Keywords: Autonomous underwater vehicles, Real-time decision systems for ocean robotics and autonomous operations, Trust, ethics, and transparency in marine AI systems
Abstract: Autonomous Underwater Vehicles (AUVs) have increasingly been utilized for underwater applications that are inaccessible for humans. However, due to the complexity of modeling hydrodynamics in uncertain underwater environments, data-driven methods are often employed for modeling and control of AUVs. One popular data-driven approach for generating optimal controllers is Reinforcement Learning (RL). However, designing a suitable reward function is often challenging, and RL lacks analytical methods for evaluating the learned policy. In many scenarios, it may be preferred to apply RL with human supervision, where it is used as a decision support method that suggests action to be taken by a human operator. In order to evaluate the suitability of a suggested action from an RL agent, the operator must be able to predict the effect of the suggested action. However, existing RL-based decision support methods do not offer justification for suggested actions that assist the operator to predict their effects. Therefore, an RL-based decision support method is proposed in this paper to provide justification for suggested actions to assist operators to validate the suitability of the suggestions. Results of this method for AUVs performing sea floor mapping under the influence of near-surface disturbances are presented.

Keywords: Real-time decision systems for ocean robotics and autonomous operations, Integration with autonomous platforms and observatories, Trust, ethics, and transparency in marine AI systems
Abstract: Despite the demonstrated utility of symbolic task-and-motion planning systems for safe and adaptable behavior, non-technical users of deployed marine autonomy generally do not use these methods because developers often assume they are complex to build and opaque to interpret. We present a mission-planning interface that addresses these concerns by abstracting a mission-planning problem for the operator into a set of composable goals with optional ordering constraints, which delegate to an onboard teleoreactive task and motion planner. Resulting behaviors can be viewed in 2D or 3D alongside plan decomposition, resource commitments, and anticipated state progression. We argue that this design places goal-level retaskable autonomy in the hands of non-technical users without demanding symbolic planning expertise. The design draws on the levels-of-robot-autonomy (LORA) and situation awareness-based agent transparency (SAT) frameworks, is deployed as a web application, and is released in part as open-source packages. We report preliminary valuations as a cognitive walkthrough against the workflow currently used for AUV Sentry, in preparation for a more comprehensive on-vehicle user study. We also discuss the implications of this work for calibrated performance-based trust, identifying it, alongside relation-based trust and psychological safety, as a motivating concern.

Keywords: Real-time decision systems for ocean robotics and autonomous operations, Vehicle navigation, AI/ML for oceanographic and environmental data analysis
Abstract: Long-duration autonomous underwater vehicle (AUV) missions generate large volumes of optical images and navigation data, whereas only a limited amount of information can be transmitted through underwater acoustic communication during a mission. Consequently, operators cannot obtain sufficient situational awareness during missions. This paper presents SummarAI, an onboard AI system that periodically generates compact mission summaries and transmits them through low-data-rate underwater acoustic communication. The system autonomously analyzes mission data and extracts representative information for acoustic transmission. Experimental validation through sea trials using a camera-equipped AUV demonstrated that the proposed system successfully generated and transmitted mission summaries while preserving the essential information required for situational awareness. The proposed system provides a practical framework for near-real-time situational awareness during long-duration AUV missions under severe underwater communication constraints.

Keywords: Real-time decision systems for ocean robotics and autonomous operations, Data fusion, uncertainty quantification, and predictive modeling, Foundation models and large-scale training for marine applications
Abstract: Ensuring that our nation’s military platforms can remain fully functional while deployed in an operationally available posture is critical to the safety and security of our soldiers. The operational availability of these platforms relies on the timeliness of required depot-level maintenance and repairs that occur prior to deployments. By democratizing depot-level knowledge in a way that enables a junior technician (with appropriate resources) to troubleshoot repairs while in the field, it is possible to revolutionize how forward deployed repairs are performed and increase platform availability. In 2024, a Johns Hopkins Applied Physics Lab (APL) research team merged enabling technologies in data analytics, generative artificial intelligence (GenAI), and mixed reality to develop a prototype AI-enabled headset that illustrates this capability. To integrate these technologies together, the team developed a generative AI-based manager (using OpenAI’s GPT-4o) that (1) leverages existing data analytics methods to statistically determine faults in a system (2) guides users interactively through a generated repair process and (3) communicates through speech and visual information via a portable augmented reality (AR) headset. The agentic AI manager uses large-language model guided finite state machines to access various AI agents for state-specific capabilities and to facilitate interactions during repair processes. The team tested the AI prototype system, with various users, on a relevant complex electronics system to demonstrate various physical fault scenarios, including multiple component repairs. While this AI prototype was designed specifically for electronics maintenance and repairs, the capability can be extended to support human operators in other, domain agnostic, complex, and multi-step processes.

