Aircraft Autopilot Behavior Cloning through Time-Series Approaching and Flaring Scenario Classification

Abstract

In this paper, we present an aircraft autopilot system focusing on a landing flight. The landing flight is composed of approaching and flaring scenarios in which different controls are required. We have presented two flight controllers based on deep neural networks. For the flight controllers, the current flight was estimated from cockpit image inputs through a proposed scenario classifier based on CNN. Due to a so-called chattering problem, however, the two flight controllers were repeatedly used. Since the speed and angle are determined by the controllers, the aircraft might fail in the landing flight. For this challenge, we propose to apply an LSTM block to the scenario classifier. The CNN with LSTM enables the aircraft to estimate the flight scenario taking time-series variation of image inputs and scenario outputs into account. In the simulation experiments, we show that the aircraft succeeds in the landing flight by switching the controllers correctly through the time-series scenario classification.