Multi-robot Coverage Path Planning (mCPP) faces critical challenges in achieving effective coordination and maintaining robustness to robot failures, with existing methods either relying on inflexible area partitioning or generating inefficient overlapping paths. This paper presents a Graph Neural Network (GNN) based distributed planning framework that enables real-time coordination of multiple robots while maintaining resilience to robot failures. Our approach combines hierarchical map representation with spatially constrained communication topology to enable each robot to learn effective coverage strategies from its local observations and neighboring robots’ states. Through imitation learning, our online planning method achieves 96% coverage efficiency when an optimal solution exists for the centralized algorithm, while reducing redundant paths by 30% compared to the baseline. The system also demonstrates strong generalization to larger robot teams, adapts to different environments, achieving 94% coverage within finite time even when there is no optimal centralized solution. It sustains coverage efficiency despite failures, making it practical for real-world deployment.