Bio-inspired versus Machine-learned Adaptations to Propulsor Damage

Natural systems of flapping propulsion display a remarkable ability to adapt to significant propulsor damage, maintaining both lift and control authority via alterations to the propulsor trajectory. To employ this desirable trait in robotic systems, one may attempt to exactly mimic these alterations. However, it is not known whether these alterations are the most efficient adaptations to damage. Biological systems are subject to additional evolutionary pressures that may not be relevant to the optimality of robotic flapping propulsion, and in addition, biological systems generally function within more limited parameter ranges. With a larger actuation search space, optimizing for efficiency alone, a robotic propulsor may adapt in a different way than natural systems to propulsor damage. In this work, we seek optimal trajectories for damaged flapping propulsors, in order to determine whether bio-inspired strategies to adapt to significant propulsor damage are indeed the most efficient. Experimental function evaluations are performed by a flexible propulsor actuated by a spherical parallel manipulator (SPM). From these evaluations, the optimal trajectory is determined via a Covariance Matrix Adaptation Evolutionary Strategy (CMA-ES). During the optimization, a portion of the propulsor is intentionally removed. How the machine learning system adapts to this damage is then compared to existing data on natural swimmers and flyers.