Learning to form cohesive schooling formations using local sensory cues

Hydrodynamic interactions govern group formations in nature, notably in birds flocks and fish schools. Recent studies have shown that stable formations can emerge passively - without the need of active control - in flapping swimmers when they exhibit closely matched flapping kinematics. In natural settings, however, fish rarely maintain the same flapping motions, due to environmental variability, sensory limitations, and biological differences. This raises the question of how fish are able to control their motion and maintain stable formations despite having limited information about their neighbors. To address this, we employ a deep model-free reinforcement learning algorithm to control the flapping motion of a flapping swimmer following another. The follower's objective is to maintain a stable distance relative to the leader, relying solely on local flow information provided by the leader's wake. We will present our findings on the control policy learned through this approach and explore their potential implications for fish schooling and the design of autonomous underwater vehicles.