Reinforcement Learning-Based Controller for Tiltrotor Aeroelastic Stability Augmentation (Presentation Only)

This study investigated the feasibility of using Deep Reinforcement Learning (DRL) for aeroelastic stability control of a Tiltrotor Aeroelastic Stability Testbed (TRAST) model. The DRL controllers use rotor swashplate inputs to minimize oscillatory wing root bending moments of the tilt rotor model. First, three DRL-based agents including Deep Deterministic Policy Gradient (DDPG), Twin Delayed Deep Deterministic Policy Gradient (TD3), and Soft Actor-Critic (SAC) were investigated to control the aeroelastic stability of the TRAST model throughout a wide range of airspeed including where the whirl flutter occurs. All three agents demonstrated the capability of stability augmentation while the SAC agent demon-strated the most robust performance. Next, the effectiveness of the SAC agent was studied further by training the SAC agent at a certain airspeed and applying the trained agent through the TRAST whirl flutter conditions. Finally, additional tuning of the SAC agent was performed to improve performance further through a hyperparameter optimization framework called Optuna.