Collective Behavior Acquisition of Real Robotic Swarms Using Deep Reinforcement Learning

Abstract

Swarm robotic systems are a type of multi-robot systems, in which robots operate without any form of centralized control. The most popular approach for SRS is the so-called ad hoc or behavior-based approach; desired collective behavior is obtained by manually by designing the behavior of individual robot in advance. On the other hand, in the principled or automatic design approach, a certain general methodology for developing appropriate collective behavior is adopted. This paper investigates a deep reinforcement learning approach to collective behavior acquisition of swarm robotics systems. Robots are expected to collect information in parallel and share their experience for accelerating the learning. We conduct real swarm robot experiments and evaluate the learning performance in a scenario where robots consecutively travel between two landmarks.