Alternating Learning for Modular Sensorimotor Control of a Flapping Wing Unmanned Aerial Vehicle

Abstract

Abstract. This paper introduces a data-driven sensorimotor control framework for a flapping-wing unmanned aerial vehicle (FWUAV). It integrates an imitation learning algorithm for optimal controls with a deep neural pose estimation scheme. Recognizing that a direct concatenation of the neural pose estimator with the learning-based controller fails, we propose an alternating learning algorithm, namely, ALICE, for the coordinated integration of the two learning schemes. In particular, we enhance the learning capability of the estimator and the controller such that they converge to a synergistic pair. The proposed framework demonstrates excellent stabilizing capabilities compared to alternative ablated strategies or even an end-to-end controller. Furthermore, the presented technique overcomes the common restrictions of existing methods for FWUAV control, particularly the requirement for high-frequency flapping to justify linearization over averaged dynamics.