Woofun AI reports that a dimly lit homestay in China has become a hub for a new form of labor where individuals wearing head-mounted iPhones and wrist grippers perform repetitive tasks to train robots. Inside these rooms, groups of three to four people fold clothes, assemble building blocks, and craft paper airplanes, repeating each action at least 20 times to generate training data. The workforce comprises new mothers, unemployed couples, and students seeking temporary income, all recruited through intermediaries promising a flexible, high-paying home-based job. After a brief training session in these facilities, workers take the equipment home to officially begin what recruitment platforms describe as an easy opportunity to "teach a robot" with a potential daily wage of 200 yuan.

The reality of this virtual data collection economy diverges sharply from the marketing promises made by third-party intermediaries. Ten days into the work, the actual pay slip revealed earnings of less than 1,000 yuan, averaging under 100 yuan per day. When calculated on an hourly basis, the rate dropped to less than 20 yuan per hour, a figure barely exceeding the income generated by shaking milk tea bottles for promotional campaigns. This discrepancy highlights a critical dynamic in the current wave of capital flowing into virtual intelligence: a new economic fuel has emerged, but the value distribution is heavily skewed. The equipment provided for this work typically includes three iPhones, two grippers, and a helmet, all of which must be worn simultaneously on the head and wrists. A young trainer explained that the robots, which move on wheels rather than legs, require dozens of three-minute videos for each task to learn basic actions, comparing the process to teaching a child to say "dad" and "mom." Despite the technical explanation, most workers remained unaware of the specific entities they were serving or the ultimate destination of the data they collected. Questions regarding the necessity of such high repetition rates and the precise calculation of pay were met with a vague summary from the trainer: earnings depend on effective working time and task diversity. While the trainer cited an anecdote of a team member earning 5,000 to 6,000 yuan in less than half a month by working six hours daily, this narrative proved to be an outlier rather than the norm for the majority of participants.

The structural mechanics of payment reveal why the advertised earnings are rarely achieved. Longer working hours do not automatically translate to higher income; the critical variable is "effective working time," which is calculated on a per-minute basis under flexible employment agreements. Generally, after working for 3 to 4 hours, the pay rate per minute drops below one yuan. Only when the working time exceeds this threshold does the rate increase to slightly over one yuan per minute.

However, this calculation is strictly tied to the quality of the recorded footage, which must meet rigorous criteria including footage integrity, correct equipment usage, and the diversity of object arrangements. For instance, when folding clothes, each of the dozens of recorded videos must display different clothing configurations, and the environment must vary, with the trainer specifying that tablecloths should be bright-colored and changed every 10 videos. Movement smoothness is also a factor; actions that are too fast create ghost images, while those that are too slow appear unnatural. Meeting these minimum standards often requires double the time of the base requirement, meaning a task with a 3-hour minimum might demand 6 hours of total effort to complete successfully.

Beyond the complex payment structure, the physical demands of the work contradict the "light physical labor" description found on job websites. The equipment is surprisingly heavy, with iPhones mounted on wrists and heads, alongside grippers and helmets that each weigh nearly as much as two eggs. Recording hundreds of videos while carrying this load for three consecutive hours is physically impossible for most, necessitating breaks after no more than half an hour of continuous work. Many new mothers in the group ended up applying plasters to their hands after just a few days, with some abandoning the job entirely due to severe soreness. The work also requires significant mental effort, as workers must plan daily tasks in advance and submit them for approval, a process far removed from simple household chores. Hardware limitations restrict mobility, making tasks like washing dishes prohibited due to the inflexibility of the grippers.

Furthermore, task requirements change unpredictably; simple sorting and folding tasks were banned after three days and replaced with more difficult assignments like sticking stickers or peeling oranges. Some materials for these new tasks had to be purchased at the worker's own expense, leading one colleague to spend money on children's paint brushes, sticker sheets, and cardboard, ultimately concluding that the job was harder than factory work and resulted in a net financial loss.

