Why Only Closed‑Loop Players Can Win the Physical AI Race at WAIC

At the 2026 WAIC exhibition, more than 200 embodied‑intelligence companies displayed robots that could dance, hand water, and tidy items, yet these polished demos mask deeper industry challenges: robustness under lighting changes, prolonged tasks, and unpredictable environments.

The critical battle has shifted to "physical AI": the ability to keep models continuously understanding a changing reality, to scale the collection of authentic operational data, and to turn robot failures into training signals.

JD.com, a leading e‑commerce giant, has entered this arena by open‑sourcing a suite of JoyAI models and committing to build the world’s largest physical‑world operation center. In a dedicated JD sub‑forum, the company unveiled the JoyAI model matrix, the JoyInside hardware stack, and its cloud AI infrastructure, forming a closed‑loop pipeline.

Three‑Stage Technical Roadmap

JD proposes a three‑step progression: Digital Intelligence (language, multimodal, agent foundations), Embodied Intelligence (embedding these capabilities into phones, cars, appliances), and Physical Intelligence (deploying models on robots to act in the real world).

Following this roadmap, the JoyAI series builds a capability chain: spatial understanding → continuous perception → real‑time interaction → action execution.

Key Model Highlights

JoyAI‑Image integrates image understanding, text‑to‑image generation, and instruction editing, with a focus on spatial relationship modeling.

JoyAI‑Echo generates long‑duration, multi‑camera audio‑visual content while mitigating error accumulation and ensuring temporal consistency.

JoyAI‑Video‑Edit enables real‑time, language‑driven video editing, allowing users to add, remove, or replace objects and characters across frames.

JoyAI‑Talker provides full‑duplex, low‑latency voice interaction, adding emotion understanding and empathetic feedback.

JoyAI‑RA is a VLA model that consumes visual observations and language commands to output executable robot actions, closing the perception‑decision‑action loop.

Data Collection at Scale

JD released the EgoLive dataset, the largest industry‑scale embodied dataset for real‑world tasks, containing 2,000 hours of high‑fidelity binocular video (60 fps), 65,866 task segments across 346 tasks, and rich annotations such as camera pose, 3‑D hand keypoints, depth, segmentation, sub‑task splits, and language descriptions.

The dataset is captured using JD’s self‑developed JoyEgoCam, a dual‑camera RGB + IMU device that records first‑person human operations without hindering hand movements. JoyEgoCam supports instant‑on‑the‑spot data capture in homes, warehouses, factories, and other core scenarios.

JD aims to collect 10 million hours of high‑quality data within two years, leveraging a nationwide embodied‑data collection community and distributed JoyEgoCam deployments.

From Data to Deployment

Collected data is uploaded to the cloud, then undergoes cleaning, alignment, conversion, and pre‑annotation before being enriched by the JoyBuilder simulation platform to cover long‑tail scenarios. The processed data fuels training of models such as JoyAI‑RA.

JoyInside embeds perception, understanding, memory, and interaction capabilities into diverse devices—from AI toys and robot dogs to smart mattresses and printers—creating an "AI space" where multiple terminals collaborate on a shared task.

JD Cloud provides the compute and engineering backbone required for continuous data processing, model training, inference, versioning, and monitoring. Any failure in this chain can stall a model at the demo stage.

Future Outlook

Physical AI must prove its reliability in uncontrolled real‑world settings—homes, stores, factories—where space changes, unexpected events, and unseen tasks occur. JD’s strategy hinges on closing the loop: real‑world data → model improvement → better data collection, enabling robots to operate continuously and scale across domains.