Baidu's Apollo World world model trains an end-to-end robotaxi driver that has 1 million driverless km

Baidu has built a virtual city in which every car and pedestrian is an AI agent that reacts to what a robotaxi does, and it already uses that city to test its Apollo Go service. Baidu Apollo showed the world model, called Apollo World, at its AI Day session in Shenzhen on Nov. 3, and said it is in use for Apollo Go's testing and validation worldwide, National Business Daily and InfoQ reported.

It matters most for the newest part of Baidu's fleet. Wang Liang, chief architect of Baidu's Intelligent Driving Group, said Apollo Go is the only platform in China running Level 4 driverless service at scale with an end-to-end system, and that this system has logged more than 1 million kilometres of fully driverless operation. Apollo Go as a whole has more than 240 million driverless kilometres, out of more than 350 million autonomous kilometres, across 28 cities and more than 23 million rides, InfoQ and NetEase Tech reported. The end-to-end driver has about 0.4% of the fleet's driverless record, and Apollo World is where Baidu means to make up the rest.

Traffic that reacts to the robotaxi

In conventional simulation, other road users act like props following a script, InfoQ wrote. In Apollo World, when the robotaxi brakes, changes lanes or yields, the agents around it can respond in different ways, and the model generates camera and lidar data in real time so the driving system perceives the virtual city directly. Yang Ze, the Baidu architect who leads the world model work, said cars can drive between any two points in the city and complete whole trips, from taking the order to dropping off the passenger, where many models in the industry simulate single scenes.

Baidu also proposed a "2+3" test for what counts as a world model. The two thresholds are causality, predicting the next moment from the current state, and interactivity, responding when an action is injected. The three measures of quality are physical accuracy, persistence over long runs, and real-time speed. On open roads, Wang said, a safety decision has to be made within 1/30 of a second.

Ten virtual days equal the real record

Industry analysts quoted by InfoQ say Apollo World can drive 100,000 kilometres a day, about what a real test car with a safety driver covers in three years. At that pace, ten days in the simulator match every fully driverless kilometre the end-to-end stack has driven on real streets.

The case against leaning on that number comes from Baidu's own list: physical accuracy, meaning whether what the driver learns in the virtual city still works on a real road. Baidu says real operating data feeds the high-value scenes in Apollo World and the virtual results feed back into the cars, which is the sim-to-real loop every world model has to close.

London will be a hard version of that test. Freenow by Lyft is testing about 24 Apollo Go cars in the borough of Brent and hopes to run them fully driverless as early as next year, as Robrief reported. Will those cars drive on the end-to-end stack that Apollo World trains?