This phenomenon represents a recurring pattern in technological evolution where a new "fuel" drives economic activity. During the rapid internet era, the fuel was low-skilled reviewers; in the era of large models, it was low-skilled data annotators; and now, in the era of virtual intelligence, it is virtual data collectors working for robots. These collectors are employed in numerous data collection centers across the country and are expanding into thousands of households. An intermediary from a Beijing labor outsourcing company noted that virtual data collection jobs in Beijing and Shenzhen are as common as factory work. On platforms like Boss Zhipin, salaries for virtual data collectors range from 4,000 to 6,000 yuan per month for full-time positions, which typically require a bachelor's degree.

However, the majority of these roles are part-time with flexible agreements, managed by intermediaries who often do not know the final end-users due to the complex chain of third-party data service providers. Many of these intermediaries previously worked in autonomous driving and large model data collection, leveraging their existing resources to profit from the booming virtual intelligence market.

The human cost of this data extraction is illustrated by the experiences of workers like Wu Yu, a 38-year-old former middle-level manager at a state-owned manufacturing company. She considered herself one of the first people eliminated by AI when her company adopted large models last year, leading to the closure of her department. After receiving compensation and resting at home, she realized she needed to support her two children and began sending out hundreds of resumes, receiving only a few responses. As her savings dwindled, she turned to virtual data collection for robots as her first part-time job. Despite her hopes to earn extra money, she was the first in her group to quit due to exhaustion and the chaotic nature of the work. New mothers like Wu Yu already faced the burden of household chores, but the constant use of grippers injured her hands within a week, forcing her family to take turns working to meet targets. The opacity of the payment system added to the frustration, as effective working time was not visible to workers until the pay failed to match expectations. This forced most to work an extra hour to ensure they met the required thresholds. Wu Yu noted that workers had to compete on the amount of time spent, a sentiment echoed by the lack of systematic training and varying salary standards provided by the data service provider. The provider's only consistent message was that "the longer you work, the more money you will earn," driven by a demand for "millions of hours of data" and a plan to recruit a team of 1,000 part-time workers across different cities.

Other workers, such as Li Li, who relied on odd jobs like tutoring and dishwashing, found the virtual data collection role unsustainable. Although she had once earned 500 yuan in a single day, she returned the equipment after five days due to frequent malfunctions and the stress of working from home. Li Li emphasized that the uncertainty of the work, with constantly changing tasks and payers, made it less reliable than traditional odd jobs. The ten days spent working for robots felt like being trapped on a new type of assembly line, characterized by mechanical repetition, video length calculations, and advance task planning. The work was driven by data requirements rather than a belief in "teaching robots," turning the promise of flexible home work into a cage with no colleagues to discuss tasks with. The pay, linked to effective working time, offered little motivation because the majority of the value was captured by upstream entities in the supply chain.

Financial data from a virtual data service provider underscores the disparity in value distribution. The current price for real-device data is around 500 to 1,000 yuan per hour, while the data collected by workers without actual device information is priced between 300 and 400 yuan per hour. Simulated data costs between 200 and 500 yuan per hour. Consequently, for the 20 yuan earned per hour by the worker, companies in the supply chain generate a profit of nearly 17 times that amount. Virtual data companies can sell the same dataset multiple times to different customers in various scenarios. Yao Maoqing, chairman and CEO of Mifeng Technology, stated that before virtual intelligence achieves large-scale commercialization, data as an infrastructure will generate commercial returns earlier than terminal applications. Significant capital has flowed into companies such as Guanglun Intelligence, Wuwen Zhike, Yiren Technology, and Zhiyuan, with some expanding overseas data collection operations. While the volume and variety of data remain in short supply, the prosperity of these firms contrasts sharply with the reality of low-skilled collectors who are concerned with lighter equipment, better training, and higher hourly pay. The fundamental question remains whether this field will become a profession where robots replace humans, leaving workers behind. As new people continue to arrive at the homestay, ready to work for the robots entering their homes, the fate of the human operators remains uncertain. This marks a stark divergence between the technological promise of liberation and the economic reality of a new digital labor force.