Baidu Unveils Apollo World Model for Its Apollo Go Robotaxis

At its recent Baidu AI Day Apollo event, the company unveiled a new technology called Apollo World, a "world model" that learns the dynamics of the real world to help train self-driving cars. The model was trained on 217.48 million miles of data collected by Baidu's Apollo Go robotaxi service. According to Baidu, "Apollo World can simulate roughly three years of road testing in a single day."

As one Omdia analyst noted in the same piece, "world models require massive compute resources, typically one GPU per user, which makes deployment very expensive." That is exactly why Baidu, which has its own robotaxi fleet and millions of miles of data, can afford to build systems like this. But if you are working on a product built on agents and language models, and you do not have access to comparable amounts of data or compute, world models for self-driving cars just are not relevant to your situation.