Pedestrian-Aware Deep RL-Based Decision-Making and Control Framework for Unprotected Left Turns

This paper addresses the challenges of guiding Autonomous Vehicles (AVs) through unprotected left turns at unsignalized, fully uncontrolled, chaotic intersections consisting of pedestrians and adversarial vehicles. These intersections present several challenges for AVs, including the absence of traffic signals or signs, the dynamic nature of the environment, including human interactions due to the presence of crosswalks, and the variability of intersection layouts. This study develops a containerized simulation environment based on a high-fidelity simulation tool, CARLA, and is made open-source at https: //github. com/faizansana/ intersection-carla-gym. This study proposes using a hierarchical Deep Reinforcement Learning (DRL) wherein the DRL policy sets the target speed for the low-level PID controllers used for actuation. To the best of our knowledge, this study stands as the pioneering effort in addressing pedes-trian interaction within a high-fidelity simulation environment, coexisting alongside adversarial vehicles within CARLA. It conducts a comprehensive comparative analysis of five model-free DRL algorithms in a hierarchical framework. It is observed that Recurrent PPO with a discretized action space has the highest success rate and lowest accident rate for unprotected left turns. However, all algorithms struggle to achieve acceptable success rates when pedestrians exist within the environment, emphasizing the importance of the high-fidelity simulation tool. The paper concludes by discussing potential extensions of the proposed system and future research directions.