Baidu has unveiled Apollo World, a virtual traffic environment in which AI agents simulate roads, vehicles and pedestrians so that autonomous driving systems can be trained and tested before they reach real streets. The model was presented at Baidu AI Day in Shenzhen on November 3, and the company said in a statement on November 4 that it is already being used to test and validate Apollo Go, its robotaxi service.
How the System Works
Unlike a fixed set of scripted scenarios, Apollo World populates its virtual cities with agents that make their own decisions. Cars, cyclists and pedestrians respond to one another, so when a driverless vehicle slows, changes lanes or yields, the surrounding traffic adjusts. The model generates camera images and LiDAR point clouds in real time, and the driving system uses those inputs to make decisions, which in turn change the simulated environment. Baidu says this creates a continuous loop of perception, decision-making, action and feedback.
The system is also designed to run complete trips. A virtual vehicle can be assigned a pickup point, carry a passenger to a destination and drop them off, with the environment reacting throughout. Baidu says this allows driverless vehicles to be tested on entire journeys before they enter new markets. The model draws on real-world operating data to identify complex traffic conditions and risk scenarios, which are then turned into repeatable tests.
Scale of Testing
Baidu says Apollo World can simulate roughly 100,000 kilometres of driving in a single day, a distance it equates to about three years of continuous road testing by a real vehicle with a safety driver on board. The company also says the model can generate scenarios that have not yet been encountered on actual roads.
Evaluation Framework
Baidu proposed a “2+3” framework for judging world models. The two basic requirements are causality, meaning the model predicts future states from past and current information, and interactivity, meaning it responds to external actions. The three advanced attributes are physical accuracy, persistence over long periods, and real-time performance. Wang Liang, chief architect of Baidu’s Intelligent Driving Group, argued that world models should reproduce physical laws and cause-and-effect behaviour rather than simply produce realistic images.
Apollo Go Background
Apollo Go operates in 28 cities worldwide. According to Baidu’s figures, it has completed more than 23 million rides and accumulated over 350 million kilometres of autonomous driving, including more than 240 million kilometres without a safety driver. The service has begun open-road testing in London with Uber and Lyft and launched fully driverless commercial operations in Dubai.
Limitations and Open Questions
Baidu did not say how much Apollo World will reduce testing costs or what effect it will have on road safety. The company’s claims about the model have not been verified by independent testing, and updated driving systems will still need to be tested on real roads.
