Do you need help? Identifying and responding to pilots’ troubleshooting through eye-tracking and Large Language Model

Abstract

In-time automation support is crucial for enhancing pilots’ performance and flight safety. While extensive research has been conducted on providing automation support to mitigate risks associated with the Out-of-the-Loop (OOTL) phenomenon, limited attention has been given to supporting pilots who are actively engaged, known as In-the-Loop (ITL) status. Despite their active engagement, ITL pilots face challenges in managing multiple tasks simultaneously without additional support. For instance, providing critical information through in-time automation support can significantly improve efficiency and flight safety when pilots need to visually troubleshoot unexpected incidents while monitoring the aircraft’s flying status. This study addresses the gap in ITL support by introducing a method that utilizes eye-tracking data tokenized into Visual Attention Matrices (VAMs), integrated with a Large Language Model (LLM) to identify and respond to troubleshooting activities of ITL pilots. We address two primary challenges: capturing the complex troubleshooting status of pilots, which blends with normal monitoring behaviors, and effectively processing non-semantic eye-tracking data using LLM. The proposed VAM approach provides a structured representation of visual attention that supports LLM reasoning, while empirical VAMs enhance the model’s ability to efficiently identify critical features. A case study involving 19 licensed pilots validates the efficacy of the proposed approach in identifying and responding to pilots’ troubleshooting activities. This research contributes significantly to adaptive Human–Computer Interaction (HCI) in aviation by improving support for ITL pilots, thereby laying a foundation for future advancements in human–AI collaboration within automated aviation systems.