京东段楠：AI走向物理世界 竞争不再只是模型之争｜WAIC 2026

This article compiles core insights from JD Group Vice President Duan Nan on AI's transition from the digital world to the physical world, as well as JD Explore Academy's latest technical achievements, shared at the JD Forum during the 2026 World Artificial Intelligence Conference. Key takeaways are as follows:

1. Core industry judgment: All existing internet data originates from the digital world, which cannot support models learning to interact with the physical world. To achieve higher-level general intelligence, AI must expand from the digital domain into the physical world. For large-scale adoption of physical AI, progress depends not only on data, models and computing power, but also on the addition of hardware scaling as a fourth dimension—all four dimensions collectively drive continuous capability improvement.

2. Clear technology evolution stages: JD divides AI's development from digital to physical into three stages. The first is digital intelligence, which completes tasks in the digital world; the second is embodied intelligence, which deploys AI to various terminals to enable natural interaction; the third is physical intelligence, which deploys AI on hardware to autonomously complete tasks in the physical world. JD has already built out a diverse model portfolio to close existing capability gaps.

3. Latest technical achievements: JD has open-sourced multiple innovative models and embodied intelligence datasets, and will launch a physical foundation model and an industry-oriented evaluation platform for use by the entire industry in the second half of the year.

The industry trend of AI expanding into the physical world will bring profound changes to brands' product R&D and scenario-based operations. Key takeaways are as follows:

1. Shifting consumer trends: In the future, AI will penetrate all kinds of physical实体 scenarios. Consumer demand for smart hardware that supports real-time natural interaction and high-empathy voice interaction will grow significantly. Products related to embodied and physical intelligence such as wearables, smart cockpits and service robots will become new consumer growth points, and brands can布局 these tracks in advance.

2. New directions for product R&D: The deployment of physical AI requires collaboration between hardware and models. When developing intelligent products, brands need to adapt to the technical requirements of real-time interaction and multimodal collaboration. They can also leverage JD's open-source models and real-world scenario datasets to lower their own R&D barriers.

3. Opportunities for operational upgrading: Leveraging physical AI technology, brands can optimize operational efficiency in physical links such as warehousing, offline stores and customer service. For example, embodied intelligence can replace部分 manual operations to improve user experience. Brands can advance technological upgrading布局 aligned with their own business scenarios.

The industry trend of AI expanding from the digital world to the physical world brings new growth opportunities for sellers. Key takeaways are as follows:

1. New growth market opportunities: Smart hardware and physical products with AI interaction capabilities will become a new hot track. Smart terminals supporting real-time response and natural empathetic interaction will see sustained growth in consumer demand. Sellers can plan product assortment and布局 in advance to capture traffic in this new track.

2. Accessible technology and cooperation resources: JD has already opened multiple open-source models and embodied intelligence datasets to the entire industry, and will launch an evaluation platform covering all sectors later. Sellers can connect to these resources to reduce operating and R&D costs for intelligent products, and achieve business upgrading through the JD ecosystem.

3. Risk warning: Physical AI is still in the stage of technical iteration, and deployment maturity varies greatly across different scenarios. When布局 related businesses, sellers need to follow the pace of technological iteration, match mature technologies with their own category and scenario, avoid盲目 investing in immature technical directions, and prevent unnecessary waste of resources.

The trend of AI expanding into the physical world brings new product demand and digital upgrading directions to manufacturing factories. Key takeaways are as follows:

1. New requirements for product design and manufacturing: The large-scale development of physical AI cannot be achieved without hardware collaboration. In the future, there will be substantial market demand for intelligent hardware such as robots, wearables, and smart cockpits. In product design and manufacturing, factories need to adjust hardware design schemes to meet AI models' requirements for real-time interaction and multi-device collaboration, to align with new industry demands.

2. New business opportunities: Factories can connect to the physical AI布局 of leading platforms such as JD, participate in the industrial division of labor for AI hardware, provide hardware support for embodied intelligence and physical AI, develop new incremental ToB business, and unlock new growth space relying on the platform's technology and scenario resources.

3. Insights for digital transformation: Factories can leverage physical AI-related simulation technology and real-world scenario data collection technology to optimize their own production processes, reduce the cost of intelligent transformation through open-source technology, optimize production links through data and model training, and promote digital upgrading on the production side.

AI's expansion from the digital world to the physical world is a new development trend for the AI industry, bringing new business directions to AI-related service providers. Key takeaways are as follows:

1. Industry development trend: AI has shifted from simply understanding and generating digital content to perceiving, making decisions and acting in the physical world. Physical AI and embodied intelligence will become the core track of future AI services, and demand for physical AI-related technical services from all sectors will continue to grow.

2. Core customer pain points: Current development of physical AI faces core pain points including high difficulty in physical data collection, inability of models to adapt to real-world physical scenario interaction, and insufficient hardware collaboration capabilities. Customers need integrated deployment solutions covering data, model and hardware collaboration.

3. Technical and business directions for布局: Service providers can布局 around the construction of real physical scenario datasets, research and development of embodied operation models, development of simulation training architecture, and construction of industrial scenario evaluation platforms. They can also connect to JD's already open open-source models and dataset resources to reduce their own R&D costs, and provide adapted physical AI technical services for customers in different industries.

The trend of AI expanding into the physical world puts forward new requirements for platforms' technological布局 and ecosystem construction, and JD's practices provide a reference for other platforms. Key takeaways are as follows:

1. New market demands for platforms: During the AI industry's upgrading toward the physical world, small and medium-sized market participants lack access to technology, data and evaluation resources. They need platforms to provide open technical tools, datasets and validation scenarios to lower industry entry barriers and accelerate technology deployment.

2. Referrable platform operation practices: JD has developed a clear deployment path: first, map out a three-stage technological路线 from digital intelligence to physical intelligence, and gradually fill capability gaps for image, video, voice and interaction models; second, open up open-source models and datasets to attract participation from across the industry; third, build an evaluation platform leveraging JD's own diverse实体 business, and validate technological value in real-world scenarios. This step-by-step, open and shared model is worth referencing.

3. Risk mitigation tips: When布局 physical AI, platforms should not only focus on models and computing power, but also develop hardware collaboration capabilities and real-scenario data capabilities in parallel, to avoid investing resources detached from actual demands of physical scenarios. At the same time, platforms should promote technology deployment in phases, and iterate gradually to reduce risks from technological investment.

This forum disclosed the latest industrial and research trends in the field of physical AI, which carries high reference value for researchers in the artificial intelligence field. Key takeaways are as follows:

1. New industry and field trends: The AI industry has reached a consensus that the next stage of AI development is the transition from digital AI to physical AI. Leading companies such as JD have clarified their布局 direction, divided technological evolution into three clear stages: digital intelligence, embodied intelligence, and physical intelligence, and are promoting technology deployment step by step.

2. New insights and unsolved problems in the field: Duan Nan proposed that the scaling law for physical AI needs to add hardware scaling as a new dimension on top of the original three dimensions of data, model, and computing power. He also put forward a new model training paradigm of "generation-assisted understanding" to solve spatial understanding problems, and identified four core problems that physical AI needs to solve: scaling of physical data, scaling of continuous learning, breakthroughs in foundation models, and scaling of hardware collaboration.

3. Accessible latest research achievements: JD has opened multiple open-source models and embodied datasets to the global research community, including the JoyAI series of image, video, speech and interaction models, as well as the multi-scenario covering EgoLive embodied dataset. The resources have already received applications from hundreds of research institutions worldwide. JD will also release a new physical foundation model and an industrial evaluation platform in the second half of the year, which are available for researchers to carry out follow-up related research.