World Context, a Pakistani startup founded in 2025, records factory workers using proprietary sensors and head-mounted cameras, selling industrial training datasets to U.S. and Chinese robotics companies .
• DoorDash launched DoorDash Tasks in March 2026, an app that pays 8 million U.S.-based Dashers to record daily chores for robotics training data .

Who Buys the Data: Tesla, Figure AI, and Agility Robotics

The first-person footage collected by gig workers is used to train humanoid robots from Tesla (Optimus), Figure AI, and Agility Robotics, among others . Figure AI in particular has ramped up production dramatically, with CEO Brett Adcock sharing that the company manufactured approximately 150 units in April 2026—a surge made possible by real-world human demonstration data .

The Numbers Are Staggering

The scale of the data pipeline is enormous and growing:

• Micro1 alone: ~4,000 contributors, 71 countries, 160,000+ hours of video per month .
• JD.com (China): Working with the city of Suqian to generate 10 million hours of robotics training data over two years .
• Global market: The humanoid robot foundation model training data infrastructure market was valued at $1.5 billion in 2025 and is projected to reach $62.8 billion by 2034 (a CAGR of 56%) .
• Investment: Over $6 billion flowed into humanoid robotics in 2025 (Tesla Optimus, Figure AI, 1X, Agility Robotics), creating almost insatiable demand for real-world training data .
• Venture estimates: Bessemer Venture Partners projects aggregate robotic data costs will exceed $3 billion by 2030 .

The Pay Divide: $15/Hour vs. $2.60/Hour

A stark pay disparity exists within this global workforce:

• Nigeria / U.S. market: Workers are paid roughly $15/hour (e.g., Micro1 contributors recording with iPhones) .
• India: Workers are paid approximately 250 rupees/hour, or about $2.60/hour—roughly one-sixth of what counterparts in Nigeria or the U.S. make for the same type of work .

Privacy and Labor Concerns

The rapid growth of this industry has raised serious ethical questions:

• Lack of transparency: Many workers are not told which companies will ultimately use their footage. Micro1 sells datasets to humanoid robotics developers, but individual workers often do not know the end buyer .
• Capturing family members: Workers film in their own homes, meaning spouses, children, and private living spaces are frequently recorded. The footage is then stored and used for AI training with limited clarity on retention or deletion policies .
• Informed consent gaps: Critics argue that gig workers in lower-income countries may not fully understand how their movement data will be used—or that they are training robots that could eventually displace similar human labor .

The Global Race: China's State-Backed Data Centers

The demand for embodied AI data is not limited to private companies. China has made it a national priority, building state-backed infrastructure at scale:

• The Beijing Innovation Center of Humanoid Robotics operates an embodied AI robot data and training base with dozens of simulated real-world scenarios, capable of generating more than 500 hours of data per day .
• In May 2026, the National Pilot Base for Embodied AI Applications was inaugurated, training over 140 robots in 40+ job-oriented scenarios including kitchens, warehouses, and farms .
• A dedicated Humanoid Robot Data Training Center in Beijing's Shijingshan district, with a budget of $30 million, began training 100 humanoid robots in October 2025 .
• Nearly 100 embodied AI data collection centers have been built across Chinese cities, backed by local state capital and tech conglomerates, making them some of the most active robot buyers on the market today .

The Bottom Line

The practice of using low-cost global gig workers as "human sensors" has become the backbone of humanoid robot training in 2026. While it generates enormous datasets at a fraction of the cost of traditional teleoperation, it raises uncomfortable questions about labor exploitation, informed consent, and privacy—particularly as workers in India earn a fraction of what their counterparts in other countries earn for identical work, often without knowing who will watch their recordings.