AI Steps Out of Screens: JD.com Unveils Full JoyAI Model Suite and Embodied AI at WAIC 2026

The 2026 World AI Conference (WAIC) in Shanghai delivered a clear message this year: AI is moving beyond screens and into factories, homes, pharmacies, and everyday devices. According to live reports from QbitAI, the exhibition remained packed even on its third day, with embodied AI, world models, AI coding, and AI hardware pavilions drawing the largest crowds.

Among the exhibitors, JD.com made one of the most notable appearances. The retail-giant-turned-tech-company showcased its full JoyAI model lineup for the first time, covering voice, image, video, real-time interaction, and embodied intelligence — forming a comprehensive AI capability matrix built for the physical world.

On the vision front, JD demonstrated its previously open-sourced JoyAI-Image-Edit, JoyAI-Echo, and JoyAI-VL-Interaction models. JoyAI-Image-Edit enables perspective transformation and geometric consistency editing; JoyAI-Echo maintains long-video coherence through cross-modal memory; and JoyAI-VL-Interaction compresses understanding and response to millisecond latency. JD also debuted JoyAI-Video-Edit, a real-time video editing model allowing users to modify frames during preview.

For human-machine interaction, JD released JoyAI-Voice (a text-to-speech foundation model) and JoyAI-Talker (an interactive voice model). JoyAI-Talker can recognize user emotions and genuine intent, adapt responses to individual speaking habits, and supports low-latency interruption — transforming AI interaction from one-shot commands into ongoing conversation.

To bridge the gap between AI decisions and physical execution, JD introduced JoyAI-RA, a mobile manipulation foundation model. It uses a unified training framework to align different robot types and human operational experience within a shared action representation, then iteratively corrects execution errors through simulation and real-world validation.

On the data infrastructure side, JD open-sourced the industry's largest first-person video dataset, containing 2,000 hours of footage across 65,866 episodes covering 346 real-world tasks. JD also announced that its embodied AI data collection center aims to gather over one million data points within two years, providing scalable training material for physical AI.

JD also unveiled JoyInside, an AI system layer connecting models to smart hardware. Currently integrated with nearly 100 home appliance, robotics, and AI toy brands, devices running JoyInside have seen average conversation rounds increase by over 30%. On-site demonstrations included an AI sleep mattress and an AI projection desk lamp powered by JoyInside, illustrating a concrete pathway from cloud AI to in-home physical deployment.

Taken together, JD's WAIC showcase mapped out a complete “model — data — device — scenario” pipeline for physical AI, marking the industry's transition from virtual capability competitions to creating tangible value in the physical world.