Safe Reinforcement Learning-Based Distributed MAC Protocol for Multi-AUV Underwater Acoustic Communication Network

With the rapid development of autonomous underwater vehicles (AUVs) technology, multi-AUV underwater acoustic communication networks (UACNs) have significantly expanded the coverage and application scenarios of traditional underwater sensor networks, becoming the key technology of Internet of Underwater Things (IoUT). However, the inherent limitations of underwater acoustic channels, such as long propagation delay, limited bandwidth, and high bit error rate, along with the dynamic characteristics of multi-AUV systems, pose significant challenges to the design of medium access control (MAC) protocols that can achieve high network throughput and ensure reliable and collision-free transmissions. Conventional handshake-based MAC protocols (e.g., SFAMA and DOTS) introduce excessive communication overhead and end-to-end delay (EED) in underwater environments. Existing learning-based scheduling MAC protocols (e.g., UW-ALOHA-QM and DCSM) exhibit strong adaptability but still suffer from severe packet collisions during the learning trial-and-error process, leading to slow convergence and significant performance fluctuations. To address these issues, this article proposes a safe reinforcement learning-based distributed MAC (SRDMAC) protocol. SRDMAC employs a spatial–temporal safety shielding mechanism to prune slots with high collision risk before scheduling, thereby achieving safe exploration. In addition, SRDMAC introduces an adaptive scheduling framework based on Lagrangian primal–dual optimization, which dynamically evaluates the comprehensive Lagrangian value of each slot by considering both rewards and collision costs. NS3 simulation results demonstrate that compared to existing MAC protocols, SRDMAC achieves faster convergence speed, higher packet delivery ratio (PDR), and superior network throughput.