JD.com Reaches Milestone with Embodied AI Data Collection
JD.com announced at the Wuzhen Summit that its embodied AI data collection initiative has covered over 100 sub-scenarios and gathered more than 100,000 real-world operation tasks. The company also detailed its comprehensive data and simulation infrastructure, including the JoyAI-Sim physics simulation engine and custom JoyEgoCam collection terminals. This progress highlights how massive retail and logistics ecosystems can bridge the gap between digital AI and physical robotics through real-world data at scale. By open-sourcing or applying these capabilities, JD.com is accelerating the development of physical AI models that can operate safely and efficiently in industrial and domestic environments. JD.com's infrastructure features a simulation-to-real consistency of approximately 90% and simulation evaluation efficiency reaching 3.2 times that of physical robots. Additionally, the company is building an AI computing cluster scaled to tens of thousands of cards to support its ambitious plan to gather over 10 million hours of human video and 1 million hours of robot operational data within two years.
## BACKGROUND
Embodied AI refers to artificial intelligence systems, such as advanced robotics, that possess a physical form and can interact directly with the physical world through perception, decision-making, and action. Training these models requires vast amounts of multimodal data reflecting real-world human tasks, which companies gather using specialized wearable capture devices and simulation toolchains.