Aircraft Autopilot Composed of Scenario Classifier and Flight Controllers Based on CNNs for Landing Flights

This paper is dedicated to the realization of an aircraft autopilot. Therefore, the authors focus on landing flights, which are considered the most difficult due to the physical contact between the aircraft and the runway. During landing, the aircraft descends toward the runway and momentarily maintains a constant altitude just before touchdown to mitigate the impact. The former scenario is referred to as “descending,” while the latter is known as the “landing flare.” The pilot performs different maneuvers for the aircraft, lowering the nose in the descending scenario and raising it in the flaring scenario. In this paper, therefore, we propose an autopilot specifically for unmanned landing flights. The autopilot is composed of a scenario classifier and two flight controllers, selected based on the classified flight scenario. This enables the aircraft to perform appropriate maneuvers autonomously depending on the scenario. Since only images from the cockpit are used as inputs, both the scenario classifier and the flight controllers are developed using convolutional neural networks. In the experiments, the proposed autopilot is applied to an aircraft in a flight simulator, and its effectiveness for landing flights is evaluated based on the results.