Deep learning for model identification and control design: a case study on an aerial soaring robot

The core topic is the study of the model of an underactuated soaring aerial robot that harvests energy from environmental airflow, and the analysis and simulation of its motion capabilities, evaluating the possibility of performing complex trajectories. Moreover, after gathering telemetry and gyroscope data, a model of the system by means of a deep learning approach is provided to beat the baseline of the dynamical model representation. Based on this, an efficient auto-regressive noise model is fit, thus providing a solid base for the subsequent control design. The latter part of the dissertation aims at exploiting deep-learned dynamical ensemble models to carry out an optimal control design via backpropagation, guaranteeing robustness to noise and initial condition perturbations, and offering an innovative viewpoint in control design starting from neural-learned dynamic twins of the physical plant.