Design of Low-Altitude Air Route Networks with Robustness Boundary via Reinforcement Learning

The rapid expansion of the low-altitude economy is reshaping urban transportation systems, while large-scale operations of dynamic and high-density low-altitude unmanned aerial vehicles pose significant challenges to performance and robustness of airspace infrastructure. Traditional free-flight and point-to-point paradigms have revealed inherent limitations in conflict resolution and congestion mitigation, making the construction of adaptive, structured, and dynamic low-altitude air route networks a critical pathway to achieve both route controllability and efficient operation. However, existing design approaches may struggle to rapidly adapt to dynamic flight requirements while ensuring high system performance. Increasing demands for robustness further complicates design of air route networks, necessitating a trade-off between performance and robustness. To address this coupled challenge, a two-stage design framework is proposed based on safe reinforcement learning (safe RL), enabling the automated construction of low-altitude air route networks with high performance and permissible robustness. The framework first constructs an initial backbone network using the shortest path sets derived from origin-destination (OD) demands to guarantee basic accessibility. Then, the initial backbone network is augmented by adding a given number of edges, which is formulated as a constrained Markov decision process (CMDP). By integrating the representation capability of graph neural networks (GNNs) with the constraint-handling mechanism of safe RL, the framework achieves adaptive network design that improves system travel performance under the robustness constraint. Experimental results in Washington demonstrate that the proposed method can effectively design air route networks across different OD scenarios. Embedding the robustness constraint into the reinforcement learning (RL)-based design paradigm, this approach provides a potential pathway for the automated design of next-generation critical infrastructure for low-altitude transportation with high performance and permissible robustness.