Understanding Training Components in Shaping AI for Loss of Control Recovery in Unforeseen Flight Conditions

This paper works towards an AI expert to mitigate the human-automation gap encountered by pilots in commercial aviation, a challenge that is increasing by the growing integration of automation systems in the cockpit. AI is proposed to create an expert system that would be able to assist the pilot by providing timely and contextually relevant information or providing direct control inputs when necessary. For this, the AI system requires a comprehensive understanding of the interconnection of various automation systems in the cockpit environment. This paper works towards such an AI system by investigating fundamental training aspects that lead to predictable and reliable output in the face of unforeseen circumstances. Reinforcement learning is implemented with the aim of understanding the aircraft environment and thereby enabling adaptation to unseen changes to aircraft dynamics, and the ability to perform upset recovery. Multiple agents were trained using different training spaces to facilitate comparative analysis. Subsequent testing assessed their adaptability to variations in aircraft weight, changes in the center of gravity, and stall recovery. The findings indicated a significant shift in the understanding gained by the agents, enabling satisfactory performance beyond their training space. Furthermore, agents demonstrating this adaptability were also capable of executing stall recovery maneuvers.