Baidu Unveils Apollo World Model Trained on 217 Million Miles of Autonomous Driving Data
Baidu has launched Apollo World, a physical AI world model capable of simulating three years of real-world autonomous driving tests per day using massive dataset from its Apollo Go fleet.
Chinese open-source AI vendor Baidu introduced its new world model, Apollo World, during its AI Day Apollo event. Designed specifically to train autonomous vehicles—including Baidu’s own Apollo Go robotaxi fleet—the model highlights how generative AI is rapidly moving beyond screen-based tasks into physical-world applications.
According to Baidu, Apollo World can generate roughly three years of real-world driving tests in just a single day. The company trained the system on 217.48 million miles of real-world operational data gathered directly from its Apollo Go autonomous fleet.
Simulating the Physical World for Self-Driving Fleets
World models allow AI systems to understand, predict, and simulate physical environments, shifting generative AI from software domains like coding and cybersecurity toward complex real-world operations.
Lian Jye Su, an analyst at Omdia (a division of Informa TechTarget), noted that autonomous driving represents one of the most practical applications for this technology today. "It’s the most reasonable application of world models because the automotive industry is one of the more mature ones," Su said, pointing out that gaming and automotive sectors offer strong commercial opportunities for world model deployment.
Baidu is far from alone in using world models for autonomous transport. Google DeepMind’s Waymo World Model—built on top of its Genie 3 foundation model—converts raw video logs into interactive 3D simulations. Autonomous trucking developer Waabi relies on its proprietary world model to train its virtual fleet prior to highway testing. Meanwhile, Nvidia offers Cosmos, an open-source world model intended to help developers integrate physical logic into robotics and automation workflows.
According to Su, the investment in building precise virtual environments pays off in real-world safety and performance. "The moment you get it right... you have strong foundational data. You can generate very accurate duplicates of a virtual version of the physical world, and that generally contributes heavily to very accurate prediction and generation," Su explained.
High Compute Costs Present Ongoing Challenges
Despite the potential of physical AI, world models face substantial commercial and operational bottlenecks—primarily compute costs.
Training world models demands vast computing resources, and running them in real time requires continuous compute power. Currently, most world models require dedicated infrastructure equivalent to one GPU per active user, making scaling an expensive endeavor for enterprise operators.
Nevertheless, as Baidu’s Apollo World deployment demonstrates, major tech operators view physical AI world models as a necessary foundation for achieving reliable commercial self-driving operations.
