Woofun AI reports that a new labor sector has emerged where humans physically demonstrate household chores to train robots, earning 200 yuan daily while the data they generate sells for up to 1000 yuan per hour. This industry relies on former IT professionals, real estate agents, and students who wear motion-capture equipment to fold clothes, make beds, and organize desks, effectively becoming the biological fuel for artificial intelligence development. The core paradox lies in the future vision where robots serve humanity, yet the current reality demands humans bend over to teach machines how to perform these exact same tasks.

The scene inside a typical data collection center involves a person wearing a helmet and gloves equipped with motion cameras, moving slowly to drag three long wires of varying thicknesses across a desk. With one hand, the worker pulls a book to the edge, picks it up, and places it on a shelf, then turns sideways to arrange a pack of wet wipes neatly beside it. At the other end of the wires, a partner sits before a computer, monitoring sensor and camera footage on a screen to adjust the mouse and change perspectives. This partner ensures that the 3D model's movements match those of a real person and that camera signals remain stable. These collectors are gathering data to organize desktops, teaching robots how to tidy up, clean, make beds, and fold clothes. Every action a human performs—what to pick up first, what posture to use, where to grasp an object, how much force to apply—is recorded by cameras and sensors and converted into data. After quality inspection and annotation, this data is used to train robot models. In simple terms, these workers are data collectors and teachers for robots. Just as young children need adults to teach them how to walk and use chopsticks, robots require a large amount of human-action data to learn. Unfortunately, such data is extremely scarce on the Internet. For robots to learn how to fold clothes, wipe tables, open doors, or organize books, someone must demonstrate these actions repeatedly. Initially, the ideal vision for robots was to serve humans, but before they could do so, humans had to bend over and become their fuel.

In a homestay setting, the workforce includes programmers who have been laid off, former real estate agents struggling with mortgage payments, and college students who have come together. They put on helmets and gloves and repeatedly fold quilts, towels, and organize desktops, turning their personal experience into data to sell. Working full-time for a month, they can earn about 6,000 yuan. For them, this becomes a temporary job with an uncertain duration. Li Chenchen, 36 years old with many years of experience in IT operations, started her own business after being laid off in 2024. She lost all her savings and even ended up in debt. Watching the person slowly organizing the desktop in front of her, she seemed to see a metaphor: once robots could do all this, what would humans be left to do? A sense of powerlessness swept through the room, and perhaps everyone there wished that day would never come. One afternoon in June, an application process for positions as robot data collectors at nearly ten companies on a recruitment website took half an hour. Soon, four companies contacted the applicant, and three arranged online interviews for the next day, including one that was a robotics company valued at tens of billions of yuan.

However, only this company offered a full-time position and provided five social insurances and one housing fund through a third-party labor service company. The other two were outsourcing companies, but later it was discovered that they both worked for the same data collection company. In fact, very few companies offer five social insurances and one housing fund; most of these robot data collection positions are available part-time, usually paying 200 yuan per day for eight hours of work. Night shifts, due to the reversed day and night schedule, pay 50 yuan more per day. Wages can be paid weekly or monthly, but if you work for too short a time, a portion of your salary may be deducted.

Becoming a robot data collector is not particularly difficult. You submit your resume, join a recruitment group, attend online interviews, and then undergo on-site trials. The whole process can be completed in as fast as 24 hours. An applicant eventually attended online interviews with two companies and was only asked about height and weight. They passed both interviews and went to the companies that did not require a full-time commitment for on-site trials. This position hardly requires any educational background or experience; the main requirement is physical ability. More than thirty people attended the same video group interview. A freelance food delivery worker, with a round face tanned by the sun, introduced himself enthusiastically. He said he used to be a programmer but started delivering food after getting laid off and still hoped to find a stable job. He wasn't sure exactly what the position was called and hesitantly said he was applying for a doll-catching job. The interviewer calmly corrected him: It's robot data collection. A recent graduate, who was relatively petite, hesitated when answering the interviewer's questions about her height and weight. The interviewer asked her to hold her palm in front of the camera and said, Let's give it a try first. At the end of the interview, only one applicant was rejected because they were too overweight to wear the data collection equipment. For this emerging industry, there is still no consensus on the ideal data collection approach. Currently, the mainstream methods can be roughly divided into three categories: real-person data, simulation data, and remote-controlled real-machine data. Remote-controlled real-machine data involves humans remotely operating robots or using exoskeleton devices to control them in real environments, with sensors on the robots recording the entire process. This type of data is closest to the actual working scenarios of robots in the future and is considered the most valuable, but it is also the most expensive, as it requires both the robot itself and human operation. Currently, this type of data collection is mainly done by robot manufacturers themselves.

