Limited bandwidth recognition of collective behaviors in bio-inspired swarms

Abstract

Models of swarming and modes of controlling them are numerous; however, to date swarm researchers have mostly ignored a fundamental problem that impedes scalable human interaction with large bio-inspired robot swarms, namely, how do you know what the swarm is doing if you can't observe every agent in the collective? We examine swarm models that exhibit multiple collective motion patterns from the same parameters. These multiple emergent behaviors provide increased expressivity, but at the cost of uncertainty about the swarm's actual behavior. Because bandwidth and time constraints limit the number of agents that can be observed in a swarm, it is desirable to be able to recognize and monitor the collective behavior of a swarm through limited samples from a small subset of agents. We present a novel framework for classifying the collective behavior of a bio-inspired robot swarm using locally-based approximations of a swarm's global features. We apply this framework to two bio-inspired models of swarming that exhibit a flock and torus behavior and a swarm, torus, and flock behavior, respectively. We present both a formal metric of expressivity and a classifier that leverages local agent-level features to accurately recognize these collective swarm behaviors while sampling from only a small number of agents.