Westlake University Establishes World Model Research Center; Yu Kaicheng Team's Three-Year Layout Lands This Year

On October 31, the Westlake World Model Research Center was established at the Yungu Campus. The name may confuse many people: they know about "large language models," but what is a "world model"?

Simply put, this is another "buzzword" in the field of artificial intelligence after "large language models."

Over the past few years, the rapid development of large language models has amazed us with their "extensive reading" and ability to "talk about anything." But being able to command language is not the same as understanding the world. The world model, by contrast, is meant to endow artificial intelligence with the wisdom to "investigate things to attain knowledge" and the insight to "see the small and know the significant"—exploring the laws of the physical world, simulating how the real world operates, and moving from "observing the world" to a "worldview."

From the end of 2025 to 2026, Li Fei-Fei, Yann LeCun, and a host of other AI leaders successively devoted themselves to world model startups. And just a few days ago, Google DeepMind released Genie4, which integrates world models and reasoning models.

The establishment of the Westlake World Model Research Center is a signal that Westlake University will focus on developing in this direction. In fact, as early as three years ago, Westlake University's Autonomous Intelligence Laboratory AutoLab began research on world models and incubated a startup company to carry out technology implementation cooperation with the business community.

Therefore, the establishment of this center has attracted not only a group of AI researchers but also nearly 90 R&D department heads and investors from industry. The "Frontiers of Physical World Models Forum" held after the founding conference was also a focus of their attention.

So, what exactly is a world model? What changes might it bring? And as something new, why do people discuss it endlessly? Of course, there is also the story of "advance planning" over the past three years.

01 World Models

From "what the world is like" to "what the world will be like"

As early as 1990, Richard Sutton, one of the main founders of the field of reinforcement learning, gave a definition of a world model: given the current state and action, predict the next state. This requires the model to answer not just "what the world is like," but "what the world will be like."

Today, getting AI to generate video is not difficult; what is harder is whether, when a cup slides toward the edge of a table, the model can judge in advance what will happen. Yu Kaicheng offers an analogy: the past way of training AI was somewhat like making AI "memorize a large number of exam questions," but world models, more than that, ultimately make precise predictions by "understanding" the laws of the physical world.

Yann LeCun, the "father of convolutional neural networks" and a Turing Award winner, put it more directly: "By merely predicting text, AI will never reach human-level intelligence. We need models that can understand physical reality."

Perhaps the earliest prototype of a world model that the public can most easily perceive is the 2024 video generation model Sora. The most classic image is a wooden sailboat riding the wind and waves inside a coffee cup, with the texture of the waves so realistic. Sora was positioned at the time as a "general-purpose simulator of the physical world."

Soon afterward, the autonomous driving field was the first to implement world models in practice. This type of world model can generate ultra-realistic driving videos and can even proactively construct "edge scenarios" such as severe weather and sudden accidents to train autonomous driving models. This makes up for a shortcoming in the autonomous driving training process, because in the real world, one cannot always create car accidents just to train a model.

But the question is: does being able to generate realistic real-world scenes mean the model truly understands the world?

Simulated rainy-day imagery for training autonomous driving (displayed in early work by Yu Kaicheng's team)

02 "Self-Negation"

—To make machines understand the world better

Born in Shandong in 1993, Yu Kaicheng has a sense of contrast: he describes himself as a "sunny, cheerful, enthusiastic big boy," but technically, he is a bit "rebellious."

In 2012 he took the college entrance examination, and his score was enough for Tsinghua University or Peking University, but in the end he applied to the University of Hong Kong. After graduating with a first-class honors bachelor's degree in engineering from HKU, Yu Kaicheng went to the Swiss Federal Institute of Technology Lausanne for a direct PhD, focusing on automated machine learning. After graduating with his doctorate, he successively conducted research in related fields at Intel Intelligent Systems Lab, Abacus.AI, and Alibaba DAMO Academy.

At DAMO Academy, Yu Kaicheng mainly worked on research related to autonomous driving. In the process, he gradually realized the limitations of traditional training methods. Traditional autonomous driving technology is essentially "rote memorization," highly dependent on training datasets: the model can handle scenes it has seen in the "question bank," but when it encounters rare emergencies, it is very prone to errors.

In 2023, he left Alibaba and joined Westlake University, where he formed the Autonomous Intelligence Laboratory AutoLab and began research in the direction of world models. World models, which are very popular this year, were still a relatively niche concept at that time.

Faced with the difficult problem of training autonomous driving on long-tail scenarios, there were then mainly two approaches in the field. The first was to generate synthetic data, create samples of rare scenarios, and then feed them to the model for training; the second was to fundamentally teach the model to learn by analogy and gain the ability to infer unknown scenarios. Of course, the latter is more difficult.

When he first joined Westlake University, Yu Kaicheng and his team gave priority to exploring the first route, which was relatively less difficult, generating videos of extreme scenarios to train autonomous driving models. This path soon produced results: in 2024 they won the world championship in the ECCV autonomous driving track, with a score far ahead of the second place.