Simulation data is generated in virtual environments and does not require a real location or real people. The cost mainly comes from hash rate, and it can be trained in large-scale parallel operations.

However, due to the differences between the virtual and real worlds, it is difficult to completely replicate details such as materials, friction, and lighting. As a result, robots trained in this way may not function well in the real world. Real-person data can be collected in two ways: either by recording videos of human actions, which is the least expensive but provides relatively limited information, or by adding motion capture and sensor tracking to the videos, which allows for the recording of more details and is a more cost-effective option. The position applied for involved collecting real-person data. A set of real-person data collection equipment consists of a cycling helmet with a motion camera, two data collection gloves with built-in sensors, a hand motion camera, multiple locators, and supporting software. Together, these cost about 100,000 yuan. The employer told us that this equipment was currently applying for a patent. Before starting formal work, workers had to undergo three days of training and on-site trials. On the first day, the project manager and team leader checked each person's hand conditions one by one. The data collection gloves come in a standard size, and hands that are too long, too short, too fat, or too soft are not suitable. More than forty candidates sat in a row, holding their hands out for inspection. After the check, four or five people were eliminated. Li Chenchen was also on the verge of being rejected. Her little finger was too short, and when she put on the gloves, the sensors would bunch up at her joints, making it impossible for the software to accurately record her movements. She looked up at the team leader, pleading for another chance. The team leader nodded, but that was just the first hurdle.

On the second day, when it came to actual operations, half of the candidates were gone. Li Chenchen and the applicant were assigned to the same group. She adjusted the software while the applicant put on the equipment. First, the applicant put on the helmet to ensure it was securely fixed; then they put on a pair of disposable sweat-proof gloves, followed by the data collection gloves with sensors, and finally a layer of knitted gloves to prevent signal interference. Three long data cables extended from the gloves and helmet and were secured around the waist with an elastic band. Next, the applicant had to keep hands flat in front of the chest while the software calibrated. Li Chenchen sat in front of the computer, adjusting the parameters of the virtual hand model on the screen, while the team leader watched and gave guidance. Ten minutes passed, but the model hand still wasn't in the right position. The team leader got impatient, took over the mouse, made a few adjustments, and said, That's it. Next person. Li Chenchen helped remove the equipment. It wasn't hot that day, and the air conditioner was on in the room, but she was already sweating profusely. I can't learn it, she said in a low voice. On the third day of the on-site trial, Li Chenchen didn't show up, so the team leader assigned a new partner, a recent graduate majoring in nursing. That day, the work location was a homestay with two bedrooms, a living room, a kitchen, and a bathroom. Li Chenchen and the applicant were responsible for making the bed and folding towels, while another group collected and organized desktop data. Some colleagues were assigned to work in places like board game cafes and kitchens—the specific locations and tasks depended on the data needs of the robotics companies.

Workers were required to move in a way that resembled robots—slowly and with minimal finger movement. It was a process of going against instincts. At first, the applicant tried to complete the tasks as efficiently as possible, just like doing household chores. But the team leader told them to slow down and move more naturally, even though the movements had to be slow. So they had to tighten muscles in the waist and hips, pick up the towel, unfold it, smooth it out, fold it, press it flat, straighten the quilt, tuck in the corners, and smooth out the wrinkles—all these movements had to be slow, complete, and continuous. Don't swing the towel or shake the quilt, the team leader added. Since there were no cameras or sensors on the forearms, robots couldn't understand these actions, so they were prohibited. Workers were also required to flexibly change the placement of objects and the way they organized them. Sometimes the towel was placed above the quilt, sometimes it was tucked between the pillow seams. Sometimes they had to lift one corner of the pillow with one hand, and other times they had to pick it up with both hands. This was done to diversify the types of data collected. Before starting work, the team leader told them that the work location was a homestay, so it was convenient to use the bathroom. But in reality, putting on and taking off the equipment and adjusting it often took at least fifteen minutes. Each trip to the bathroom would waste nearly half an hour for two people, and missing even one minute of data collection could affect performance evaluation. Although there was no penalty for collecting too little data, they would only receive a reward of 50 yuan if they collected 5 hours, or 18,000 seconds, of effective data per day. Time here was measured in seconds. There are 86,400 seconds in a day, and an 8-hour shift equals 28,800 seconds. As novices, they needed to collect about 9,000 seconds of effective data per day, but just putting on and adjusting the equipment took more than 1,000 seconds, and they already felt tired by then.

