Evolution of Signaling in a Multi-Robot System: Categorization and Communication

Abstract

Communication is of central importance in collective robotics, as it is integral to the switch from solitary to social behavior. In this article, we study emergent communication behaviors that are not predetermined by the experimenter, but are shaped by artificial evolution, together with the rest of the behavioral repertoire of the robots. In particular, we describe a set of experiments in which artificial evolution is used as a means to engineer robot neuro-controllers capable of guiding groups of robots in a categorization task by producing appropriate actions. The categorization is a result of how robots' sensory inputs unfold in time, and, more specifically, of the integration over time of sensory input. In spite of the absence of explicit selective pressure (coded into the fitness function), which would favor signaling over non-signaling groups, communicative behavior emerges. Post-evaluation analyses illustrate the adaptive function of the evolved signals and show that these signals are tightly linked to the behavioral repertoire of the agents. Signals evolve because communication enhances group performance, revealing a “hidden” benefit for social behavior. This benefit is related to obtaining robust and fast decision-making mechanisms. More generally, we show how processes requiring the categorization of noisy dynamical information might be improved by social interactions mediated by communication. In a further series of experiments, we successfully download evolved controllers onto real s-bots. We discuss the challenges involved in porting neuro-controllers displaying time-based decision-making processes onto real robots. Finally, the beneficial effect of communication is shown to transfer to the case of a real robot, and the robustness of the behavior against inter-robot differences is discussed.