Autonomous Driving Agents for Safe Road Junctions Through Deep Reinforcement Learning

Autonomous driving agents have garnered significant public interest, particularly concerning their safety within transportation systems. Deep reinforcement learning (DRL) has emerged as a promising approach for developing these agents, especially in tasks such as obstacle avoidance. However, agents often operate with partial observations in complex traffic scenarios like road junctions, requiring DRL algorithms to rely on long-term rewards to train effective models. This reliance can introduce risks, particularly when agents explore new actions, potentially leading to collisions and penalties. Although safety considerations are typically integrated into reward functions, they are not always the primary metric for evaluating the effectiveness of DRL algorithms in autonomous driving agents. This study addresses this gap by assessing the comprehensive safety performance of three baseline DRL models, DQN, A2C, and PPO, based on collision, success, and freezing rates. We also propose a self-awareness DRL model incorporating a multi-attention mechanism for autonomous driving agents. Based on experiments in intersection and roundabout scenarios, the proposed model demonstrates enhanced safety performance compared to the best baseline model, DQN.<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>