A Survey of Research and Applications of Optimal Path Planning Based on Deep Reinforcement Learning

University College London, Electrical and Electronic Engineering Department, WC1E 6BT, United Kingdom

^* Corresponding author: Zceeonb@ucl.ac.uk

Abstract

This study focuses on the comparison between conventional path planning techniques and Deep Reinforcement Learning-based path planning technology. Traditional algorithms including the A* algorithm, Dijkstra algorithm, and random sampling techniques, which perform well in static contexts, become inefficient in dynamic and complicated high-dimensional settings due to high computing costs. In contrast to DRL, a robot is charged with detecting sensor data linked with optimum path-planning using MDP (Markov Decision Process). Such robots independently study their surroundings while choosing the proper path. This implies it gets better at adapting to new settings and changes. This research looks at the real-time performance of three standard DRL algorithms: Proximal Policy Optimization, Deep Deterministic Policy Gradient, and Deep Q-Network in real-world settings. This work also discusses the drawbacks of DRL in path planning, i.e., high processing requirements, extended training times, weak generalization capabilities, and so on. Future initiatives include creating efficient training algorithms, mechanisms that increase model generalization, and optimizing them using classical techniques.