West Lake World Model Research Center was established to focus on the cutting-edge layout of physical intelligence

China News Service, Hangzhou, October 31 (Linbo) On October 31, the West Lake World Model Research Center was established at the Yungu Campus of West Lake University. The center is a school-level cross-research platform established by West Lake University in the field of physical intelligence. It is headed by Assistant Professor Yu Kaicheng of West Lake University and Chair Professor Jin Yaochu of West Lake University as chief scientist. It focuses on world models and physical intelligence and is committed to breaking through the core bottlenecks of intelligent systems in understanding physical laws, predicting environmental changes, and making interactive decisions.

What is a world model? Simply put, this is another hot word in the field of artificial intelligence after large language models. Large language models are good at probabilistic prediction of language symbols, but "being able to control language is not the same as understanding the world." The world model attempts to give artificial intelligence the ability to "investigate things to gain knowledge", explore the laws of the physical world, deduce the operation of the real world, and move from "viewing the world" to "worldview".

As early as 1990, Richard Sutton, one of the main founders of the field of reinforcement learning, proposed that the world model can be understood as predicting the next state given the current state and action. This requires the model to answer not just “what the world is like” but “what the world will be like”.

Westlake University has been planning in this direction for three years. In 2023, Yu Kaicheng joined Westlake University and established an independent intelligent laboratory, AutoLab, to start world model research. At that time, the concept was still relatively unpopular. The team initially chose to generate extreme scene videos to train the autonomous driving model, and won the ECCV autonomous driving track world championship in 2024.

But during the deployment process, the team began to reflect: Since the model can generate realistic data, it means that it has understood the world to a certain extent. So what exactly is the point of continuing to render pixels?

At the end of 2024, AutoLab shifted to the "conceptual world model" route. It no longer generates pixels, but establishes physical concept representations in abstract mathematical implicit spaces, disassembles complex scenes into reusable concepts such as spatial relationships, material properties, and contact states, and then recombines them to deduce new scenes.

In early 2026, Xihu Digital Intelligence Awomo was established to position itself as a large-scale basic model company for the physical world. Its "conceptual world model" ranked first in four authoritative international robot evaluations, verifying the route's generalization ability in environmental disturbances, fine operations, and long-term tasks.

In addition to autonomous driving, the world model can also become the "physical brain" of the robot and bring changes in the data training model. In Yu Kaicheng's vision, if they rely on a universal physical base, downstream developers or small businesses no longer need to collect millions of hours of data, and may only need 20 to 30 hours of post-training to quickly customize models in specific fields.