WAIC 2026 Roundtable: General-purpose embodied intelligence needs to break through specialized scenarios first; in the future, the competitive focus will shift to high-quality data acquisition and scenario closed-loop validation

According to Beating 監測 monitoring, Fudan University Vice President Jiang Yugang, Yuanzhi Robot partner Yao Ouqing, Itshi Zhihang CEO Chen Yilun, and Liangyuan Xinchuang CEO Jiang Xu held a roundtable discussion at the 2026 World Artificial Intelligence Conference, focusing on world models. The guests reached a consensus that the core of a world model lies in understanding the underlying laws of the physical world and predicting the next state or action, rather than merely rendering images; it must naturally master multimodal fusion, physical laws, causal reasoning, and long-horizon prediction capabilities. The biggest current bottleneck is data—Chen Yilun noted that video data lacks key modalities such as force and touch. Ideal training data should meet three conditions: complete modalities, high-frequency interactions, and deriving from real-world scenarios—embodied intelligence either due to the complexity of operation, or needs tens of millions of hours of real interaction data. Yao Ouqing compared it to the billions of hours of speech training volume for large language models, estimating that the physical world may require “more than one hundred million hours” of real data to learn common-sense physical prediction. On the architecture level, Jiang Xu pointed out that current mainstream architectures conflate state prediction and action prediction, causing conflicts between generation and understanding capabilities, making it difficult to optimize both simultaneously.

In terms of deployment pathways, all three guests view manufacturing as the most certain large-scale scenario for the next three years:

Yao Ouqing revealed that Yuanzhi Robot has achieved six days and sixty thousand operations on a production line, with 99.99% success rate for robot swarm operations;

Chen Yilun is betting on manufacturing, citing reasons including high data density, clear task completion standards, and a large amount of human demonstration data. Itshi Zhihang has already cooperated with automakers to advance the deployment of an industrial embodied robot cluster at the one thousand unit scale, and emphasized that China’s manufacturing industry is the most globally concentrated, making it an ideal testing ground for physical AI;

Jiang Xu believes that embodied intelligence is an extension of multimodal large models. The internet already has ten billion hours of video data suitable for pretraining. The capability jump will first appear in everyday scenarios such as homes and offices, but commercialization requires meeting conditions for high fault tolerance—finding scenarios for large models is no easier than training models.

The three parties’ consensus is that we are still far from general embodied intelligence. Breakthroughs in specialized scenarios are a necessary stage; in the future, competition focus will shift from model architecture to the ability to obtain high-quality data and validate scenario closed loops.

Click the original text link below to join the Beating · Feishu AI News channel and get 24/7 nonstop monitoring of global AI hotspots and news.