Swarm AGV Optimization Using Deep Reinforcement Learning

Abstract

Behavior design for Automated Guided Vehicles (AGV) systems is an active research area, fundamental for robotics, industrial systems automation. The rise of machine learning neural systems and deep learning make promising results in a multitude of areas including warehouse environments.In this paper, several different policies will be obtained by using reinforcement learning on a heterogeneous swarm robotic system, applied for solving logistical tasks in Automated Guided Vehicles. More specifically, two different types of agents will be used: the vehicles that collect, transport and deposit their package and the traffic lights that regulate the number of vehicles that circulate on the tracks. The main objective of our work is to learn simultaneously two different control policies, one for each kind of agent.The obtained policies have shown their ability to correctly learn the package transport behavior in addition to balance traffic flow to facilitate agent mobility and avoid collisions. Furthermore, the scalability of the system and the behavior performance for different number of vehicles has been shown.