Morphing control of a new bionic morphing UAV with deep reinforcement learning

Abstract With rapid development of aviation technology, materials science and artificial intelligence, aircraft design is pursuing higher requirements both in civil and military fields. The new generation of aircraft should have the autonomous capable of performing a variety of tasks (such as take-off and landing, cruising, maneuvering, hover, attack, etc.) under a highly variable flight environment (height, Mach number, etc.) and meanwhile maintaining good performance. Morphing aircraft can use smart materials and actuators to autonomously deform the shape according to the changes in flight environment and mission, and always maintain an optimal aerodynamic shape, therefore get flourished developments. Based on the ability of birds to stretch wings when flying at low speed and to constrict wings at high speed, a new bionic morphing UAV has been designed and developed as the study model by our team. In order to make this new aircraft be able to complete rapid autonomous morphing and aerodynamic performance optimization under different missions and flight conditions, we developed deep neural networks and reinforcement learning techniques as a control strategy. Considering the continuity of the state and action spaces for model, the Deep Deterministic Policy Gradient (DDPG) algorithm based on the actor-critic, model-free algorithm was adopted and verified on the classic nonlinear Pendulum model and Cart Pole game. After the feasibility was verified, morphing aircraft model was controlled to complete prescribed deformation using DDPG algorithm. Furthermore, on the condition that the DDPG algorithm can control morphing well, through training and testing on model using simulation data from wind tunnel tests and actual flight, the autonomous morphing control for the shape optimization of the bionic morphing UAV model could be realized.