Control of Morphing Wing Shapes with Deep Reinforcement Learning
- Additional Document Info
- View All
© 2018 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved. Traditional model-based feedback control techniques are of limited utility for the control of many shape changing systems due to the high reconfigurability, high dimensionality, and nonlinear properties of the plant and actuators of these systems. Computational intelli- gence and learning techniques offer the promise of effectively leveraging the use of both smart materials and controls for application in aerospace systems such as morphing air vehicles. This paper addresses the challenge of controlling morphing air vehicles by developing a deep neural networks and reinforcement learning technique as a control strategy for shapememory alloy (SMA) actuators in the context of a morphing wing. The control objective is to minimize the error between an objective and the actual airfoil trailing edge de ection. The proposed controller is evaluated on a simple inverted pendulum for validation, on a 3D printed wing section that is actuated by a composite SMA actuator in a wind tunnel, and on a simulation based on wind tunnel data. Results show that the learning algorithm is capable of learning how to morph the wing. It is also able to control shape changes from arbitrary initial shapes to arbitrary goal shapes using the same trained learn- ing algorithm. The results provide a proof of concept for the use of learning algorithms to control more complex morphing aircraft with continuous states and actions for the outer mold line configuration.
author list (cited authors)
Goecks, V. G., Leal, P. B., White, T., Valasek, J., & Hartl, D. J.