Multi Agent Reinforcement Learning Based Swarm Navigation

Abstract

Swarm navigation in dynamic and complex environments represents a critical challenge for both decentralized control systems and dynamic obstacle avoidance. Therefore, this study presents a robust, decentralized Multi Agent Reinforcement Learning (MARL) model which uses a distributed-proximal policy optimization (PPO) algorithm and enables the navigation of swarms collectively in a decentralized manner, using a shared policy without explicit communication among agents. In order to train each agent to learn how to transform an 8 dimensional vector into a specific action from their perspective in a two dimensional arena (thus enabling them to take action in the arena), we represent the system as a decentralized Multi Agent Markov Decision Process (MAMDP) and use a multi component reward function that provides penalty against collisions, rewards agents for successful progress toward the goals, and dampens velocity to promote stability. Our simulation results show quick convergence during training, and our agents achieve high success rates in all randomized configurations of the environment. Furthermore, our agents are able to generalize well across new, unseen obstacles; therefore, they are able to maintain their ability to navigate efficiently and safely in previously unseen obstacle configurations. © 2026 IEEE.