Pursuit-evasion using reinforcement learning

Abstract

In pursuit-evasion problems, we are presented with one or more pursuers attempting to capture one or more evaders. We consider the same problem in a bounded and convex physics-based environment. In our case, the agents are limited by obstacles, the field of view and a maximum speed limit in the 2D-plane. We propose a Voronoi partition, distance-based cooperative pursuit in which the pursuers try to minimise the distance, the area of the Voronoi region possessed by the evader to trap it in a corner. The agents are trained using Proximal Policy Optimization (PPO) and curriculum learning. By Multi-Agent Asynchronous Training we attempt to train multiple models parallelly to improve the training process. We simulate this 3D- environment using Unity.