Reinforcement Learning Based Parking Space Egress for Autonomous Driving 2024-26-0088

Automated parking systems for cars have become the need of the hour globally, gaining wide acceptance from customers, and hence OEMs are working towards achieving precise/accurate automated parking. Various algorithms are being developed to plan the trajectory of the vehicle to be moved in/out of the desired parking slot. Most of these algorithms assume a static environment and don’t account for highly dynamic objects. Accounting for such objects is vital especially when autonomously exiting a parking slot and merging with traffic. This paper summarizes our initial efforts in addressing dynamic objects, specifically the ‘right of way’ aspects, while autonomously exiting a parking slot. In this study, we propose a novel approach for generating linear and angular velocity profiles using Deep Reinforcement Learning (DRL) in conjunction with Hybrid A* path planning for autonomous vehicles (AVs) navigating parking maneuvers. The aim is to address challenges faced by traditional Model Predictive Control (MPC) methods in trajectory planning, such as the lack of consideration for the right of way of the traffic participant. The proposed DRL algorithm employs a reward function that considers safety, path efficiency, and right of way aspects. The proposed system was trained using Reinforcement Learning toolbox in Matlab and tested using the Automated Driving toolbox. The results are then presented for parallel park-out case which show the effectiveness of the proposed solution.