HRL-CoMindNet: A Hierarchical Framework for Intent-Based Communication and Coordination

This paper introduces Hierarchical Reinforcement Learning with Collaborative Mind Network (HRL-CoMindNet), a novel framework for multi-agent systems that enhances communication and coordination through a hierarchical design. At its core, HRL-CoMindNet decouples high-level coordination from low-level execution, shifting the focus of communication from raw observations to shared intentions. This intent-driven approach promotes deeper mutual understanding and synchronized behavior among agents. To realize this paradigm, we design a dedicated communication engine consisting of a Collaborative Mind Network (CMN), which predicts intentions of other agents, and a Message Passing Neural Network (MPNN), which propagates and refines these predictions into a coherent shared mental model. To enhance robustness and practical deployment, HRL-CoMindNet incorporates two key mechanisms: an Echo State Network (ESN) that captures rich temporal features for more accurate intention forecasting, and an adaptive trust mechanism that allows agents to dynamically weigh received intentions based on reliability, increasing resilience to communication failures. The framework aims to offer a predictive, reusable, and generalizable communication capability. The experimental results demonstrate that in the target tracking task, the system achieves a target coverage rate of 76.31% under ideal communication conditions; even with a 50% message dropout rate, it maintains 58.44% coverage, reflecting the effect of maximizing coverage. Ablation studies highlight the essential roles of the ESN, CMN, and MPNN in improving coordination performance and system robustness.