Vision-Based Autonomous Driving: A Hierarchical Reinforcement Learning Approach

Abstract

Human drivers have excellent perception and reaction abilities in complex environments such as dangerous highways, busy intersections, and harsh weather conditions. To achieve human-level driving performance, autonomous driving systems require powerful environmental perception systems and the ability to make accurate decisions in difficult situations, enabling vehicles to maintain smooth driving. However, making decisions based on visual perception is still a daunting challenge for autonomous vehicles, also current end-to-end networks and modular frameworks have limitations in effectively addressing perception, decision-making, and control, such as a lack of interpretability and weak generalization ability in complex environments. This paper proposes an elaborate modular pipeline for autonomous driving that effectively integrates semantic perception information, multi-level decision tasks, and control modules. The decision-making module comprehensively considers high-level maneuver selection and low-level motion control in both horizontal and vertical directions. The proposed MP framework is trained end-to-end by a novel hierarchical reinforcement learning method with a new action sampling mechanism. To adapt to the experimental scenarios, we use the CARLA simulation platform to collect data to evaluate the proposed autonomous driving framework and its new training method, considering various environmental factors such as sunny, dusk, and rainy weather. The results show that the framework exhibits smooth and effective driving strategies in different environments and can converge quickly and stably. The method improves learning efficiency and reduces unnecessary coupling and error propagation. Overall, the proposed framework and training method provide new ideas for improving existing autonomous driving systems.