Semantic‐Aware Multi‐Agent Reinforcement Learning Toward Adaptive and Resilient Industry 5.0 Systems

ABSTRACT The Industrial Internet of Things (IIoT) and the emerging Industry 5.0 paradigm demand intelligent systems that are not only efficient and autonomous, but also adaptive, resilient, and semantically aware. In such environments, large‐scale industrial spatial systems—ranging from smart factories and logistics infrastructures to cyber–physical production networks—exhibit complex interactions, dynamic topology changes, and strong coordination requirements among distributed agents. To address these challenges, this paper proposes a semantic‐aware multi‐agent reinforcement learning framework that tightly couples graph‐based semantic encoding with cooperative policy optimization. Unlike conventional MARL architectures that treat perception and control as loosely connected components, the proposed framework introduces a bidirectional semantic–decision structure in which attention‐based graph reasoning directly guides actor–critic updates, enabling semantically grounded coordination rather than purely reward‐driven learning. Extensive evaluations on Matterport3D, SMACv2, and POGEMA demonstrate faster convergence, improved stability, and stronger generalization compared with VDN, QMIX, and MADDPG. The explicit modeling of inter‐agent relevance and multi‐scale abstraction enhances interpretability by exposing structural decision factors, improving transparency and robustness in dynamic environments. These results highlight the potential of semantic‐aware multi‐agent learning for next‐generation IIoT and Industry 5.0 systems.