The Robot Training Data Supply Chain: The Ultimate Guide

A woman in Shandong straps a phone to her head and films herself doing chores for six hours a day at 20 yuan an hour. A contractor in Manila steers shelf-restocking robots inside Tokyo convenience stores for about $300 a month. In a garment factory outside Delhi, workers sew with cameras on their heads and are paid nothing extra for the footage. In San Jose, a humanoid company posts jobs for "data creators" at $30 an hour. All four are the same industry: the supply chain that turns human movement into training data for robots.

Language models were trained on text that already existed. Robots have no such archive. The motion data a humanoid needs, how a hand reaches, grips, folds and places, has to be recorded, one demonstration at a time, by people. This guide maps who does that recording, where, for how much, and who buys the result. It draws on the sourced dataset behind the Robot Data Supply Chain tool, where every company, pay figure and event links to the page it came from.

The take: robot data is a labour market before it is a technology. Three supply routes dominate in 2026: teleoperation (a person drives a real robot), egocentric video (a person wears a camera while doing the task), and state-funded training grounds in China that combine both at industrial scale. Pay follows geography far more than skill. Reported robot-data pay outside the US has a median near $3.50 an hour, against about $31 inside it, and the price buyers pay for raw first-person footage has already fallen by more than half in a year. The open questions are consent, who owns the recordings, and whether a worker filming their own job is training their replacement.

Companion reading: robot teleoperation, imitation learning for robotics, and foundation models and VLAs for robotics.

Table of contents

1. Key takeaways
2. Why robots need recorded human data
3. The three supply routes
4. Who collects it: the company map
5. Where the work happens
6. What it pays
7. China's training grounds
8. Southeast Asia and Malaysia
9. The datasets it produces
10. Labour, consent and the open questions
11. Frequently asked questions
12. Sources

Why robots need recorded human data

A vision-language-action model learns to map what a robot sees to what it should do with its body. The internet holds billions of images and words but very little paired data of the form "here is the scene, here is exactly how the arm and fingers moved". That pairing has to be manufactured.

The research datasets that started the field were small and hand-built. Google's RT-1 was trained on 130,000 episodes collected with a fleet of 13 robots over 17 months. The DROID project gathered 76,000 trajectories with 50 data collectors across three continents. The Open X-Embodiment effort pooled more than a million trajectories from 22 robot types. Commercial labs now want orders of magnitude more: Physical Intelligence reports pretraining on more than 10,000 hours of robot data, and Dyna Robotics says its latest model saw over a million hours of egocentric human video.

Hours at that scale cannot come from lab students. They come from a workforce.

The three supply routes

Teleoperation is the most useful data and the most expensive, because every operator needs a robot. Egocentric video is the opposite: a phone strapped to the forehead is enough, and the robot learns to translate human hands into its own body later. Figure says its Project Go-Big navigation model used 100% egocentric human video collected in real homes. Tesla moved Optimus data collection from motion-capture suits to helmet-and-backpack camera rigs. The cheaper route is winning on volume.

A fourth stream sits alongside these: remote operation as a service. Contractors in Manila operate restocking robots in more than 300 FamilyMart and Lawson stores in Tokyo, and the company passes that intervention data to Physical Intelligence. Avatar Robotics runs warehouse robots through teleoperators in Mexico, Colombia and the Philippines. The paid work and the training data are the same session.

Who collects it: the company map

Scale AI says it has gathered more than 100,000 hours of physical-world data and collects more than 1,000 hours of demonstration data a day. Appen told ABC it can collect around 60,000 hours of video a month from up to 5,000 contributors in India, Canada, the US, Brazil, Mexico and the Philippines. Korea's Config runs a data operation of nearly 300 people in Seoul and Hanoi. XDOF, three months out of stealth, reported 20 customers including several frontier AI labs that it cannot name.

That last detail is typical. Most vendors do not disclose who buys their data, so any map of buyer-to-vendor links is partial. The Flow view in the tool shows only links a source names.

Where the work happens

The tool tracks sourced activity in 56 countries. The heaviest are China, the US and India, followed by Malaysia, Kenya, the Philippines, Singapore and Colombia. Coverage follows press attention, so a quiet country may simply be under-reported.

What it pays

Across the 44 robot-data pay points in the dataset, the median outside the US is about $3.50 an hour and the median inside the US is about $31.50. That is roughly a ninefold gap for work that is close to identical: perform a task while a camera or a robot records it.

Measured against local minimum wages from ILOSTAT, almost every robot-data rate sits above the floor, often by a factor of two to five. That is the honest case for the work. The case against it is the margin: in one of the few contracts where both sides are known, OpenAI paid Sama $12.50 an hour per worker while the workers took home $1.32 to $2.

Rates are also falling. A Bengaluru vendor told the Business Times that raw egocentric footage has gone from $6 to $10 an hour of video to $2 to $3, and CNBC reported that collection contract prices halved within months.

