Global path planning algorithm design using Deep Q-Networks and grid maps constructed from KML data for autonomous vehicles

Abstract

In modern autonomous vehicle algorithms, deep learning plays a vital role by enabling vehicles to process and interpret tremendous data, allowing for accurate decision-making and safe navigation in complex environments. Path planning is a crucial step during decision-making since it controls how vehicles navigate on the road. It includes two parts: global path planning, which considers the entire environment for a long-term route, and local path planning, which focuses on immediate surroundings for real-time navigation and obstacle avoidance. To achieve both environmental and economic benefits, it is essential to improve the efficiency and accuracy of the path planning algorithm. In this research, the primary goal is to improve global path planning by including fuel cost factors for autonomous vehicles. The algorithm used in this thesis is Deep Q-Network (i.e., using deep learning to implement Q-learning) with a modified reward function and a grid map environment model. The model includes speed limits, traffic lights, and four-way stop signs on the road to produce different travel time for different paths. By considering global path planning on the grid maps extracted and preprocessed from KML data, the new path planning design outputs effective results. These results show that the algorithm considers fuel cost factors, whereas the traditional A* algorithm only looks for the shortest distance. This study demonstrates DQN’s ability to find the optimal path on a real-life structured traffic road network while considering factors to minimize fuel consumption. While this research offers a novel method to use KML data for deep reinforcement learning, there remains significant scope for further investigation and refinement.