According to Dongcha Beating monitoring, Fudan University Vice President Jiang Yugang, Zhuyuan Robot Partner Yao Maoqing, TESI Aerospace CEO Chen Yilun, and Bright Source Innovation CEO Jiang Xu participated in a roundtable discussion at the 2026 World Artificial Intelligence Conference, focusing on the world model. The guests unanimously agreed that the core of the world model lies in understanding the laws of the physical world's operation to predict the next state or action, rather than just rendering images. It requires native integration of multimodal fusion, physical laws, causal reasoning, and long-term prediction capabilities. The current major bottleneck is data—Chen Yilun pointed out that video data lacks key modalities such as force and touch. Ideally, training data should meet three conditions: complete modalities, high-frequency interaction, and originating from real scenarios, with embodied intelligence requiring complex operations or tens of millions of hours of real interaction data; Yao Maoqing analogized to the language model's hundred billion hours of speech training and estimated that mastering common sense physical predictions in the physical world may require "over one hundred million hours" of real data. At the architectural level, Jiang Xu pointed out that the current mainstream architecture conflates state prediction with action prediction, leading to a conflict between generative and understanding capabilities, making it difficult to optimize both simultaneously.

Regarding the implementation path, all three guests see the manufacturing industry as the most certain large-scale scenario in the next three years:

Yao Maoqing revealed that Zhuyuan Robot has achieved robot fleet operations on the production line for six days with sixty thousand actions and a 99.99% success rate;

Chen Yilun is betting on the manufacturing industry, citing reasons such as high data density, clearly defined task completion standards, and the existence of a large amount of human demonstration data. TESI Aerospace has collaborated with automotive companies to promote the deployment of a thousand-unit industrial embodied robot cluster. He emphasized that China's manufacturing industry is globally concentrated, making it an ideal experimental field for physical AI;

Jiang Xu believes that embodied intelligence is an extension of multimodal large models, with the Internet already having 100 billion hours of video data suitable for pre-training. The leap in capability will first appear in daily scenarios such as homes and offices, but commercialization needs to meet high fault tolerance conditions. Finding scenarios for large models is no easier than training the models.

The consensus among the three parties is that we are still far from universal embodied intelligence, and breakthroughs in specific scenarios are a necessary stage. The future competitive focus will shift from model architecture to high-quality data acquisition and scenario closed-loop verification capabilities.

---------------------------------

Click on the original article link below to join Dongcha Beating · Feishu AI News Channel, monitoring global AI hot topics and news 24/7.