Baidu unveils Apollo World multimodal world model with 2+3 criteria
Baidu unveiled Apollo World, an agent-driven multimodal world model for physical AI, and proposed a 2+3 criteria covering causality, interactivity, physical accuracy, persistence and real-time performance.
At the Baidu AI Day Apollo event in Shenzhen on November 3, Baidu Apollo for the first time disclosed Apollo World, an agent-driven multimodal world model for physical AI, and proposed a 2+3 criteria for judging world models.
The criteria span five dimensions. Causality requires a world model to respect the one-way nature of time, with the past determining the future and predictions based on current state rather than stitching together frames or reasoning backwards. Interactivity requires dynamic response to external intervention, with injected actions affecting future world state. Physical accuracy asks how reliable simulated states are and whether they conform to the laws of the world, and whether learned abilities still work back in the physical world. Persistence concerns whether sustained runtime and stability keep up and whether objects and details drift over long runs. Real-time performance requires low latency so that growing world complexity does not increase inference time.
In the virtual cities built by Apollo World, vehicles and pedestrians are not props following fixed scripts but AI agents with their own decision-making and interaction. The system generates multimodal observations including camera imagery and lidar point clouds in real time, with the driving system making steering, acceleration and braking decisions and feeding actions back, forming an end-to-end loop of perception, decision, action and feedback. Apollo World supports vehicles driving from any origin to any destination in the virtual city, completing the full journey from accepting a ride through carrying passengers to arrival.
Baidu said its Apollo Go platform covers 28 cities globally with over 350 million kilometers of autonomous driving, of which more than 240 million kilometers are fully driverless, and Apollo World is already used for global testing and validation of the service.