I Serve as "Fuel" for Robots: 30 Yuan Per Hour for Folding Clothes, Data Trading Valued at 15 Billion

Behind the massive data gap in embodied AI, a hidden industrial chain is growing at a frantic pace.

On social platforms, recruitment postings for "embodied AI data collectors" and "robot data collection" priced at 240 yuan per day are ubiquitous. Collectors wear specialized devices and repeat actions such as folding clothes, tying shoelaces, opening cabinet doors, and picking up packages day after day in home or outdoor scenarios. These trivial daily action data are packaged into standardized training data by intermediaries, becoming the "fuel" fed to robots, and eventually flowing to various embodied AI enterprises.

At the other end of the industrial chain, a different landscape unfolds. An embodied AI data company valued at over 15 billion yuan has secured two consecutive rounds of financing totaling 2 billion yuan within two weeks. Over the past year, a portion of the hot money in the embodied AI industry has been flowing to these companies that do not manufacture robots but only "sell data".

While robot-making enterprises are still burning huge sums on R&D, data-selling service providers have already started collecting payments from clients. Some enterprises have signed contracts worth 550 million yuan in a single quarter, exceeding the total amount of the previous whole year.

Folding Quilts at Home, Restocking Shelves in Supermarkets: I'm Making "Cheap Fuel" for Robots

Zhang Yue's day starts with clocking in at 8 a.m. Her job sounds simple: tying shoelaces, sweeping the floor, folding quilts... But the execution is far more than just "housework". She must fully mimic the movements of a robot, wearing either 5 cameras or a motion camera on her body, with every angle and every movement of her hands strictly recorded.

Embodied AI data collectors, also known as "robot trainers", collect real-world human behavior data by operating robots or demonstrating actions themselves, which will later be used to train robots to master daily or industrial operational skills.

In Suzhou, home-based data collectors like Zhang Yue earn a daily wage of 250 yuan. Newcomers are paid by clock-in time, from 8 a.m. to 8 p.m., but Zhang Yue says even working nonstop can only yield 6 hours of valid footage, as the actual effective duration depends on how well the equipment cooperates.

According to company regulations, home-based data collectors have to commute to the office daily to deliver USB drives and equipment, and the equipment quality depends on luck. Zhang Yue told Tech Planet that sometimes she gets fully functional devices, but other times the whole day is plagued by malfunctions: "The device suddenly shuts down mid-recording, sensor errors, USB drive read failures, you don't even notice the lens is tilted, and the main unit overheats..."

The most time-consuming step is putting on the wearable devices. The 5 cameras all over the body need their angles adjusted one by one, which takes Zhang Yue half an hour to complete. "You're not even allowed to scratch an itch while wearing the gear. You can take it off if you don't mind the hassle of putting it back on," Zhang Yue explained.

Compared to other scenarios, the home scenario has the strictest time requirements: outdoor recording only needs 6 hours of footage, while home recording must reach 8 hours. This means that even when the equipment keeps breaking down, she has to rush to meet the progress during the limited periods when everything works properly.

The recorded footage will eventually be packaged and sold to robot companies across the country. "Big companies place orders, specifying exactly what scenarios they need, and we film them," Zhang Yue said. "All the recruiters are intermediaries. As long as the data we collect meets the standards, it becomes their training material."

In outdoor scenarios, Wang Lei's work operates slightly differently. Although his daily wage is also 250 yuan, his position has inspectors responsible for adjusting and delivering equipment, so collectors only need to focus on filming, with scenarios changing every certain period. Wang Lei was previously assigned to a supermarket 300 meters from his home, and now he is posted to an auto repair shop 10 kilometers away.

These sites are negotiated by dedicated company sales representatives, including supermarkets, courier stations, logistics hubs, factories, auto repair shops, hotels, and restaurants.

