A scene where people contribute data to global artificial intelligence development while performing tasks like wiping tables and folding clothes is now a reality in Suqian, Jiangsu province. JD.com announced on May 20 that the country's first embodied intelligence data collection community has officially commenced operations in Suqian, offering a new approach to address the "data scarcity" faced by the embodied AI industry.
The embodied AI industry has seen rapid development this year, but a shortage of high-quality, practical operational data has become a major bottleneck. Industry estimates indicate that training a generalizable embodied model requires at least tens of millions of hours of real-world scenario data. Currently, available high-quality practical data amounts to only a few hundred thousand hours, revealing a significant gap. Concurrently, issues such as high data costs, lack of standardization, and low reusability further hinder technological iteration and large-scale implementation.
To tackle this challenge, JD.com has partnered with Suqian to establish the embodied intelligence data collection community. Located in Suqian's Hubin New District, the community has seen enthusiastic participation from residents since its trial operation began in April. After receiving professional training, data collectors can perform data collection during their routine household chores. While engaging in normal activities like wiping tables, folding clothes, organizing storage, and floor cleaning, collectors wear JD.com's self-developed JoyEgoCam terminal. This device captures critical parameters such as upper limb trajectories, force distribution, and interactions between humans and home environments.
Weighing only 220 grams, the device is equipped with an inference unit and automotive-grade IMU (Inertial Measurement Unit), enabling millimeter-level precision data collection in various environments including homes, outdoors, and production lines without disrupting normal life. The data collection work is flexible and straightforward, and participants can earn considerable income, making it popular among stay-at-home groups. The collected data, after undergoing processes like upload, quality inspection, and annotation, becomes high-quality "data fuel" for training embodied AI models, making robots smarter and more capable.
Currently, JD.com has established a complete data pipeline covering "collection—annotation—training—validation," spanning five core scenarios: logistics and warehousing, industrial manufacturing, health services, home services, and urban operations. This system records multi-dimensional data including visual, tactile, and spatial trajectory information. JD.com plans to mobilize over 100,000 internal employees across various professions and 500,000 external personnel from different industries for large-scale data collection initiatives. In Suqian alone, the initiative aims to involve over 100,000 citizens, covering more than a hundred detailed scenarios across households, offices, factories, logistics, stores, and sanitation services.
In nursing homes, collectors record caregiving actions such as assisting the elderly to stand up, administering medication, and conducting rehabilitation exercises. In farmlands, they capture hand-eye coordination trajectories for fruit and vegetable picking and mechanical data for tool handling. In garment factories, they document precise hand operations like sewing and cutting. By sourcing data from real-world scenarios and applying it back to them, JD.com is accelerating the development of the "world's largest physical world operation center," contributing to the construction of an artificial intelligence ecosystem.
Comments