Baidu Develops World Model Tailored for Autonomous Driving

The game has only just begun.

After large models enter the second half of their development, the competitive focus of artificial intelligence is quietly shifting: from the "digital world" that understands language to the "real world" that understands physical rules.

According to recent media reports, Baidu is developing the "World Model", and it is expected to announce relevant research progress soon. It is understood that this model is designed to learn the representation of the physical environment and predict its changes, and is expected to become the underlying infrastructure for systems that need to operate in the real world.

The report points out that autonomous driving is one of the core application directions of this technology. By predicting the dynamic changes of surrounding vehicles, pedestrians and road conditions, the World Model can greatly improve the planning and decision-making capabilities of autonomous driving systems.

Although Baidu has not yet disclosed the specific application methods of this model and the details of its deep integration with its self-driving taxi business Apollo Go, this move itself has sent an extremely clear signal to the industry: after the end-to-end large model, autonomous driving is brewing a new round of underlying technological paradigm shift, and Baidu is trying to occupy the technological high ground in this cycle.

Farewell to "rule-based code": The World Model reshapes autonomous driving

To understand the strategic significance of the World Model, we must first understand the "data bottleneck" that current autonomous driving is facing.

In the past few years, the technical route of autonomous driving has undergone an evolution from "rule-based" to "data-driven (end-to-end large model)". Although the end-to-end large model makes the driving behavior of vehicles more human-like, it still has a fatal flaw: it cannot accurately predict the future.

When human drivers are driving, they will naturally build a "physical world model" in their minds: when seeing a child playing with a ball ahead, they will instinctively predict that the child may rush onto the road; when seeing the brake light of the preceding vehicle turn on, they will predict the congestion ahead. However, traditional autonomous driving models often perform probability fitting on historical data, lacking deep reasoning of physical causality.

The function of the "World Model" is to endow the autonomous driving system with this "foresight". By learning massive amounts of real physical environment data, it can generate and predict the evolution of the world in the next few seconds. This is not just about identifying obstacles, but about understanding gravity, inertia, friction, and the dynamic game of complex traffic participants.

Once the World Model matures, the autonomous driving system will have more powerful generalization capabilities in the real physical world. It no longer needs to exhaust all extreme scenarios (Corner Cases), but can make the safest planning decisions based on physical common sense when encountering complex road conditions that have never been seen before.

Baidu's hole card: real data and scale effect from 28 cities

In Silicon Valley, Nvidia, Tesla and a number of top autonomous driving startups (such as Wayve) have regarded the World Model as an important path to L4-level autonomous driving. Baidu's high-profile entry at this moment draws its greatest confidence from its huge amount of real road test data.

According to public information, Baidu's self-driving taxi business Apollo Go has expanded to 28 cities around the world. In China's first-tier cities and some complex right-hand drive markets, Apollo Go has accumulated a wealth of real driving data. Public information shows that the total public autonomous driving test mileage of Apollo has exceeded 350 million kilometers.

This is exactly the scarcest fuel for training the World Model.

Different from other companies that are still using simulation data or small-scale road test data to train models, Baidu has massive real road data. These data include extremely complex Chinese road conditions, bad weather, and various unexpected scenarios of human-vehicle interaction.

With these data, Baidu can build a digital twin world with extremely high precision. In this virtual world, the autonomous driving model can carry out highly efficient reinforcement learning, thus greatly reducing the cost and time of real vehicle road tests.

It can be said that while the commercial operation of Apollo Go has verified the Robotaxi business model, it is also continuously generating real road data that can be used to train underlying AI models.

Baidu's anxiety and counterattack: from language to physical AI

In the early stage of the AI large model wave, Baidu took the first-mover advantage with ERNIE Bot. However, with the full counterattack of giants such as ByteDance, Tencent and Alibaba, as well as the open-source involution of startups such as DeepSeek, Baidu's competitive pressure in the general-purpose language large model track has increased sharply.

