Autonomous Autorotation of a Multirotor Vehicle Using Output Feedback Model Predictive Control with Online Learning

View Video Presentation: https://doi.org/10.2514/6.2021-0379.vid Autonomous autorotation of a multirotor vehicle using output model predictive control (MPC) is explored. A reduced order model (ROM) consisting of convolutional neural network based rotor models is used to predict the output dynamics of a high fidelity vehicle simulation. A base rotor model is learned offline using data created via Monte Carlo simulation of a blade element theory based model. The reduction in the state space allows the controller to function without knowledge of the inflow and blade flapping, which are difficult to measure or estimate in practice. Online learning is performed by adding an additional multi-layer perception network to the ROM, training it offline using vehicle simulation data, and feeding it online measurements. A generalized policy is used to constrain the control inputs within the MPC framework, and the derivatives needed to perform an iterative optimization technique are presented.