Grasper: A Generalist Pursuer for Pursuit-Evasion Problems

Pursuit-evasion games (PEGs) model interactions between a team of pursuersand an evader in graph-based environments such as urban street networks. Recentadvancements have demonstrated the effectiveness of the pre-training andfine-tuning paradigm in PSRO to improve scalability in solving large-scalePEGs. However, these methods primarily focus on specific PEGs with fixedinitial conditions that may vary substantially in real-world scenarios, whichsignificantly hinders the applicability of the traditional methods. To addressthis issue, we introduce Grasper, a GeneRAlist purSuer for Pursuit-EvasionpRoblems, capable of efficiently generating pursuer policies tailored tospecific PEGs. Our contributions are threefold: First, we present a novelarchitecture that offers high-quality solutions for diverse PEGs, comprisingcritical components such as (i) a graph neural network (GNN) to encode PEGsinto hidden vectors, and (ii) a hypernetwork to generate pursuer policies basedon these hidden vectors. As a second contribution, we develop an efficientthree-stage training method involving (i) a pre-pretraining stage for learningrobust PEG representations through self-supervised graph learning techniqueslike GraphMAE, (ii) a pre-training stage utilizing heuristic-guided multi-taskpre-training (HMP) where heuristic-derived reference policies (e.g., throughDijkstra's algorithm) regularize pursuer policies, and (iii) a fine-tuningstage that employs PSRO to generate pursuer policies on designated PEGs.Finally, we perform extensive experiments on synthetic and real-world maps,showcasing Grasper's significant superiority over baselines in terms ofsolution quality and generalizability. We demonstrate that Grasper provides aversatile approach for solving pursuit-evasion problems across a broad range ofscenarios, enabling practical deployment in real-world situations.