Role-Structured Reinforcement Learning for Heterogeneous Airship-Guided UAV Pursuit Tracking

This paper investigates heterogeneous aerial pursuit tracking with one high-endurance airship, multiple UAV pursuers, and evasive targets. The task is formulated as a partially observed mixed cooperative-adversarial game under centralized training and decentralized execution. A role-difference-potential multi-agent reinforcement learning framework is proposed, in which each agent maintains an individual actor and critics are shared at the role level with identity conditioning. To improve cooperative credit assignment, pursuer optimization combines potential-based shaping with counterfactual one-step difference rewards. The potential function integrates coverage guidance, approach pressure, capture-neighborhood entry, encirclement quality, and safety regularization; actions are parameterized by bounded speed and yaw rate for kinematic consistency. Simulation results show stable coordinated closure and reliable capture, indicating that role-structured entropy-regularized learning with difference-potential rewards is a practical baseline for heterogeneous pursuit tasks.