A Review of Adaptive Sampling Technologies and Multi-Glider Cooperative Applications for Underwater Gliders

To track ocean phenomena whose locations and structures evolve over time, underwater gliders must adapt their sampling missions in response to observational feedback. Existing reviews have largely focused on vehicle control or path planning, with limited synthesis of the adaptive sampling process, the functional scope of single-glider methods, and the mechanisms of multi-glider coordination. This review uses the adaptive sampling feedback loop as its organizing principle to examine how observations are translated into subsequent sampling actions and to synthesize system architectures, evaluation criteria, and tiered validation approaches. Based on their dominant roles in the sampling decision process, single-glider methods are classified into rule-based triggering and feature tracking, information-driven active sampling, and model-based forecasting and constrained path planning. These categories focus, respectively, on responding to local observations, assessing the value of candidate sampling locations, and generating executable paths. Reinforcement learning is treated as an auxiliary approach that can be embedded across these stages. Multi-glider coordination is categorized as centralized, distributed, or hybrid according to decision location and information exchange, and the deployment requirements for formation control, coverage sampling, and cooperative observation with heterogeneous platforms are examined. Across applications, observations of mesoscale eddies and ocean fronts emphasize estimating feature locations, identifying boundaries, and adjusting sampling paths. Thermocline sampling relies mainly on identifying the target layer from vertical temperature gradients and adjusting the profiling range. Internal-wave observations depend more strongly on deployment location, sampling schedule, and array configuration, whereas plume sampling focuses on anomaly detection and the localization of source regions and plume boundaries. Near-term engineering efforts should prioritize path executability in strong currents and real-time onboard computation. Subsequent work should better account for multiple sources of error, reduce the communication burden of coordination, and develop task-oriented methods for assessing sampling value across multiple variables. Tiered validation will be essential for reliable long-duration operation.