Baidu Apollo Unveils Apollo World Model for Autonomous Driving
Baidu Apollo unveiled its agent-driven, all-modal world model named Apollo World at the Shenzhen AI Day event, designed to train autonomous driving systems. Additionally, Baidu introduced a '2+3' evaluation standard for physical AI world models consisting of two basic thresholds and three advanced metrics. This development creates a continuous data flywheel that allows autonomous vehicles to safely and repeatedly practice complex or risky scenarios in virtual cities. By integrating multi-modal observations and closed-loop testing, it helps accelerate the large-scale deployment of L4 autonomous driving globally. The '2+3' evaluation standard includes causality and interactivity as basic thresholds, alongside physical accuracy, persistence, and real-time performance as advanced metrics. Apollo World is already being utilized in Baidu's Apollo Go (Luobo Kuaipao) platform for global testing and validation.
## BACKGROUND
A world model in autonomous driving is an AI architecture designed to simulate and predict real-world dynamics, including environment changes and multi-agent behaviors, to train and evaluate driving policies. Baidu's Apollo Go operates an L4 autonomous ride-hailing service that has accumulated hundreds of millions of autonomous driving kilometers.