In the competition between C-end applications and B-end cloud services, Baidu needs to find a new moat that can form an absolute barrier.

The combination of the "World Model" and autonomous driving is precisely the key move for Baidu to reshape its leading position in technology.

From the perspective of industrial logic, the commercial realization path of chatbots in the purely digital world is relatively crowded. Extending AI capabilities to the physical world to solve practical problems in heavy-asset industries such as travel, transportation and manufacturing has far-reaching commercial and social value. This is also the continuation of Li Yanhong's strategy of "empowering thousands of industries with AI" that he has adhered to for many years.

The World Model is actually the bridge for Baidu's transition from "language intelligence" to "physical intelligence". If Baidu can take the lead in closing the loop of the World Model in the field of autonomous driving, it will not only greatly improve the safety and operational efficiency of Apollo Go, but also export this set of technical base externally to empower more robotics industries that require physical environment interaction.

Global shadow war: technological competition between Chinese and American giants

In the research and development of the World Model, Chinese and American technology giants are in a shadow war.

On the other side of the ocean, Nvidia has launched and continuously iterated the Cosmos World Model platform, which aims to provide tools for autonomous driving and robotics companies to generate virtual physical environments; behind Tesla's FSD, there is also a huge world simulator carrying out high-intensity training.

In China, in addition to Baidu, automakers such as Huawei, Xpeng and NIO are also constantly mentioning the concept of the "World Model". However, Baidu's particularity lies in that it not only has the hardware carrier of autonomous driving (Apollo Go vehicles), but also has the underlying computing power infrastructure (Smart Cloud) and basic large model (ERNIE), with full-stack self-research capabilities.

But this does not mean that Baidu can rest easy. The R&D threshold of the World Model is extremely high. It not only requires extremely huge computing power support, but also requires top AI scientists to achieve technological breakthroughs in cross-cutting fields such as physical simulation, generative AI and reinforcement learning.

At present, Baidu has not yet disclosed the specific technical parameters and progress of this model. In the stage of increasingly fierce industry competition, "grabbing the lead" in technology release is often the key to seizing market mindshare.

The ultimate test of commercial closed loop: cost reduction and efficiency improvement

Returning to the essence of business, any breakthrough in underlying technology must ultimately serve the real profit statement.

For Apollo Go, the greatest value of the World Model lies in "cost reduction and efficiency improvement".

At present, the cost of Robotaxi is still high. In addition to the hardware cost of the vehicle itself (such as lidar and computing power chips), the huge road test team and operation cost are also huge expenses. By building a high-fidelity simulation environment with the World Model, Baidu can greatly reduce the mileage of real vehicle road tests, allowing AI to complete the training of most extreme scenarios in the virtual world.

In addition, the World Model can also improve the generalization ability of autonomous driving. When Apollo Go enters a new city, it does not need to re-accumulate local data for several months, but can adapt to local traffic habits faster based on the common sense reasoning of the World Model. The improvement of this capability directly determines the cross-city expansion speed of Apollo Go.

More importantly, only when self-driving is truly "safer than human drivers" can public doubts be completely eliminated and real large-scale commercial operation be realized. The World Model is exactly the key to absolute safety.

Conclusion

From large language models to the World Model, AI is undergoing a profound leap from "understanding text" to "understanding physics".

Baidu's entry into the World Model at this moment is not only a charge to the technical ceiling of autonomous driving, but also a strategic defense to maintain competitiveness in the second half of the AI era. It has data accumulation from 28 cities and a full-stack AI technical base, but this is still an exploration road full of unknowns.

In front of the laws of the physical world, there are no shortcuts. Whether Baidu can win the marathon of autonomous driving with the World Model depends not only on the technological breakthroughs in the laboratory, but also on whether this model can be transformed into a lower-cost, safer travel service in real commercial operations. This is a long process of technological accumulation, and the game has just begun.

This article comes from the WeChat Official Account "Tech Business", authored by Tech Business, and published with authorization from 36Kr.