Learning Aircraft Pilot Skills by Adversarial Inverse Reinforcement Learning

The aging population in Japan is causing a shortage of pilots in the aviation industry. This is expected to continue, leading to a severe lack of human resources in the future. The shortage will impact airlines’ operations and pilot training. Since training is costly and time-consuming, efficient pilot training is crucial. To address these future problems, incorporating artificial intelligence tools to support instructor pilots and trainees is an effective solution. In this study, we applied Adversarial Inverse Reinforcement Learning (AIRL), a practical and scalable method for imitating expert operations, to solve a landing problem for aircraft, which is an important and difficult task in training as well as actual operations. To obtain the operation model of expert pilots, we developed a new framework based on AIRL that incorporates multimodal information. This framework allows us to build an operation model with structures similar to human recognition systems, which can infer physical information. When expert pilots try to land, it is required to keep the aircraft states desired values under changing circumstances from time to time. Before learning a model of expert pilots, we analyzed time series data of expert pilots to find out their skilled control techniques. Our learning results demonstrate the effectiveness of our framework and the successful acquisition of expert piloting skills. By learning AIRL based on multimodal information developed in this study, a model that can generate maneuvers satisfying all the research objectives (constraints of speed, glide path angle, and touchdown point) was obtained. We also aimed to improve the performance of flare operations just before touchdown by conducting pre-training.