JD.com scales data collection for embodied AI initiatives

JD.com announced at the World Internet Conference that it has expanded its AI development into physical operational environments. By leveraging its network of 3,800 warehouses and thousands of retail outlets, the company has collected data on 100,000 specific tasks to train its embodied AI models.

The firm is working toward a two-year goal of gathering 10 million hours of video data documenting human activities and 1 million hours of operational robot data. This information is processed through a proprietary physical simulation engine, which the company claims accelerates testing efficiency by 3.2 times compared to physical-only evaluations.

To support the heavy computational demands, JD.com is scaling its AI processing platforms, including plans to build large-scale clusters capable of managing hundreds of thousands of units. The company intends to integrate this technology into its supply chain, with long-term plans to deploy millions of robots, drones, and autonomous vehicles across its logistics network over the next five years.