AI Goes Physical: Robots Racing to Gain Real‑World Experience at WAIC

WAIC this year was packed to the brim, with crowds comparable to Shanghai’s weather‑driven foot traffic and even scalped tickets reaching 3,000 CNY, underscoring the intense enthusiasm for AI technologies.

Beyond the hype, the conference highlighted a clear shift: AI is moving out of the screen and into factories, homes, pharmacies, wearables, and other real‑world environments, requiring not only perception but also continuous decision‑making and physical execution.

JD seized the spotlight by unveiling the full JoyAI “model family” for physical AI, including:

JoyAI‑Image‑Edit : enables view‑point changes, spatial roaming, and geometry‑consistent editing, pushing AI from flat‑pixel manipulation toward spatial understanding.

JoyAI‑Echo and JoyAI‑VL‑Interaction : provide long‑video consistency and millisecond‑level interaction, allowing real‑time observation and agent calls.

JoyAI‑Video‑Edit : lets users customise video frames and edit them on‑the‑fly, delivering instant feedback and dynamic adjustments.

JoyAI‑Voice and JoyAI‑Talker : recognise user emotions and intents, adapt responses to individual speaking habits, and support low‑latency, interruptible dialogue.

JoyAI‑RA : a mobile manipulation model that aligns robot actions with human experience through a unified training framework, simulation, and real‑world validation.

JD also shared a concrete experiment comparing 10,000 hours of high‑quality data with 10,000 hours of flawed data for model training. The result showed that merely increasing data volume does not guarantee improvement; the flawed dataset actually degraded performance, emphasizing that the value lies in accurately captured actions, trajectories, task outcomes, and scene variations.

To close the loop, JD has built a massive embodied‑data pipeline: more than 600 k participants are contributing to what JD calls the world’s largest embodied data collection centre, targeting 10 million hours of real‑world human data within two years. The publicly released EgoLive dataset contains 2,000 hours of video, 65,866 episodes, and 346 distinct real‑world tasks, providing first‑person recordings of environment, hand motions, and task intent.

All collected data undergo automatic annotation, quality optimisation, and structuring before being fed into digital‑twin environments for training and evaluation, after which the refined models are deployed to physical robots.

JD further demonstrated how these capabilities are being integrated into everyday devices through the JoyInside AI brain. Examples include a smart mattress that continuously monitors sleep signs and proactively queries the user, and an AI‑projector lamp that recognises handwritten problems, tracks learning progress, and projects model‑generated explanations directly onto the desk. Devices equipped with JoyInside have seen dialogue turns increase by over 120 % and now span nearly 200 brands across home appliances, robots, and AI toys, with a target of tens of millions of terminals in the near future.

In summary, JD’s showcase at WAIC illustrates a concrete pathway from cloud‑based AI research to embodied intelligence that perceives, understands, decides, and acts in the physical world, supported by a closed‑loop data‑model‑deployment ecosystem that is already delivering double‑digit accuracy gains in real‑world tasks.