To prevent the camera on the head from shaking during movements, the applicant had to tighten the adjustment strap of the helmet as much as possible, which made it feel like a tight-fitting collar around the head. The disposable sweat-proof gloves created a hot and humid environment. After just 2,000 seconds of data collection, both the gloves and hands were wet and wrinkled. By evening, the applicant couldn't remember how many times they had folded quilts or towels. Shoulders and neck ached from carrying the heavy helmet, and the back was stiff from bending over for so long. Before robots could learn to work in the same way, the human had already become like them. During that day of on-site trials, the equipment that was applying for a patent kept breaking down. Sometimes the locators lost connection frequently, sometimes the sensors deformed and could not be calibrated, and different hand shapes also caused discrepancies in the mapping results. An operations staff member ran back and forth between several buildings, constantly restarting and adjusting the equipment, with sweat running down his forehead non-stop. Since the equipment was newly developed and there was no standard operating procedure, everything had to be adjusted manually. He told the applicant that just half a month ago, he was still working as a video editor, and he had had to learn about equipment maintenance on the spot. It's been two hours, and the data collection gloves still haven't connected, a member of the next group said helplessly. He stood still, holding his arms out to assist with the adjustment, moving his shoulders occasionally to relieve the tension before continuing to wait. Nearly half of the eight-hour workday was spent adjusting the equipment. Everyone hoped that the equipment would return to normal as soon as possible. There were only 24 sets of equipment in total, and they were the most expensive assets in the entire area. To make the most efficient use of these equipment, the company arranged shifts during the day and at night, with four collectors rotating among each set of equipment. Every minute the equipment was idle meant one less minute of data could be collected.

In the field of embodied intelligence, data collected through human operations, along with visual and sensory information, is extremely valuable. According to reports from The Paper, the current market price for embodied intelligence data ranges from 200 to 500 yuan per hour, and some real-machine data collected in real-world scenarios can cost up to 1,000 yuan per hour. Theoretically, a group of robot data collectors working 8 hours a day could generate data worth up to 1,600 to 8,000 yuan.

However, the term effective data essentially means that only certain parts of the data are considered useful. During an 8-hour shift, if the video footage is lost, the movement path is poorly designed, the operations are repetitive, or the camera captures a person's face, the data is considered invalid and must be discarded and started over. Skilled collectors can generate 4 to 5 hours of effective data per day, while novices usually manage only 2 to 3 hours. Before these data can be sold on the market, they must go through quality inspection, cleaning, and annotation processes, and a significant amount is lost in these steps. As a result, far fewer data sets can be sold at the desired price. Even after significant discounts, the data is still valuable.

However, it is the data itself that is valuable, not the people who produce it. The labor service companies told workers that the daily wage for this position was 200 yuan during the day and 250 yuan at night. But the employers actually paid the labor service companies 300 yuan per person per day, which was considered quite generous. From the workers' daily wage of 200 yuan to the data's minimum price of 200 yuan per hour, various intermediaries such as labor service companies and data service providers took a cut, leaving the collectors at the bottom of the value chain.

Woofun AI data shows that this disparity highlights a structural inefficiency where the physical labor required to generate high-value AI training data is compensated at a fraction of the market value of the output. The industry relies on a transient workforce of displaced professionals who provide the essential biological input for a technological future they may not share. This marks a critical juncture where the economic model of embodied intelligence depends entirely on the exploitation of human physicality to simulate the very automation intended to replace it.