W. Laurito, R. Titze, W. Schiffmann

The fatality rate for flights involving light, fixed-wing aircraft is relatively high, and most accidents occur after a loss of thrust in the lower airspace. In these emergency situations, with a lack of potential energy, an assistance or automated system could lead the aircraft safely to an appropriate landing field. Some solutions for path planning and guidance exist, however most of them rely on a simplified model of the environment and the aircraft´s dynamics to generate a path. Thus, those solutions tend to be rigid and less reliable during an emergency; especially in the occurrence of wind. Moreover, many solutions don´t consider the expected landing direction and the heading of the aircraft has to have, when reaching the landing field. In this work, we tackled these issues by focusing on the creation of a real-time guidance system for 3D trajectory planning after a loss of thrust based on deep reinforcement learning (DRL). In DRL, an agent learns through trial and error by interacting with an environment. DRL is especially useful in uncertain environments, where many parameters can´t be calculated in advance, which is the case in an emergency. Therefore, to train the agent to guide a fixed-wing, engines-off aircraft to an arbitrary target position, we developed and implemented multiple simulation environments. Furthermore, we incorporated wind into one of these environments. The created software package of the environments can be found online 1. Usually, complex calculations are needed to model the engines-off flight dynamics and to generate a 3D path (guidance) under wind. With DRL these calculations can be avoided. By using shaped reward functions, we trained a neural network to successfully select directions and glide angles to lead the aircraft to an arbitrary, chosen landing field in real-time while avoiding accidents. The success rate during our experiments was high: In most cases, the aircraft reaches the target position from the correct direction and with the expected heading.

Deutscher Luft- und Raumfahrtkongress 2022, Dresden

Deutsche Gesellschaft für Luft- und Raumfahrt - Lilienthal-Oberth e.V., Bonn, 2023

Conference Paper

englisch

21,0 x 29,7 cm, 32 Seiten

urn:nbn:de:101:1-2023053112260278129209

10.25967/570497

Deep Reinforcement Learning, Aviation, Flight Guidance, Emergency Landing, Trajectory Planning

Download - Bitte beachten Sie die Nutzungsbedingungen dieses Dokuments: Copyright protected

Laurito, W.; Titze, R.; Schiffmann, W. (2023): Emergency Pilot: Automated Flight Guidance After a Loss of Thrust Based on Deep Reinforcement Learning. Deutsche Gesellschaft für Luft- und Raumfahrt - Lilienthal-Oberth e.V.. (Text). https://doi.org/10.25967/570497. urn:nbn:de:101:1-2023053112260278129209.

31.05.2023