When the mob sees the light - distributed learning of phototaxis without local gradients using robot morphology

A typical phototactic strategy requires a local intensity gradient monitoring, either through fore-aft differential light sensing, or through temporal differentiation on the fly. Using their sheer size, schools and flocks can respond to minute gradients the small individual could not detect accurately on its own. For this, the swarm needs cohesivity (through an effective attraction or an alignment interaction) in order to maintain a continuously connected communication network. We show that a completely decentralized population of kilobots can collectively find a successful phototactic strategy even in the total absence of a local gradient-field. Using an unsupervised online learning algorithm, the robots can reach a consensus for a phototactic policy, while their communication network remains sparse, intermittent, and disconnected, allowing the individual to maintain a high degree of autonomy. Using an exoskeleton to shape the robots' morphology, the swarm can promote cohesion-through-collision at the phototactic destination. We find that the steady state of a successful strategy can give an insight to the ability of the learning process to converge.