A recurrent neural network for nonlinear control of a fixed-wing UAV

Abstract

Application of echo state network (ESN) for the nonlinear control of a fixed-wing unmanned aerial vehicle (UAV) is presented. The data required for the network training is generated using a validated flight dynamics model of the UAV. The ESN is used for both offline and online training. The offline training realizes the inversion required for the feedback linearlization. The online training is then used to reduce the inversion error that results due to the modeling deficiencies. The data required for the online training is obtained in real-time from the FlightGear flight simulator while the UAV is flying in the simulation. The FlightGear model of the UAV is based on the flight validated nonlinear model of the UAV. Results show that the ESN performs very well for the nonlinear control of the UAV with significantly less computational time and simplicity.