Baidu's Apollo World Simulates 100,000 km of Robotaxi Driving a Day, With Every Other Car and Pedestrian Run as an AI Agent
Baidu launched Apollo World on November 3, an AI world model that simulates city traffic for training and testing its Apollo Go robotaxis. Baidu says it can generate up to 100,000 km of simulated driving a day, about three years of real-world testing by one car with a safety driver, and produces live camera and lidar data so the driving system reacts as it would on a real road. Other road users in the simulation are modelled as AI agents that respond to the robotaxi.
For robotaxi companies, real road miles are slow and expensive to collect, and Baidu is betting it can generate many of them in simulation instead. On November 3, at a Baidu AI Day session in Shenzhen, the company launched Apollo World, a "world model" that simulates real traffic for training and testing its Apollo Go robotaxis. Baidu says it can generate "up to 100,000 km" of simulated driving a day, which it says equals about three years of real-world testing by a car with a safety driver, Reuters reported.
What Baidu says it does#
Apollo World produces real-time camera and lidar data, so the driving software senses the simulated street as it would a real one, Reuters said. When the car brakes, changes lane or yields, the virtual world changes and feeds new data back to it. Baidu says the simulation is already being used to train Apollo Go's fleet.
The difference from traditional simulators, according to Yang Ze, the Baidu architect who leads the world-model work, is that other road users are not scripted. "Apollo World models all traffic participants in the virtual scene as AI agents with decision-making and interaction capabilities," he said, as reported by the Chinese auto outlet Gasgoo, which means a pedestrian or another car can respond differently each time the robotaxi acts. Cars can also drive complete trips between any two points, from pickup to drop-off, rather than isolated scenarios.
Wang Liang, chief architect of Baidu's intelligent driving group, said a world model should be judged less on how good its graphics look than on whether it is an "examiner" that can verify safety. Baidu proposed its own standard: a world model must be causal and interactive, and is better the more physically accurate, persistent over long runs and real-time it is.
Apollo Go's real-world data, and rivals' plans#
Apollo Go says it operates in 28 cities and has completed more than 23 million rides, with more than 240 million km of fully driverless driving, figures that supply the real-world data Apollo World is built from. Rivals are also moving: Reuters noted that Xpeng plans to start robotaxi operations in 2027, after it said it had won approval for L4 robotaxi demonstrations across Guangzhou. "They still have the capex of training models … and vehicles," Frank McCleary of Arthur D. Little told Reuters.
Limits#
- The 100,000 km figure and the three-year comparison are Baidu's; it has not published evidence that simulated kilometres predict real-world safety.
- Apollo Go's ride and distance totals are company figures.
- Quotes from Wang Liang and Yang Ze are our translations of Gasgoo's Chinese-language report.
- Baidu did not give technical details such as model size or how often simulated results are checked against real driving.