Decentralized Autonomous Traffic Management through Corridor Networks

As Advanced Air Mobility (AAM) scales, centralized air traffic management faces bottlenecks in communication and responsiveness. This paper investigates whether decentralized, autonomous aircraft can effectively navigate complex, multi-corridor airspace networks using only local observations and learned coordination behaviors, rather than relying on global scheduling or centralized control.

The researchers utilize a Multi-Agent Reinforcement Learning (MARL) framework to train autonomous agents in a simulated corridor environment. The methodology introduces three key innovations: a rotation-invariant policy representation, curriculum-based training to ensure corridor conformance, and a decentralized observation model. Agents are trained in a single-corridor setting and then tested in a zero-shot manner on more complex network topologies featuring various merges, splits, and heterogeneous vehicle performance envelopes. The policy relies solely on local relative states of nearby aircraft and the agent's own progress within the corridor, rather than global network information.

The study finds that policies trained in simple, single-corridor environments generalize effectively to complex, multi-corridor networks. Despite the lack of centralized coordination, the agents successfully maintain corridor boundaries, manage inter-aircraft separation, and achieve efficient traffic flow. The authors report that tactical safety interventions (e.g., emergency collision avoidance) are required less than 5% of the time, even in congested scenarios, suggesting that learned decentralized behaviors can provide a robust foundation for scalable AAM traffic management.