According to Zhang Yue, the company she applied for provides a short group training after the interview, mainly teaching how to wear the equipment properly. The people training with her are all young adults in their 20s, including some college students. All positions in the company, including collectors, inspectors, and sales representatives, are part-time roles. The highest-paying position is "robot QC inspector", which involves reviewing videos in the company office and can earn over 10,000 yuan per month.

Compared to entry-level data collectors, according to recruitment postings on platforms like Liepin, some data collection-related positions requiring technical backgrounds or on-site work generally offer monthly salaries of 8,000 to 15,000 yuan, with senior or management roles reaching over 20,000 yuan.

Chen is a "real robot remote operator", working in a professional data collection center built by a major internet company. His task is to sit at a desk and control a robotic arm to complete a series of instructions: for example, using the left and right hands in sequence to pick up three items and throw them into a trash can.

Chen chose the night shift, which pays 370 yuan per day. The daily tasks themselves are not difficult, similar to a factory assembly line, but he finds it more flexible: "During breaks, I watch the World Cup, then earn 300 yuan and head back to the dorm to sleep," Chen said.

According to Chen, the entry threshold is not high, but there is an implicit age filter: almost all part-time workers alongside him are young people under 25. He remembers on his first day, a 37-year-old woman quit the next day. "60% to 70% of people leave on the first day, because they find the work too tedious to stick with."

Although the client is a major internet company, the data collection business is outsourced to multiple contractors, and Chen signed an outsourcing contract. He has been resting for a week recently because the previous project was suddenly halted, and he is waiting for the next one. However, he doesn't mind these unstable factors: "For part-time work, you can't find anything better than this."

"Screw" Data at the Bottom of the Industrial Chain, Resold by Intermediaries at 10 Times the Price

Embodied AI data collectors are like "screws" on an assembly line, with a single action collected hundreds or thousands of times, providing indispensable training data for humanoid robots learning to fold clothes and open cabinet doors.

Behind this lies a market with extreme imbalance between supply and demand. Industry estimates show that training a robot "brain" approaching human-level capabilities requires 1 billion hours of real operation data, while the current global effective supply is only around 5 million hours, creating a 200-fold gap.

The reason for the explosion of part-time home-based data collector roles, as seen with Zhang Yue, is twofold: on one hand, data collected in laboratories and factories cannot cover real-world environments like homes and shopping malls. Having collectors perform household chores and move around in these scenarios is to obtain the real data that algorithms need for these "capillary-level" scenarios.

On the other hand, compared to "real robot remote operation" which requires expensive robots and specialized venues, the "non-entity collection" model of distributing devices for part-timers to collect data at home is recognized as an effective way to lower the threshold and cost of data collection.

Different from large language models that can draw on massive internet corpora, robots require 3D action data in the real world, such as "grasping, placing, moving, gripping, obstacle avoidance, and operation". The complexity and difficulty of data acquisition are much higher, leading to a prominent "data famine" in the industry. The sector even regards 2026 as the "first year of embodied AI data".

Currently, embodied AI data sources mainly follow three models, showing obvious pyramid characteristics. At the top tier is real robot data, where operators control actual robots via VR devices and exoskeletons, precisely recording every movement and force feedback of the robots. It boasts the highest data quality but also the highest cost, yet it is the key to the deployment of humanoid robots. In the middle tier is simulation data, which refers to the mass generation of robot interaction data in virtual environments for model training. It is low-cost, scalable, and can make up for the current shortage of real robot data. At the bottom of the pyramid are internet videos and human behavior data, which have extensive sources and strong generalization capabilities.

Home-based data collectors are almost at the very bottom of the entire data collection industrial chain. For the same set of data, from collection to sale, the value distribution across the chain is roughly as follows: at the collection stage, the cost paid to data collectors like Zhang Yue and Wang Lei is 30 yuan per hour. After markup by intermediaries, it is sold to robot companies at 300 to 500 yuan per hour. The higher up the chain, the thicker the profit and the higher the valuation.

In addition, the industry is in a seller's market where supply falls short of demand, and the scarcity of data itself further supports its high pricing.

