Learning-Based High-Level Decision Making for Abortable Overtaking in Autonomous Vehicles

Abstract

Autonomous driving technology aims to enhance safety, efficiency, and decision-making in complex scenarios such as overtaking maneuvers. However, existing methods often lack adaptability in dynamic environments, leading to unsafe maneuvers or failed attempts. This paper introduces a reinforcement learning-based high-level decision-making framework that incorporates an abortable overtaking mechanism to improve safety and maneuver success rates. Using a Deep Q-Network (DQN), the system dynamically assesses traffic conditions to decide whether to proceed with or abort an overtaking maneuver, reducing collision risks. The proposed method was evaluated in 3,000 simulation scenarios, achieving a 92.2% successful overtaking rate, a 40% improvement over OpenPlanner’s 52.7%. Additionally, our approach significantly reduced collisions with oncoming traffic from 83.5% to 17.4%, demonstrating its effectiveness in handling dynamic interactions. The system was further validated in real-world experiments using the iseAuto autonomous shuttle, confirming its feasibility for deployment in real-world driving conditions.