Multi-UAV Multi-Target Cooperative Interception via Multi-Agent RL with Reward Shaping and LLM-Based Curriculum Learning

ABSTRACT

Multi-UAV cooperative interception of multiple evasive targets is a critical problem in defense and security applications. We address the challenge of maximizing the interception success rate within a limited time horizon using a team of autonomous UAVs against maneuvering targets headed toward a protected base. We propose a multi-agent reinforcement learning (MARL) approach based on Proximal Policy Optimization (PPO) with centralized training and decentralized execution (MAPPO). Our method incorporates potential-based reward shaping to inject domain knowledge and a four-stage large-language-model-based (LLM-based) curriculum learning scheme to gradually increase task difficulty. A local greedy heuristic serves as a baseline for comparison. Simulation results in a high-fidelity environment demonstrate that the proposed curriculum-guided MARL significantly improves the target neutralization rate and reduces missed targets compared to the greedy baseline (approximately 81.1% vs 66.2% success rate). The learned policy exhibits stable training curves and coordinated UAV behaviors that efficiently distribute targets among team members. These findings highlight the benefit of combining reward shaping and LLM-based curriculum learning in MARL for complex multi-UAV multi-target interception tasks.