Deep reinforcement learning for turbulent gust rejection on flexible wing

• Aeroelastic model is used to model the flexible wing and reinforcement learning is used to train a policy for turbulent gust rejection. • Three sets of observations and two actuator time constants are investigated. • Inclusion of current control surface deflection to the observation does not result in any significant performance improvement. • Observing the incoming gust velocity through an alpha probe significantly improves the performance of the controller. • Actuator with lower time constant is more responsive to the commands, resulting in better control performance. • Reinforcement learning policy out-performs PD and LQG controllers. In this paper, reinforcement learning is implemented to train a neural network-based feedback controller for turbulent gust rejection on a flexible wing. The aeroelastic model of the wing is modeled by coupling the unsteady vortex lattice method for unsteady aerodynamics with finite-element based structural dynamics. Thereafter, reinforcement learning via the proximal policy optimization algorithm is used to train a neural network to minimize gust-induced tip deflections by directly manipulating the control surface. Results from simulation show that the trained policies are able to reduce the mean-squared-error in the tip displacements compared to the open-loop responses. Additionally, observing the incoming gust magnitudes, measured by an alpha probe, significantly improves the performance of the controller by allowing it to take preemptive actions. The reinforcement learning policy is able to better adapt to and learn complex dynamics, resulting in better overall performance when compared to PD and LQG controllers.