According to reports from The Paper, the current overall pricing of embodied AI data is 200 to 500 yuan per hour, with real robot data being the most expensive, reaching 500 to 1000 yuan per hour on the market.

In April this year, Yao Maoqing, Chairman and CEO of Mifeng Technology, stated that before embodied AI achieves true large-scale commercialization, data as infrastructure will generate commercial returns earlier than end-user applications. He also predicted that the price of non-entity data, which does not rely on specific robot hardware, will eventually converge to one-third to one-half of real robot data prices. For example, if real robot data sells for 1000 yuan per hour, non-entity data will likely stabilize at 300 to 400 yuan per hour in the future.

Robot Companies Enrich Data Service Providers: Who Is the Real "Money Printer"

With exploding demand and intensifying internal competition, the hourly wage of home-based data collectors keeps declining. In sharp contrast, the valuations of data service providers that sell "shovels" and "water" are continuously rising.

One category is independent data service providers, with a typical example being Photon Intelligence. Founded in January 2023, it is now valued at over 15 billion yuan, and recently completed two consecutive rounds of financing totaling 2 billion yuan within two weeks, with a valuation exceeding many star enterprises that directly manufacture robots.

The essence of its business model is "data resale". The same high-quality data can be processed into standardized products and sold simultaneously to multiple leading robot companies such as Agibot and Galaxy Universal. Some high-quality scenario data can achieve a resale rate of over 10 times. This means the collection cost is one-time, but the data assets can be monetized infinitely. In the first quarter of 2026, Photon Intelligence's new orders reached 550 million yuan, exceeding the total of 2025.

Another category is data service enterprises incubated by leading robot companies and operating independently. For example, in February 2026, Agibot internally incubated and established Mifeng Technology. Its founder Yao Maoqing revealed that current data demanders are mainly large model teams, domestic and international tech giants, and startup robot companies, all in a frenzy of "we'll buy as much as you have, and we need it immediately".

Mifeng Technology's strategy is highly aggressive: bypassing the costly real robot remote operation path, it launched the MEgo series of non-entity collection hardware. The company not only sells the hardware devices externally but also uses these devices through its own collection network to generate data, creating a dual revenue stream of "selling shovels + producing data". Its goal is to build a real data production capacity of 10 million hours by 2026.

Unitree Robotics, which is pushing for an IPO, also admitted in its prospectus that its early R&D focused on robot hardware and the "cerebellum" (motion control), with limited investment in the "brain" (embodied large model), and has not carried out large-scale data collection and factory deployment training. To fill this gap, Unitree Robotics plans to invest 2.022 billion yuan (nearly half of its total fundraising) into its intelligent robot model R&D project, with the core task being to "build a large-scale real dataset", further confirming the strategic importance of data in the embodied AI era.

The third category is major internet companies, which leverage their scenario and channel advantages to seize market share through different paths. In March this year, JD announced the launch of an embodied AI data collection center, involving over 100,000 internal employees of various professions and up to 500,000 external practitioners from different industries, with the goal of accumulating 10 million hours of human video data within two years. Baidu launched the "Embodied AI Data Supermarket", building a hierarchical data labeling system to aggregate multi-source data.

The primary market is also increasing its focus on core components and data infrastructure in embodied AI. Statistics from Tech Planet show that from November 2025 to July this year, more than a dozen enterprises including Jianzhi Robotics, Yiren Technology, Luming Robotics, Photon Intelligence, Wuwen Zhike, and Qianxun Intelligence completed financing rounds ranging from tens of millions to hundreds of millions of yuan. Among them, Jianzhi Robotics, founded in July 2025, completed three rounds of intensive financing within 4 months, and secured lead investment from institutions including Ant Group, Didi, and Delian Capital just 5 months after its establishment.

This gold rush centered on embodied data has only just begun.

(Note: Zhang Yue, Wang Lei, and Chen in this article are pseudonyms.)

This article is from the WeChat official account "Tech Planet" (ID: tech618), author: Lin Jing, published with authorization from 36Kr.