Baidu has introduced Apollo World, a virtual traffic environment for training and evaluating autonomous-driving systems. The company said on November 4 that the model is already supporting global testing and validation for its Apollo Go robotaxi service. Baidu says a day of virtual driving in Apollo World represents roughly three years of continuous testing by a real vehicle with a safety driver.
The simulation is designed to cover an entire journey, from accepting a ride request and collecting a passenger to completing the drop-off. Traffic participants—including pedestrians and other vehicles—are represented by AI agents that respond to the autonomous vehicle’s actions instead of following fixed scripts. The system generates live camera and LiDAR data, then updates the virtual traffic after the vehicle acts, creating a repeating cycle of perception, decisions, driving and feedback.
At a Baidu AI Day session in Shenzhen, Wang Liang, chief architect of the company’s Intelligent Driving Group, said a world model should capture physical rules and the causes of behavior, not just produce realistic-looking scenes. Baidu’s proposed evaluation framework sets causality and interaction as basic requirements, with physical accuracy, persistence and real-time performance as further measures. The company says real-world traffic and risk cases can be turned into repeatable virtual tests, but updated driving systems still need road testing; it did not quantify cost savings or safety gains.
Apollo Go operates in 28 cities worldwide. Baidu reported in its August second-quarter earnings release that the service had logged more than 350 million kilometres of autonomous driving, including over 240 million fully driverless kilometres. The company has also begun open-road testing in London with Uber and Lyft and started fully driverless commercial operations in Dubai.