Towards Naturalistic Human-Machine Teaming With LLM Agents: A Case Study in Air Traffic Management

Natural language is the dominant form of interaction modality in Air Traffic Management (ATM), currently operated by human controllers. However, most AI research in this domain uses low-level interfaces that are disconnected from how controllers communicate. This paper introduces an LLM-manned team framework: multiple large language model (LLM) agents assuming specialized ATM roles that collaborate via natural language within the open-source BlueSky simulator environment. We instantiate two roles: (1) a communication agent that translates controller-style natural language commands using retrieval-augmented generation, and (2) a deconfliction surrogate fine-tuned with reinforcement learning to propose collision-avoidance maneuvers in horizontal conflict scenarios. The interface role achieves 71% translation accuracy, while the deconfliction agent achieves a success rate up to 96.1% from 65.9% compared to baselines. These results demonstrate the promising potential for LLM-manned teams to serve as decision-support tools and testbeds for human-machine teaming in ATM.