Underwater acoustic networks enable critical applications including environmental monitoring, offshore infrastructure inspection, maritime security, and Autonomous Underwater Vehicle (AUV) operations. Their design, however, is fundamentally constrained by the acoustic propagation environment, characterized by severe bandwidth limitations, long and variable delays, strong Doppler effects, complex multipath, non-Gaussian noise, and the absence of widely accepted statistical channel models. These challenges significantly limit the effectiveness of traditional model-driven approaches. In terrestrial Radio Frequency (RF) systems, Artificial Intelligence (AI) has successfully enabled data-driven inference and control across the protocol stack. Motivated by these advances, recent work has begun exploring AI-driven underwater acoustic communications, yet adoption remains limited due to strong non-stationarity and environmental dependence. Using motivating examples based on real underwater data, we show that learned models exhibit limited generalization across environments. This paper presents a research roadmap toward AI-enabled autonomous underwater acoustic networks and sensing applications, identifying key challenges and outlining directions for robust, adaptive inference and control under stringent energy, computational, and environmental constraints.
Alshawabka et al. (Fri,) studied this question.