Pursuit-evasion with Decentralized Robotic Swarm in Continuous State Space and Action Space via Deep Reinforcement Learning

In this paper we address the pursuit-evasion problem using deep reinforcement learning techniques. The goal of this project is to train each agent in a swarm of pursuers to learn a control strategy to capture the evaders in optimal time while displaying collaborative behavior. Additional challenges addressed in this paper include the use of continuous agent state and action spaces, and the requirement that agents in the swarm must take actions in a decentralized fashion. Our technique builds on the actor-critic model-free Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm that operates over continuous spaces. The evader strategy is not learned and is based on Voronoi regions, which the pursuers try to minimize and the evader tries to maximize. We assume global visibility of all agents at all times. We implement the algorithm and train the models using Python Pytorch machine learning library. Our results show that the pursuers can learn a control strategy to capture evaders.