China's training grounds

China treats robot data as infrastructure. Rest of World counted more than 40 state-owned data collection centres announced by December 2025; by June 2026 the China Academy of Information and Communications Technology counted more than 70 in operation and 46 more under construction or planned. Shandong alone listed 26 provincial training grounds in one notice.

Inside, trainers wear VR headsets or exoskeletons and repeat a motion on rows of humanoids. A Hubei centre runs nearly 100 human-controlled humanoids folding clothes and wiping tables hundreds of times a day. In Liuzhou, a novice trainer needs about 300 attempts per usable movement.

The money loops back. Robot makers sell humanoids to these centres: UBTech sold 566 million yuan of robots to three data centres. In September 2026 regulators began slowing humanoid IPOs, and one person estimated that some valuations could fall 60% to 70% without data-centre revenue. The state is also standardising the output: the Ministry of Industry and Information Technology approved the first embodied-AI dataset quality standard, effective 1 November 2026, and recognised "embodied intelligence robot technician" as a new occupation.

Southeast Asia and Malaysia

Southeast Asia is becoming the second hub after India. The Business Times described egocentric recording as a new side hustle spreading across the region; Jakarta's MyRobin pays workers up to 3.3 million rupiah a month (about $187) to record five to six hours a day of their own jobs.

Malaysia has moved quickly. MR. ROBOT, the automation arm of MR D.I.Y., runs a data collection centre with 20 AgiBot G2 humanoids operating around the clock, and signed an agreement with staffing network TROOPERS, which offers access to more than 500,000 registered workers, to capture first-person recordings in retail, warehousing and food service. Nasdaq-listed YY Group is building a training and data collection facility in Johor. AgiBot pledged a potential RM1 billion investment in Malaysia by 2035. The Gig Workers Act, in force since April 2026, sets the first legal framework for the platform workers who do much of this recording.

Singapore takes the state route: HTX's new Humanoid Robotics Research Centre has zones for motion capture, teleoperation, human demonstration and simulation.

The datasets it produces

The open datasets are the public face of the field. The proprietary corpora, collected by vendors and gig apps in lower-wage countries, are where most of the hours now sit, and their collection terms are rarely published.

Labour, consent and the open questions

Three issues recur in the reporting. The first is consent and ownership: AI Weekly noted that worker consent, data ownership and revenue sharing for the more than 1,000 Indian gig workers wearing Human Archive headsets are not disclosed. The second is pay disputes: Scale AI agreed a $12.5 million settlement over misclassifying its annotation workforce, and robot startup The Bot Company faces a class action over unpaid overtime. The third is quality and fraud: the egocentric marketplace Kled pulled out of Nigeria after about 95% of uploads proved duplicate or fraudulent.

Underneath all three sits the uncomfortable fact that many recorders are filming the exact work a humanoid is being trained to do. AFP's report from Tamil Nadu carried the quote that sums up the bargain: "Who else will give you 250 rupees an hour just for doing housework?"

Frequently asked questions

What is robot training data? Recordings of a task paired with the actions that accomplish it: camera views plus joint positions, gripper states or human hand poses. A robot policy learns to reproduce the actions from the observations.

What is egocentric video? First-person video recorded from a camera on the head or chest of a person doing a task. It shows the hands and objects from roughly the viewpoint a robot will have, and it is far cheaper to collect than teleoperation because no robot is involved.

Why is so much robot data collected in lower-income countries? Because the work is labour-intensive and pay tracks local wages. A company can pay a recorder in India or Indonesia a fraction of a US rate while still paying above the local minimum wage, and phone-based capture needs no special equipment.

What do robot data collectors earn? In the sourced dataset, robot-data pay runs from about $1 an hour in India to $48 an hour for a US teleoperation role. The median outside the US is about $3.50 an hour and inside the US about $31.50.

What are China's robot training grounds? State-backed facilities where trainers drive rows of humanoid robots through scripted tasks to generate data. More than 70 were operating by mid-2026, and the government has started issuing quality standards for the datasets they produce.

Is this data open? Some research datasets are, such as Open X-Embodiment, DROID, Ego4D and AgiBot World. Most commercial collection is proprietary and sold to robot makers and AI labs under undisclosed terms.

Sources

1. Robot Data Supply Chain tool - the sourced dataset behind this guide, with every quote linked.
2. MIT Technology Review: the humanoid data gig economy - gig workers in Nigeria, India and Argentina filming chores.
3. Rest of World: China's state-funded robot training centres - inside the training grounds.
4. Rest of World: Manila teleoperators running Tokyo store robots - pay and data flow to Physical Intelligence.
5. ABC: Indonesian gig workers recording chores - Appen's collection volumes and contributor pay.
6. The Business Times: why the robot data gold rush is on borrowed time - falling prices and the Southeast Asian side hustle.
7. Time: OpenAI's Kenyan labellers - the vendor margin in a known contract.
8. TechCrunch: Config, the robot data firm backed by Korea's manufacturers - the Hanoi data operation.
9. Xinhua: China's first embodied-AI dataset quality standard - standardisation of training-ground output.
10. Open X-Embodiment and DROID - the open research datasets.