THE ESSENTIALS IN 30 SECONDS

  • Baidu unveiled Apollo World on November 3, an AI system that simulates road environments and predicts how they change in order to train self-driving software.
  • The company says it can produce up to 100,000 km of simulated driving per day, which it equates to about three years of real-world testing with a safety driver.
  • Baidu says the model already trains the Apollo Go robotaxi fleet; it has not published independent results or technical details.

Baidu on November 3 introduced Apollo World, an AI system designed to simulate real road environments and predict how they evolve, so that autonomous-driving software can be trained without sending cars onto public streets. Reuters reported the launch from Beijing. The announcement follows reports in late October, carried by TechNode and 36Kr, that the company was building such a model and would soon share progress.

What Baidu says the model does

According to Baidu, Apollo World can generate up to 100,000 km of simulated driving each day. The company equates that volume to roughly three years of real-world testing by an autonomous vehicle with a safety driver on board. The system produces multimodal data in real time, including camera and lidar observations, so that a self-driving stack can perceive and navigate the generated scene as it would a real one. Baidu says the model is already being used to train the fleet of Apollo Go, its robotaxi service, which Reuters describes as China's largest robotaxi operator.

These figures come from Baidu. Neither Reuters nor the earlier reports provide independent benchmarks, technical parameters or evidence of how much simulated training improves on-road safety.

Why world models matter for robotaxis

World models learn a representation of a physical environment and forecast what happens next, such as how nearby vehicles and pedestrians will move. A 36Kr article published before the launch argued that the approach could let Baidu run reinforcement learning in a digital twin, reduce costly road testing, and help Apollo Go adapt to new cities faster than the months of local data collection it currently needs. The same piece noted that Apollo Go operates in 28 cities and that Apollo's public road-test mileage has passed 350 million km, data that can feed such a model. Those are the outlet's assessments, not Baidu's stated plans.

Competitive context

Baidu is not alone: 36Kr points to similar efforts at Nvidia, with its Cosmos platform, at Tesla and at the startup Wayve, and says Huawei, Xpeng and NIO have also discussed world models. In China, Apollo Go competes with WeRide, Pony.ai and Momenta, while Xpeng plans to start robotaxi operations in 2027, Reuters reports. Training cost is a central issue: Frank McCleary of Arthur D. Little told Reuters it is a "challenge that every autonomous vehicle company is going to face."