A neuro-evolutionary approach to control surface segmentation for micro aerial vehicles

This paper addresses control surface segmentation in micro aerial vehicles (MAVs) by leveraging neuro-evolutionary techniques that allow the control of a higher number of control surfaces. Applying classical control methods to MAVs is a difficult process due to the complexity of the control laws with fast and highly non-linear dynamics. These methods are mostly based on models that are difficult to obtain for dynamic and stochastic environments. Moreover, these problems are exacerbated when both the number of control surfaces increases and the model’s accuracy in determining the impact of each control surface decreases. Instead, we focus on neuro-evolutionary techniques that have been successfully applied in many domains with limited models and highly non-linear dynamics. Wind tunnel simulations with Athena Vortex Lattice show that MAV performances are improved in terms of both reduced deflection angles and reduced drag (up to 5%) over a simplified model in two sets of experiments with different objective functions. We also show robustness to actuator failure with desired roll moment values still attained with failed actuators in the system through the neuro-controller.