Baidu Unveils Physical AI World Model Apollo World for the First Time; Nearly 100 Apollo Go Driverless Vehicles Operate with New Architecture

On November 3, at the Baidu AI Day · Apollo Special Event held in Shenzhen, Baidu unveiled Apollo World, a world model for physical AI, for the first time. Wang Liang, chief architect of Baidu's Intelligent Driving Group, said that about 100 Apollo Go robotaxis are already equipped with the next-generation end-to-end architecture and are taking part in operations, with cumulative safe mileage in fully driverless operations approaching one million kilometers.

The next-generation architecture uses an end-to-end model that learns driving policies directly from driving data and predicts control signals such as steering, throttle, and braking. To validate the model's safety, Baidu introduced Apollo World, which builds vehicles, pedestrians, and other traffic participants into AI agents capable of autonomous decision-making, simulates complex traffic scenarios and full trips, and can complete about 100,000 kilometers of virtual road testing per day. However, Baidu also noted that the world model cannot yet fully simulate the real environment, and virtual data still needs to be combined with real road data.

Currently, companies such as Waymo, Pony.ai, and WeRide are also exploring generative simulation technologies. Baidu said that L4 autonomous driving must not only ensure safety but also balance traffic efficiency, avoiding overly conservative behavior that disrupts traffic order. At the same time, model training also needs to filter human driving data to reduce interference from irregular driving habits.

As general-purpose large models enter the autonomous driving field, Baidu believes that real-time performance remains a key obstacle. Driverless systems need to make decisions on a cycle of about 33 milliseconds, while general-purpose models still have limitations in computing power, perception capability, and response speed. As autonomous driving moves from road testing to large-scale operations, world models and end-to-end architectures are becoming important technical directions for improving safety and generalization.