Hierarchical Deep Reinforcement Learning Agent for Autonomous Overtaking

Abstract

Over the past few decades, the rapid production of electric vehicles and advanced transportation system technologies has facilitated the development of Autonomous Vehicles (AVs) in intelli gent transport systems. The introduction of AVs can resolve several issues, such as reducing road accidents, minimizing pollution, and following traffic rules, among other benefits. How ever, the major concern regarding AVs is traffic safety in the presence of manually operated surrounding vehicles. Several driving maneuvers, such as lane change, lane following, over taking, and tailgating, play a crucial role in ensuring traffic safety. Among these maneuvers, overtaking is considered the most complex, involving several sub-maneuvers, including lane changing, acceleration or deceleration, and lane following. To address the driving-related is sues in AVs, reinforcement learning plays a crucial role in handling basic scenarios that involve a limited number of vehicles and a static environment. To handle dynamic traffic with mov ing surrounding vehicles, we presented a Hierarchical Deep Reinforcement Learning (HDRL) methodology for performing overtaking maneuver that divides overtaking maneuvers into sev eral sub-maneuvers and executes each maneuver where necessary. Several traffic rules are also integrated into the proposed HDRL approach for efficient handling of overtaking maneuvers by following all general traffic rules. Experiments are conducted and evaluated using the CARLA platform for AVs to test the efficiency of the proposed approach.