Enhancing Autonomous Driving Following Motion Decision-Making through Model-Based Policy Optimization

Reinforcement Learning (RL) presents formidable challenges with respect to safety and efficiency in the context of autonomous driving. In response to these challenges, we introduce a novel algorithm known as Adaptive Rollout Model-Based Policy Optimization (AR-MBPO). Our primary objective is to enhance the decision-making strategy for autonomous vehicles (AVs) that follow other vehicles, with a focus on elevating efficiency, safety, and passenger comfort. Our innovative algorithm incorporates an ensemble environment model designed for handling uncertainties. By using branched rollout techniques for sample collection, we enhance policy performance and sampling efficiency. Additionally, we introduce adaptive rollout length, a dynamic parameter that adjusts based on predictive performance to mitigate model inaccuracies. Our testing of AR-MBPO in AV following simulations demonstrates its superior performance with rapid convergence, reducing real-environment interaction and safety risks. These consistent improvements underscore the importance of adaptive rollout length in algorithm performance, particularly in the presence of imprecise environment models. Comprehensive tests validate its effectiveness in enhancing passenger comfort and operational efficiency.