At the AI Day·Apollo special event held in Shenzhen, Baidu introduced Apollo World for the first time—a world model designed for physical AI—and announced the operational progress of its next-generation end-to-end autonomous driving architecture. Currently, around 100 Apollo Go robotaxis equipped with this architecture are in active service, accumulating nearly one million kilometers of fully driverless operational safety mileage. This milestone marks the architecture's transition from technical testing to real passenger-carrying services. The next-generation end-to-end architecture draws on technical approaches from large language models, enabling the model to learn driving strategies directly from massive driving data. Compared with traditional autonomous driving systems that are divided into modules and rely on rule-based constraints, it is more flexible, but it also brings safety validation challenges in complex scenarios. Apollo World is precisely the virtual testing environment Baidu has built for this purpose, modeling traffic participants as AI agents with decision-making and interaction capabilities, supporting virtual road testing for complete trips, and capable of completing about 100,000 kilometers of virtual road testing per day. However, at the current stage, it still cannot fully simulate all the changes in the real world and needs to be used in combination with real road data. At present, world models and generative simulation have become directions commonly explored by the autonomous driving industry, with overseas companies such as Waymo and Wayve and domestic companies such as Pony.ai and WeRide all having launched related initiatives. Baidu stated that Apollo Go's L4 driverless driving needs to avoid being overly conservative while ensuring safety, improve traffic efficiency to better integrate into local traffic flows, and at the same time screen high-quality driving training data to adapt to traffic rules and driving habits in different regions. Recent research experiments show that general-purpose large models have initially demonstrated certain real vehicle control capabilities, but limited by perception capabilities, computing power, and especially real-time response requirements, general foundation models will still find it difficult to directly undertake L4 driverless driving tasks for a relatively long period of time.