In a community in Huishui County, Guizhou Province, a middle-aged resident named Wu Xiumei wears a data collection device on her head while performing household chores such as folding quilts, wiping tables, and sweeping floors, with her every action captured by the device's camera.

Similarly, in a residential area in Suqian, Jiangsu Province, a young mother identified as Wang Liqin has her daily routines wiped tables, arranged fruit, and mopped floors meticulously recorded by the same type of equipment. Meanwhile, in a kiwifruit orchard, several workers wear collection devices while pruning branches and applying tape. Across cities like Xi'an, Ningbo, Shanghai, and Hangzhou, similar data collection operations are unfolding in diverse settings, from cleaning and tidying to supermarket stocking and hotel housekeeping. These everyday tasks now serve an additional purpose: providing learning material for robots. These individuals, effectively working for embodied intelligence systems, share a common identity as data collectors.

Is the promise of earning money by doing housework realistic? What are the barriers to becoming a data collector, and how are time investments and compensation calculated? Following roles like food delivery drivers and ride-hailing drivers, can data collection emerge as a viable flexible employment option for ordinary people?

Zhao Hongmin, a well-known tech blogger and founder of Yipian Network, shared his experience on a crowdsourced data platform. Over two weeks, he collected 76 minutes of video covering scenarios like package unpacking, kitchen cleaning, tea brewing, and desk organization. The platform verified 66 minutes as valid, translating to an efficiency rate of about 87%, earning him 26.37 yuan. Zhao explained that the operation is straightforward: wear the device, connect it, record, upload, and withdraw earnings after review. Pricing varies by task, with package unpacking at 30 yuan per hour, while pet feeding, tea brewing, and kitchen cleaning are priced at 20 yuan per hour.

Wu Xiumei, a standout collector in the pilot program by Guizhou Mengqing Technology Co., captures roughly 4.55 hours of valid data daily. Qiu Gang, the operations manager, noted that the company recruits local residents, trains them through standardized procedures, and settles payments daily or next-day, addressing industry challenges of scattered labor and instability while creating local employment opportunities. In Suqian, JD.com established an embodied intelligence data center last October, attracting stay-at-home mothers, orchard workers, nursing home caregivers, and dance teachers as part-time collectors. Trained residents can take equipment home to gather household data, earning between three to four thousand yuan monthly, with some approaching ten thousand yuan. Wang Liqin describes the work as easy, flexible, and a helpful supplement to her family income.

Job listings for data collectors are widespread on platforms like Boss Zhipin, offering both short-term gigs paying 150 to 300 yuan daily and full-time roles with monthly salaries ranging from 6,000 to 7,000 yuan.

Collectors wear first-person perspective (Ego) devices, representing a no-embodiment approach to data collection, where physical robots aren't present. Other methods include end-effector collection (like UMI) where humans hold camera-equipped grippers to demonstrate tasks, and motion capture requiring specialized suits and gloves. In contrast, embodiment-based collection involves teleoperation or autonomous task data. Currently, Ego and UMI methods offer low cost and scalability, making them the most accessible for widespread participation.

Shi Jinlong, AI product director at Taoding Digital, points out that first-person video collection has lower entry barriers than teleoperation or professional motion capture. Ordinary individuals can handle daily tasks after standard training, enabling mass expansion. However, he emphasizes, "The entry threshold is relatively low, but the quality bar is not." Embodied intelligence requires high-quality data for model training, involving viewpoint completeness, action consistency, task continuity, environmental diversity, and temporal alignment across multimodal data like video, depth, IMU, and voice. Advanced data pipeline development demands stronger task comprehension and execution standards. Qiu Gang adds that the biggest hidden cost is labor, as traditional methods relying on senior engineers are inefficient and costly. Mengqing Data uses a crowd network model, leveraging county-level talent pools with a structured recruitment, training, scheduling, and settlement system. Zhao Hongmin predicts a surge in crowd-sourced collectors as a new flexible work form, especially suited for stay-at-home individuals or service workers seeking supplementary income.

The industry is pushing toward accumulating millions or tens of millions of hours of data. Beijing Humanoid Robot Innovation Center announced delivery of nearly 30,000 hours of high-quality data, aiming for 1 million hours. Mifeng Technology unveiled its 20,000th Ego device, having produced over 1 million hours of data. JD.com targets 10 million hours within two years, mobilizing around 600,000 participants. Overseas, Figure AI launched the Index crowdsourcing app, paying users to upload first-person videos of cooking, laundry, and stocking tasks, committing over $1 billion for data and compute over the next year.

Despite rapid expansion, the effectiveness of this data in improving robot capabilities remains unverified. Closing the loop from collection and training to application and feedback is a critical hurdle. Hong Shaodun, investment director at Jinding Capital, warns that investors are unlikely to back companies built solely around data collection. Whether data collection is a transient need or a lasting profession is debated. Shi Jinlong believes repetitive, low-value collection may be temporary, but demand for high-quality real-world data will persist. While simulation and generative technologies could replace some standardized data, complex tasks, failure cases, and authentic interactions remain valuable. Roles may evolve into specialized tiers like basic collectors, scenario-specific experts, and teleoperation engineers, with job longevity tied to skill adaptability. Zhao Hongmin notes that demand growth is linked to industry and capital cycles; after data saturation, reliance on manual collection may wane, though new scenarios will create fresh needs.

On September 9, the Ministry of Human Resources and Social Security announced 11 new professions, including "Embodied Intelligent Robot Application Technician," signaling official recognition. State media reports project over 1 million new jobs in fields like data collection and robot team coordination over the next five years. Career pathways are opening, but individual continuity, income stability, and skill conversion remain tied to the sector's evolution, a development worthy of ongoing attention.