Abstract
As the demand for skilled pilots continues to grow, the need for innovative approaches to pilot training has become increasingly urgent. Virtual flight instructors, when integrated into flight simulators, have promised the best of all worlds: tailored feedback adapted to the student, and self-paced access to learning materials without need for human instructor oversight. Current systems provide detailed analysis by making use of state-based architectures. Development requires engineers and pilot subject matter experts to manually design, implement, and script each training scenario and the corresponding feedback. This manual approach creates excellent demo capabilities, but limits the scalability, flexibility, and affordability of these virtual instructors to meet full syllabi or address more complex scenarios.
In order to address these challenges, this paper presents PilotGPT, a simulator agnostic virtual instructor testbed built upon emerging commercial and open-source Large Language Models (LLMs). The prototype application ingests raw simulator data and generates detailed analysis and feedback of student pilot performance. PilotGPT leverages techniques such as multi-agent architecture, chain prompting, and retrieval augmented generation (RAG). The RAG mechanism searches a comprehensive database of vectorized flight information including aircraft manuals, FAA documents, mission data, and other relevant flight data. This approach more closely mimics real-life instructors by focusing on embedding comprehensive understanding of the aircraft and mission sets into the virtual instructor, instead of relying on rote evaluation against defined rules. This paper will review how various design trades made during the development of PilotGPT impacted the measures of reliability, quality of feedback generated, and scalability to other applications.