Correction: A Mobile Reinforcement Learning-Cyber-Physical Fluid Dynamics-based Flapping Wing Platform: Simulation Component

Animals and micro-aerial vehicles that use flapping wing kinematics exploit underlying, unsteady aerodynamics that are difficult to computationally and experimentally model because of their large, dynamic, and nonlinear parameter spaces. Cyber-physical fluid dynamic systems address the need for a high-resolution fluid model that captures these nonlinearities by using a physical fluid and an experimental apparatus to recreate computationally intensive fluid-structure interactions. Reinforcement learning has had success optimizing control tasks in a variety of simulated environments with large and dynamic parameter spaces. This paper presents the results of a simulated, mobile cyber-physical fluid dynamic platform that uses reinforcement learning to discover power-optimal flapping wing kinematics. For fast and efficient training, the simulation uses a low-fidelity lifting line solver and the ISAAC Gym RL package to simulate the aerodynamic forces and rigid-body kinematics of flapping flight. The training agent successfully controlled the wing stroke and the stroke plane angle to maximize its forward velocity and minimize its instability. The learned kinematics exhibited periodicity, a large-scale feature that characterizes efficient flapping flight. In future studies, the trained agent can be transferred to a cyber-physical system that mimics the simulated agent in form and functionality, on which it can continue training in a physical fluid to refine its control policy and discover the small-scale features of the optimal kinematics. The combined simulation and cyber-physical platform could provide insight into the power-optimal flapping kinematics of a variety of rigid wings and the fluid dynamics responsible for their efficiency.