Adaptive Learning by Robot Swarms in Unfamiliar Environments

Abstract

Although the goal of this book is to design swarms methodically, it is not always possible to theoretically determine the values of all parameter settings a priori. In this situation, the behavior of the swarm can be improved via optimization of the parameters. Traditionally, the optimization of swarm behaviors occurs offline (i.e., in simulation prior to actual task execution), and the swarm must be retrained whenever environmental conditions change. However, online adaptation (i.e., during the process of task execution) is much more practical. Offline approaches are unrealistic because, in general, the environment cannot be anticipated a priori due to numerous factors. In this chapter, we demonstrate the utility of an online swarm adaptation technique, called Daedalus, that is combined with physicomimetics. The task addressed here is that of a swarm of robots that needs to learn to avoid obstacles and get to a goal location. In our approach, swarms exhibit fault-tolerance, adaptability and survivability in unfamiliar environments. The success of our approach is greater than that of previous approaches that have been applied to this task.