Learning a Swarm Foraging Behavior with Microscopic Fuzzy Controllers Using Deep Reinforcement Learning

摘要

This article presents a macroscopic swarm foraging behavior obtained using deep reinforcement learning. The selected behavior is a complex task in which a group of simple agents must be directed towards an object to move it to a target position without the use of special gripping mechanisms, using only their own bodies. Our system has been designed to use and combine basic fuzzy behaviors to control obstacle avoidance and the low-level rendezvous processes needed for the foraging task. We use a realistically modeled swarm based on differential robots equipped with light detection and ranging (LiDAR) sensors. It is important to highlight that the obtained macroscopic behavior, in contrast to that of end-to-end systems, combines existing microscopic tasks, which allows us to apply these learning techniques even with the dimensionality and complexity of the problem in a realistic robotic swarm system. The presented behavior is capable of correctly developing the macroscopic foraging task in a robust and scalable way, even in situations that have not been seen in the training phase. An exhaustive analysis of the obtained behavior is carried out, where both the movement of the swarm while performing the task and the swarm scalability are analyzed.