Text2Traffic: Retrieval-Enhanced In-Context Learning for Complex Air Traffic Scenario Generation

Efficient and customizable air traffic scenario generation is crucial for advancing research and enabling the evaluation of advanced air traffic management (ATM) concepts, as well as supporting the strategic development of future ATM systems. Traditional methods based on modifying historical data, creating randomized traffic, or hand-crafting specific scenarios are time-consuming and offer limited customizability in tailoring specific research or operational needs, such as testing novel ATM concepts or human-in-the-loop experiments involving complex air traffic scenarios. This paper proposes Text2Traffic, a novel methodology that leverages large language models and retrieval-augmented generation to dynamically generate traffic scenarios for high-fidelity air traffic simulations based on natural language descriptions. The methodology enables rapid, iterative customization of traffic parameters—such as aircraft density, routing, separations, aircraft types, and operational constraints—to generate complex air traffic scenarios. Experimental evaluations compared Command R, Llama3.1-8b-Instruct, and GPT-4o-mini across prompts of varying complexity. Command R achieved the best performance, with 100% syntactic accuracy and the highest semantic accuracy. It successfully generated a complex traffic scenario with an overall complexity score of 0.13, characterized by dense horizontal interactions (0.41). These results demonstrate that the method is capable of constructing intricate, realistic scenarios through sequential, natural language inputs, supporting flexible modification and contextual adaptation. By bridging natural language interaction with scenario generation, Text2Traffic offers a scalable and intuitive framework for ATM research, simulation, training, and operational planning, enhancing the pace and precision of scenario-based experimentation in future air traffic systems.