Value-Decomposed Reinforcement Learning Framework for Taxiway Routing with Hierarchical Conflict-Aware Observations

Researchers have introduced Conflict-aware Taxiway Routing (CaTR), a novel reinforcement learning framework designed to optimize real-time multi-aircraft taxiway routing and enhance on-surface conflict avoidance in airport operations. Addressing the limitations of existing planning methods, which often suffer from high computational costs, and standard reinforcement learning approaches that struggle with downstream traffic conflicts, CaTR employs a grid-based environment with action masking. The framework features a hierarchical foresight traffic representation to encode current and future conflict-related conditions, alongside a value-decomposed strategy to prioritize safety-critical objectives. Validated through experiments simulating Changsha Huanghua International Airport under various traffic densities, CaTR demonstrates superior safety-efficiency trade-offs compared to representative baselines while maintaining practical runtime performance. This advancement offers a promising solution for improving the safety and efficiency of complex airport surface operations.