Embodied AI at WAIC 2026: The Inflection Point for AI in the Physical World

On July 18, 2026, the World Artificial Intelligence Conference (WAIC) opened in Shanghai. If last year's keyword was "large language models," this year the spotlight unquestionably fell on "embodied AI."

Tencent presented its full-stack embodied AI strategy for the first time, spanning cloud infrastructure, model layers, platform layers, and application layers. Meanwhile, outside the main venue, the SimFoundry system — jointly released by NVIDIA's GEAR Lab and Fei-Fei Li's Stanford team — was generating significant buzz. Taken together, these developments send an unmistakable signal: AI is moving from algorithms on screens to real-world, physical applications.

Tencent's Full-Stack Play: From Models to Ecosystem

At WAIC, Tencent Chief Scientist Zhang Zhengyou stated: "True intelligence requires connecting language, vision, spatial cognition, bodily control, and environmental feedback — validated within a closed loop of perception, embodiment, and action."

Behind this statement, Tencent has assembled three key assets in embodied AI:

Model Layer: Tencent released three embodied foundation models — Hy-Embodied-VLA-0.5 (unified vision-language-action modeling), Hy-Embodied-VLM-1.0 (a perception model achieving flagship-level performance at just 1/10th the computational cost), and Hy-Embodied-RxBrain-1.0 (a cognitive brain that pioneers deep collaboration between textual reasoning and visual imagination). All three have been validated in real industrial settings and tasks including shopping guidance and elderly care services.

Platform Layer: The TairosAgent embodied agent framework and Apexio agent integrate the "left brain," "right brain," "cerebrum," and "cerebellum" with the physical body into a unified system, enabling robots to form a coherent loop of continuous perception, decision-making, and action. The open-source Tairos platform has also been upgraded, lowering the full-chain development barrier from models to hardware.

Infrastructure Layer: Tencent Cloud introduced the industry's first EaaS (Embodied-AI-as-a-Service), covering the entire pipeline from computing infrastructure to perception and interaction, from R&D to training to deployment.

Tencent VP Lin Songtao summarized it well: "The sign of AI democratization is not making the strongest even stronger, but enabling ordinary people to become extraordinary with AI."

SimFoundry: Ending the Sim2Real Bottleneck with a Single Video

If Tencent is building the "operating system" for embodied AI, the SimFoundry system from NVIDIA and Fei-Fei Li's team solves the most painful fuel problem: training data.

The Sim2Real gap has plagued robotics for nearly two decades. Simulation environments require manual 3D modeling, taking weeks or even months to build. When policies trained in simulation are deployed in the real world, performance collapses due to subtle differences in lighting, friction, and object deformation.

SimFoundry flips the problem on its head. With just a single ordinary RGB video, the system automatically generates an interactive, trainable, and evaluable simulation environment. More critically, its "digital cousin" mechanism can automatically swap objects, rearrange layouts, and generate new tasks while preserving functionality and affordance — turning one real-world video into a near-infinite data generation space.

The experimental results are compelling: the average Pearson correlation coefficient between SimFoundry's simulation evaluations and real-world performance reached 0.911. Policies trained entirely in simulation can be deployed zero-shot onto real robots.

This transforms the one-way "simulate first, deploy later" pipeline into a complete closed loop: real-world capture to simulation expansion to real-world deployment.

An Industry Inflection: Three Forces Driving a Trillion-Yuan Market

The concentrated explosion of embodied AI in 2026 is driven by three converging forces:

Policy: China's 15th Five-Year Plan has designated embodied AI as a future industry. The Ministry of Industry and Information Technology's "Real-Scene Training Initiative" is accelerating data accumulation and model iteration in authentic environments.

Industry: According to sector reports, China's embodied AI market is projected to surpass one trillion RMB in 2026. Companies like Unitree, UBTECH, Agibot, and Leju Robotics have already navigated the path from prototypes to mass production. In April, a humanoid robot completed the Beijing Yizhuang Half Marathon faster than the human champion — not a gimmick, but a demonstration of engineering capability.

Technology: Breakthroughs in foundation models are providing crucial momentum for embodied AI. Tencent's Hunyuan Hy3 saw a 68x increase in total API calls within its first week compared to its predecessor, topping the OpenRouter global leaderboard. The stronger the foundation model, the better robots can understand the world and plan actions.

The industry roadmap is becoming clear: 2025-2026 for mass production validation, 2027-2028 for ecosystem formation, and 2029-2030 for an inflection point. Multiple industry leaders predict the AGI inflection point will arrive around 2030.

True Intelligence Must Enter the Physical World

Judging from WAIC 2026, the competitive logic of embodied AI has fundamentally shifted. It is not another rerun of the LLM race — the contest is no longer about parameter counts and benchmark scores, but about "can it actually get things done in the real world?"

Tencent's EaaS, NVIDIA's SimFoundry, SenseTime's "Galaxy Plan" with over 20 domestic chip companies — these moves all point in the same direction: lowering barriers, accelerating deployment, and bringing embodied AI from a handful of tech company labs into thousands of industries.

As one industry observer put it, the "ChatGPT moment" for embodied AI may still be a few years away. But when that day comes, what walks toward us won't be a talking dialog box — it will be an intelligent agent that can truly take action.