Data-Driven Aviation Scenario Generation using Generative Adversarial Networks

The use of scenario-based methodologies for testing autonomous systems in virtual environments is becoming increasingly vital. These approaches are becoming integral to the verification and validation processes of ADS. By employing scenario- and simulation-based techniques, it is possible to significantly reduce both the costs and the time associated with these efforts. Traditional techniques called Knowledge-Driven Scenario generation, although effective, required significant manual efforts to be representative. That is why Data-Driven Scenario Generation (DDSG) has emerged as a crucial method, especially in the development of Automated Driving Systems (ADS), attracting significant attention for its potential to capture a wide range of situations these systems may encounter using real-life data. By leveraging data-driven scenario generation, real-life data can be parameterized and applied in aviation simulation test procedures, subsequently populating a scenario database that closely mirrors actual flight paths and edge cases. This method offers high efficiency, flexibility, and reliability by leveraging authentic datasets, augmenting expert-derived inputs, ensuring completeness, and representativeness. Research in the aviation field on this topic is nonexistent compared to ADS. This paper presents a framework for DDSG using Generative Adversarial Networks trained on real-world data to generate realistic aviation scenarios that aid in the verification and validation of autonomous flight systems. In our case, we demonstrate the approach's efficacy in an aircraft landing use case. Leveraging data-driven scenario generation to assist experts in deriving meaningful scenarios enhances the completeness of cases for the safety assurance of such systems.