During a deep deployment with a leading domestic automaker, Yu Kaicheng and his team began to doubt the foundation of this route. Since the model could already generate realistic data, that meant it had, to some extent, understood this world. So then, what exactly was the point of continuing to render pixels?

At the end of 2024, AutoLab shifted to a new technical route and proposed the "concept world model"—that is, no longer generating pixels, but instead establishing representations of physical concepts in an abstract mathematical latent space. Yu Kaicheng said that at this point the model needed to, like a human, break complex scenes down into a finite set of concepts, then recombine these concepts, and thereby "understand" new scenes that had never appeared in full in the training data.

In other words, the "concept world model" decomposes complex physical environments, object attributes, spatial relationships, and action consequences into reusable physical concepts, such as spatial relationships, material properties, contact states, and so on. When encountering an entirely new scene, the model does not need to relearn the entire scene, but instead recombines existing concepts in latent space for inference.

As the research moved into 2025, the team began to observe in experiments that, once the concept decomposition and association mechanisms were established, the model's success rate on complex tasks improved significantly compared with traditional imitation learning and reinforcement learning baselines, while maintaining low prediction costs.

This is a technical route of implicit world models, and AutoLab has also become one of the few teams in China to complete preliminary validation. These research advances also laid the groundwork for subsequently incubating a company.

03 More than just autonomous driving

In early 2026, Westlake Digital Intelligence Awomo was officially established. The company positions itself as a foundation model company for the physical world, benchmarked against large language models, with the goal of building a native physical large model rather than merely delivering customized solutions to automakers.

The "concept world model" developed by AutoLab also acquired an official name in 2026—Awomo. Awomo ranked first in all four internationally authoritative robotics benchmark suites, further validating the route's generalization capability under environmental perturbations, fine manipulation, and long-horizon tasks.

At present, this startup has secured cumulative seed and angel funding exceeding 100 million yuan, with investors including hard-tech investment institutions such as Inno Fund, Orient Jiafu Fund, Zhengxuan Investment, Tianqi Capital, Westlake Innovation Investment Fund, and Jinma Investment.

Autonomous driving, however, is only one application scenario of world models. World models can also serve as robots' "physical brain": when facing complex perturbations, a robot can rapidly infer object attributes and action consequences in latent space, and precisely complete grasping, assembly, and long-horizon collaborative tasks.

A bigger change will also occur in data training models. Traditional AI training requires massive amounts of training data, which means extremely high computing costs. But in Yu Kaicheng's vision, if a general physical foundation is relied upon, downstream developers or small enterprises will no longer need to collect millions of hours of data; possibly only 20 to 30 hours of post-training may be needed to quickly customize a model for a specific domain.

In the long run, world models will ultimately serve ordinary people's personalized creative needs—this is something Yu Kaicheng firmly believes. Because there is not too much burden of massive training, users can leverage the ability to "build worlds with models" to construct their own personal robots or personalized intelligent assistants, much like setting up a private garage workshop.

All these applications can be understood as the actions AI attempts to give when facing the world, and all actions will leave trajectories. This is also why Yu Kaicheng is simultaneously studying ancient calligraphy. Chinese characters are pictographs, possessing an inherent sense of structure and spatiality. From oracle bone script to the present, Chinese characters have essentially been a process of compressing entities in the real world into two-dimensional planar symbols.

"Oracle bone script is the earliest latent space," Yu Kaicheng said.

And the process of transforming intent or concepts into strokes is fundamentally connected to the mathematical logic behind robot motion control and trajectory planning. Thousands of years have passed, and we are still trying to abstract the laws of this world—just with different tools.

Appendix: Introduction to the Westlake World Model Center

The Westlake World Model Center was officially established in October 2026 and is a university-level interdisciplinary research platform set up by Westlake University for the field of physical intelligence. Professor Yu Kaicheng serves as director and Academician Jin Yaochu as chief scientist, with multiple PIs jointly undertaking research tasks in various directions.

Oriented toward the development direction of artificial intelligence moving from language cognition to physical interaction, the center focuses on world models and physical intelligence, is committed to breaking through the core bottlenecks in intelligent systems' understanding of physical laws, prediction of environmental changes, and conduct of interactive decision-making, and aims to build an internationally leading academic highland and innovation engine for physical intelligence.

With world models as its core research direction, the center explores intelligent systems that combine perception, cognition, prediction, and action, and promotes artificial intelligence's ability to understand the world, infer the future, and act autonomously in real environments. Relying on Westlake University's research foundation and interdisciplinary advantages, the center will bring together research forces from multiple fields, promote the coordinated development of basic theory, key technologies, and application exploration, and develop general world models capable of understanding physical laws, predicting action consequences, and continuously learning from real feedback, thereby supporting China's independent innovation and frontier layout in the new generation of artificial intelligence.