CaTR: Reinforcement Learning Framework for Multi-Aircraft Taxiway Routing
A novel reinforcement learning approach named Conflict-aware Taxiway Routing (CaTR) has been developed for the real-time routing of multiple aircraft on taxiways. This framework creates a grid-based model of the airport's surface environment featuring action masking, utilizes a hierarchical foresight traffic representation to capture both current and future conflict-related traffic scenarios, and implements a value-decomposed reinforcement learning technique to emphasize sparse yet crucial safety objectives. Testing was carried out in a realistic setting modeled after Changsha Huanghua International Airport, across various traffic densities. This study tackles the intertwined safety-critical challenges of taxiway routing and conflict avoidance on airport surfaces, addressing the limitations of existing planning methods and the difficulties reinforcement learning faces in managing downstream traffic conflicts and balancing multiple objectives.
Key facts
- CaTR is a reinforcement learning framework for taxiway routing.
- It uses a grid-based airport surface environment with action masking.
- Hierarchical foresight traffic representation encodes current and downstream conflicts.
- Value-decomposed reinforcement learning prioritizes safety-critical objectives.
- Experiments were conducted at Changsha Huanghua International Airport.
- The framework addresses coupled safety-critical decision problems.
- Existing methods are limited by online computational cost.
- Reinforcement learning methods struggle with downstream traffic conflicts.
Entities
Locations
- Changsha Huanghua International Airport
- China