Learning to school through local sensory information

Fish schooling is a widely observed phenomenon in nature, providing benefits to individual fish through collective behavior. Recent studies have shown that stable formations can emerge through passive hydrodynamic interactions in schools of flapping swimmers, without the need for active control. These passive interactions, however, only lead to stable formations when the swimmers are highly coordinated and similar. Differences in their flapping motion and/or size can destabilize these formations and increase the cost of transport for the school. Here, we use a deep reinforcement learning algorithm to control the fapping motion of uncoordinated swimming agents to achieve cohesive schooling formations in a reduced-order model. The swimmers objective is to minimize the cost of transport for the whole group while only sensing the local flow field. We then test the robustness of our controller to perturbations in the swimmers motion and the flow field. Our results contribute to a deeper understating of fish schooling in nature and have potential applications in the design of autonomous underwater vehicle systems.