Article · AI translation
IT Home reported on November 4 that Cao Peng, chairman of JD Group’s Technology Committee, delivered a speech today at the Robotics and Embodied Intelligence Forum of the 2026 World Internet Conference Wuzhen Summit, sharing JD’s latest efforts to bring AI from the digital world into the physical world. JD’s embodied intelligence data collection now covers more than 100 specialized production and service scenarios and has collected data on more than 100,000 real-world work tasks.
Cao Peng said competition in physical AI is not just about individual models and robots, but also about the capabilities of the entire system. JD will continue to pursue a “self-developed foundation + open ecosystem” approach, opening capabilities validated in real business operations to industry so that AI becomes a productive force that improves industrial efficiency and everyday experiences.
IT Home noted that JD’s AI is now used in more than 3,000 scenarios across retail, logistics, healthcare, industry and other fields. JD also has more than 3,800 warehouses, over 200,000 partner brick-and-mortar pharmacies, and more than 10,000 stores of various kinds, including JD MALL, 7FRESH and auto service outlets. These diverse business settings provide real environments and continuous feedback for robots learning tasks such as handling merchandise, warehouse operations and household services.
To turn these settings into training data, JD is developing an embodied intelligence data collection center. It plans to collect more than 10 million hours of high-quality video data from real-world human activities and more than 1 million hours of robot operational data over 2 years. The collection will strictly comply with laws and regulations to ensure that it is lawful, compliant, safe and trustworthy. So far, it has covered more than 100 specialized production and service scenarios and collected data on more than 100,000 real-world work tasks.
For data use, JD Cloud has developed collection devices such as JoyEgoCam and built an embodied data service system covering collection, storage, annotation, training, evaluation, simulation and testing. JD has also launched the JoyAI-Sim physics simulation engine, which connects human operation data, simulation environments and robot evaluation to support data augmentation and policy validation. In tests completed so far, simulation evaluations agreed with real-world evaluation results by approximately 90%, and evaluation efficiency reached 3.2 times that of evaluations on physical robots. Tests of JD’s self-developed foundational embodied manipulation model, JoyAI-RA, showed that its scores on tasks performed by physical robots rose steadily as the volume of videos of people performing tasks increased.
For computing power, JD Cloud’s JoyScale AI computing platform uses pooled computing resources, heterogeneous scheduling, and coordinated hardware and software optimization to manage diverse computing resources through one system and schedule them efficiently. JD Cloud has worked with domestic partners to build a cluster of 10,000 domestically produced accelerator cards and plans to build a cluster of 100,000 cards to support large-scale model training and physical AI applications.
For training, the JoyBuilder model development platform provides end-to-end services for embodied model training, improves the efficiency of multimodal data loading and training, and connects training, evaluation and simulation in a closed loop, allowing companies to devote more effort to model and business innovation.
For inference and model services, JD Cloud has developed its own inference engine, substantially improving throughput and resource utilization in high-concurrency scenarios and reducing inference costs for typical tasks by as much as 80%. The one-stop large model service platform JoyMaaS has brought together more than 100 mainstream open-source models and JD-developed models, offering unified access, enterprise-level management and cost controls to lower the barriers to development and adoption for businesses.
On the application side, JD is working to integrate AI and robots throughout the supply chain. For example, JD Logistics plans to purchase 3 million robots, 1 million autonomous vehicles and 100,000 drones over the next 5 years. AgiBot’s embodied robots have entered JD Logistics’ Zhiong warehouse, where the 2 companies are testing the products’ limits and improving reliability in real operations, helping robots move from isolated tasks to coordinated workflows.