JD.com Bets on a Closed-Loop Physical AI

机器之心 · wechat · 2026-07-18

JD.com Splits Physical AI into Three Layers

This article focuses on JD.com's "Physical AI" roadmap showcased at WAIC: transitioning from "digital intelligence" to "attached intelligence" and ultimately "embodied intelligence." JD.com aims to weave models, data, terminals, and cloud infrastructure into a continuously iterating closed loop. The emphasis is that in the physical world, the key isn't a one-off demo, but whether models can continuously understand a changing reality, acquire real operational experience, and feed failures back as training data.

Closing the Loop: Models, Data, Terminals, and Cloud

The article introduces the JoyAI capability matrix:

On the data front, JD.com open-sourced the embodied dataset EgoLive, featuring 2,000 hours of 60fps binocular first-person video, covering 65,866 task segments across 346 real-world tasks, paired with the proprietary capture terminal JoyEgoCam. JD.com plans to expand data collection using its stores, warehouses, and pharmacies, aiming to hit tens of millions of hours of data within two years.

Terminals and Scenarios are Key to Deployment

JD.com positions JoyInside as the entry point for attached intelligence, embedding AI into mattresses, tea bar machines, desk lamps, toys, and robot dogs to enable proactive collaboration around user intent. Ultimately, JD.com is betting heavily on real-world business scenarios. Its vast network across retail, logistics, industrial, domestic services, food delivery, and healthcare provides the perfect proving ground for training, validating, and iterating. In physical AI, the real differentiator is who can sustain operations in the real world.

Related event: JD.com Showcases Full-Stack Physical AI Ecosystem at WAIC(3 posts)→

Original post →

More from Embodied

Embodied channel →