Baidu introduces apollo world, a physical AI model, and reports progress with driverless fleet

Baidu debuted Apollo World, a novel world model designed for physical AI, during its AI Day Apollo Special Event in Shenzhen on November 3. Wang Liang, chief architect of Baidu's Intelligent Driving Group, confirmed that approximately 100 Apollo Go robotaxis are now utilizing the new end-to-end architecture, achieving nearly one million kilometers in cumulative safe, fully driverless mileage. Apollo World functions by modeling vehicles and pedestrians as autonomous AI agents, allowing for the simulation of complex traffic scenarios and full trips, capable of running around 100,000 virtual kilometers daily. Despite this capability, Baidu noted that the model still requires integration with real-world data as it cannot perfectly replicate the actual environment. While competitors like Waymo, Pony.ai, and WeRide are also exploring generative simulation, Baidu emphasized that L4 autonomous driving must balance safety with traffic efficiency, avoiding overly cautious driving. Furthermore, the company highlighted that while general-purpose large models are entering the sector, the 33-millisecond decision cycle required for driverless systems remains a significant computational hurdle.