Learning to blindly follow hydrodynamic trails

Many aquatic animals can follow hydrodynamic trails by sensing and responding to flow signals. Despite numerous studies on this topic, an understanding of how this behavior can be enacted using feedback control strategies that require only local and instantaneous flow sensing remains elusive. Here, we apply deep Reinforcement Learning to solve the problem of following vortical wakes to their generating source. We find that the trained swimmer reaches the source of the wake by turning towards larger flow speed, and that the location of the flow sensor is crucial for successful trail following. Through analysis in a reduced order signal field, we map the sensor location to the stability of the controller in locating the source. Importantly, the sensory control strategy is generalizable to thrust and drag wakes of different Strouhal and Reynolds numbers and to 3D wakes. This work emphasizes the importance of both sensor location and sensor type and has implication on other source seeking control problems with traveling-wave characteristic.