Baidu Apollo World, Baidu's AI world model, was launched on Tuesday, November 3, 2026, to simulate real environments and train autonomous driving systems.
What happened?
During the Baidu AI Day in Shenzhen, the company revealed Apollo World, which generates up to 100,000 kilometers of driving simulation per day — equivalent to about three years of real-world testing with a safety driver. The system produces real-time multimodal data, such as camera images and lidar point clouds, allowing autonomous driving systems to perceive and navigate virtual environments.
Baidu proposed a "2+3" evaluation standard for world models: the two basic criteria are causality (the future depends on the past, not the other way around) and interactivity (responds to external actions). The three advanced ones include physical accuracy, persistence, and real-time performance.
Why does this matter?
World models like Apollo World drastically reduce the cost and time of training autonomous vehicles. Instead of relying solely on real-road tests, robotaxis can practice complex and rare scenarios in simulation, with AI agents representing other vehicles and pedestrians that make dynamic decisions.
The model is already being used in training Baidu's Apollo Go fleet, which operates in 28 cities globally and has accumulated more than 350 million kilometers of autonomous driving, 240 million of which are driverless.
What changes in practice?
For the autonomous vehicle industry, this means faster development cycles and safety validation at scale. Apollo World supports complete trip tasks, from order to delivery, in continuous virtual cities, creating a data loop that feeds improvements to the real system.
Sources: IT Home and Reuters.
Image credit: Reuters — Source: IT Home / Baidu
By GeekikiBot
