Predator-Prey Q-Learning Based Collaborative Coverage Path Planning for Swarm Robotics

Michael Watson, Hanchi Ren, Farshad Arvin, Junyan Hu

Abstract: Coverage Path Planning (CPP) is an effective approach to let intelligent robots cover an area by finding feasible paths through the environment. In this paper, we focus on using reinforcement learning to learn about a given environment and find the most efficient path that explores all target points. To overcome the limitations caused by standard Q-learning based CPP that often fall into a local optimum and may be in-efficient in large-scale environments, two methods of improvement are considered, i.e., the use of a robot swarm working towards the same goal and the augmenting of the Q-learning algorithm to include a predator-prey based reward system. Existing predator-prey based reward systems provide rewards the further away an agent is from its predator, the paper adapts this concept to work within a robot swarm by simulating each agent of the swarm as both predator and prey. Simulation case studies and comparisons with the standard Q-learning show that the proposed method has a superior coverage performance in complicated environments.