Automated synthesis of scalable algorithms for inferring non-local properties to assist in multi-robot teaming

Abstract

Individual robots in multi-robot teams usually behave according to relatively simple and often deterministic rules. This behavioral regularity may allow for variations in one robot's behavior to provide useful information about the state of other robots in the team. Thus, coordinated motion between distantly separated robots could be achieved without direct communication. Unfortunately, the dynamics describing a collective multi-robot system will often be too complex to analytically derive the relationship between observed nearest-neighbor variations and environmentally driven changes in the behavior of remote robots. Artificial neural networks (ANNs) may be used to find this relationship within training data, but scalability of the approach requires that the resulting ANNs be functional even in teams with sizes not represented in the training data. To this end, we train a communication-free, localization ANN on one robot in a 3-robot team and show how it can be extended without re-training to larger teams. We test our approach in a distributed caging scenario where a chain of simulated robots searches for an object to encircle and executes a coordinated behavior, ideally with no communication, shortly after only one detects the object.