Learning to Fly: Upgraded Aerodynamics and Control Surfaces

Abstract: In recent times the unmanned quadcopter aircraft has been used for a widening range of applications, but for longer distances it still falls short of conventional airplanes in terms of energy usage. There exists hybrid configurations of unmanned aircraft which combine the mobility of quadcopters and the range of fixed-wing aircraft. The transition between the hovering mode and the gliding mode, however, is a complex non-linear control problem. To solve this a recent study applied a neural network as a closed loop controller. This controller was capable of seamless mode transition and could be trained for any copter configuration using reinforcement learning. The work presented here focuses on improvements to the method of controller design established by said study, mainly focusing on increased realism of the aerodynamic simulations and the addition of control surfaces for increased maneuverability. These improvements resulted in a lift of 37% of the total copter weight and a higher achievable top speed of 8 m/s before instability occurs. To verify these improvements were not only present in the simulations a physical prototype was also constructed which when tested succeeded in hovering flight but failed to sustain any significant forward flight.