Differential-Game-Guided Event-Triggered Multi-Task Multi-Agent Reinforcement Learning for Cooperative Pursuit under Partial Observability

This paper studies cooperative pursuit-evasion game in obstacle-rich 2-D environments, where line-of-sight occlusions and unknow evader strategies lead to incomplete information. An expert-driven multi-agent reinforcement learning (MARL) framework is proposed which utilizes Wavefront to represent the evaders’ potential motion region. The pursuit process is formulated as a stage-conditioned semi-Markov decision problem and integrates learning with expert knowledge in a unified loop. In the first stage, pursuers learn to achieve effective visibility occupation and coordinated coverage over the evaders’ reachability boundary. In the second stage, the policy organizes a shrinking search inside the Wavefront to improve search efficiency while reducing redundant exploration. In the third stage, an Apollonius Circle-Based pursuit policy is employed for efficient capture when lines of sight become available between pursuers and evaders, with adaptive switching between chase and search to enhance robustness under occlusions. The proposed pursuit strategy integrates expert policies as priors with MARL, offering a scalable paradigm for expert-augmented MARL in complex pursuit–evasion scenarios, and its effectiveness is evaluated in simulation.