JD.com Unveils JoyAI Full-Series Large Models, Embodied Data Collection and Training Pipeline at WAIC

en.Wedoany.com Reported - JD.com, China's leading e-commerce company, showcased for the first time at the World Artificial Intelligence Conference (WAIC) its full series of JoyAI large models, embodied data collection and training pipeline, and the AI Home built on the JoyInside platform, marking the extension of physical AI from digital content understanding to environmental perception, real-time decision-making, and concrete execution in the physical world.

At WAIC, JD.com launched the "full family bucket" of JoyAI models for the physical world, covering voice, image, video, real-time interaction, and embodied intelligence. Among them, in the visual model category, the open-source JoyAI-Image-Edit enables perspective transformation and spatial roaming, JoyAI-Echo maintains long-video consistency through cross-modal memory, and JoyAI-VL-Interaction compresses understanding and response to millisecond levels. The latest research achievement, the real-time video editing model JoyAI-Video-Edit, allows users to modify the footage in real-time during preview. In voice interaction, JD.com released the foundational model JoyAI-Voice and a new voice interaction model JoyAI-Talker, the latter capable of recognizing user emotions and true intentions, featuring low-latency interaction and the ability to be interrupted at any time. In the execution phase from decision to action, JD.com demonstrated the mobile manipulation foundational model JoyAI-RA, which aligns different robots with human operational experience through a unified training framework.

To support the operation of physical AI in the real world, JD.com has built the world's first embodied data infrastructure covering the entire chain of collection, storage, annotation, training, evaluation, simulation, and testing. Currently, 600,000 people have been mobilized to participate in data collection, and the world's largest embodied data collection center has been established, with a plan to collect 10 million hours of real human data within two years. At WAIC, JD.com open-sourced the industry's largest human first-person perspective dataset, EgoLive, containing 2,000 hours of video and 65,866 episodes, covering 346 real-world tasks. Visitors can wear the self-developed ultra-high-definition collection terminal JoyEgoCam to complete tasks in front of shelves. After automatic annotation and structured processing, the data is fed into a digital twin environment for training and then deployed onto physical robots, forming a complete closed loop.

JD.com has also encapsulated the capabilities of the JoyAI large model into the intelligent hardware AI system JoyInside, connecting perception, understanding, memory, and interaction capabilities to handle voice recognition, intent understanding, long-term memory, and multi-device coordination in home scenarios. At WAIC, JD.com showcased a smart sleep mattress equipped with JoyInside, capable of capturing sleep signs and proactively inquiring; as well as the JoyInside AI projection desk lamp, which provides real-time Q&A and corrections through screenless diffuse projection and multimodal homework scanning and grading. Currently, the average number of conversation turns for smart hardware connected to JoyInside has increased by over 120%, and the collaboration scope has covered nearly 200 brands in home appliances, home furnishings, robotics, and AI toys. It is expected that over 10 million terminal devices will be connected this year.