Keywords: Vehicle navigation
Abstract: Autonomous Surface Vehicles (ASVs) operate without an onboard crew and must comply with the International Regulations for Preventing Collisions at Sea (COLREGs) when encountering other vessels. Classical planners, optimization-based methods, and deep reinforcement learning can perform effectively within predefined operating conditions but may struggle to generalize to unfamiliar encounter geometries. Large Language Models (LLMs) offer a potential alternative by interpreting heterogeneous situational information and reasoning over navigation rules. This paper proposes an LLM-guided navigation framework that varies situational input along two dimensions: temporal and visual context. We evaluate three configurations: zero-shot text, text augmented with a rolling decision history, and decision history paired with a rendered tactical image, using GPT-5.2, Claude Sonnet~4.5, and DeepSeek-chat across six COLREGs scenarios. We also introduce a rule-citation alignment metric that determines whether a compliant action is justified using the correct rule, rather than assessing action correctness alone. Across 1,060 queries generated from 480 simulation runs, multimodal input achieves the highest per-query accuracy of 71.3%, compared with 61.7% for memory-augmented input and 55.7% for zero-shot prompting. Model choice also substantially affects performance: under a fixed prompt, GPT-5.2 achieves 68.5% accuracy, whereas DeepSeek-chat achieves 46.6%. The alignment analysis reveals a gap concealed by action accuracy: GPT-5.2 cites the correct rule in 91-96% of its compliant decisions, while Claude does so in only 44-49%. Thus, current LLMs produce correct actions and rule-aligned justifications at markedly different rates, underscoring the need for domain-specific training and rigorous safety validation before real-world maritime deployment.

Keywords: Autonomous underwater vehicles, Low-cost, scalable oceanographic sensors and platforms, Marine optics technology and instrumentation
Abstract: Visual place recognition (VPR) is an important component of visual simultaneous localization and mapping (SLAM), enabling systems to recognize previously visited locations and correct resulting trajectories through loop closure. In underwater environments, low-texture seafloor imagery, turbidity, and limited lighting make reliable place recognition particularly challenging. As a result, missed loop closures can lead to accumulated drift and fragmented maps. This work replaces the Bag-of-Words (BoW) VPR method used by ORB-SLAM3 with Pair-VPR to improve loop-closure candidate retrieval for underwater visual SLAM. Using downward-facing stereo imagery collected in Puget Sound, we compare default and trained DBoW2 configurations with the transformer-based Pair-VPR method. Candidate keyframe pairs are evaluated against ground-truth camera trajectories. The results show that Pair-VPR retrieves a greater proportion of verified matches while producing fewer overall candidates, indicating improved candidate quality in visually ambiguous underwater scenes. When integrated into ORB-SLAM3, Pair-VPR also enables split-map merges to be applied earlier and results in more consistent recovery across the entirety of our underwater datasets.

Keywords: Imaging and vision, Real-time decision systems for ocean robotics and autonomous operations, Data fusion, uncertainty quantification, and predictive modeling
Abstract: Simultaneous Localization and Mapping (SLAM) has been a pertinent research field over the past several decades. More recently, with the advent of self-driving cars and related technologies, visual SLAM has emerged as a rapidly growing area of research. However, the problem of visual SLAM for autonomous surface vessels and other marine platforms remains largely unexplored while presenting unique challenges. Maritime environments are complex due to their prevalence of dynamic water surfaces and sky regions that degrade feature extraction and pose estimation. This work presents a learned semantic segmentation approach for masking sky and water regions from monocular camera frames prior to SLAM processing. We train a SegFormer (MIT-B0) model on 300 hand labeled coastal images and evaluate impact on three state-of-the-art visual SLAM systems spanning distinct algorithmic paradigms: ORB-SLAM3 (classical feature-based), DROID-SLAM (deep optical flow) and WildGS-SLAM (Gaussian Splatting). Our masking method transforms ORB-SLAM3 from complete tracking failure (zero frames) to successful operation (635 frames). For DROID-SLAM, masking reduces trajectory drift by 32%, and increases processing speed by 1.8x. WildGS-SLAM achieves a 41% drift reduction, 36% smoother trajectories, 2.8x faster processing and4.8x more Gaussian map primitives. These results demonstrate that learned semantic masking is an effective preprocessing step that significantly improves visual SLAM robustness across multiple algorithmic paradigms in maritime settings. Index Terms—visual SLAM, maritime robotics, semantic seg- mentation, autonomous vehicles

Keywords: Imaging and vision, Vehicle navigation, AI/ML for oceanographic and environmental data analysis
Abstract: Visual SLAM, visual place recognition and Structure-from-Motion (SfM) underpin a wide range of ocean and Earth-observation missions, but underwater their performance is bounded almost entirely by the data association front end. We present a controlled comparison of nine feature matching pipelines - SIFT, ORB, AKAZE, Super- Point+SuperGlue, ALIKED, DISK+LightGlue, LoFTR, DKM and RoMa (the tiny_roma_v1_outdoor variant) - within a unified COLMAP-based SfM framework on three sequences: a UAV pass over a marine-terminating glacier in Svalbard, a deep-sea survey of the Bio9 hydrothermal vent on the East Pacific Rise, and a UAV orbit around a complex heritage building as an in-distribution control. The glacier and vent sequences lie far outside the training distribution of every learned method. All pipelines run under identical resolution, keypoint and match budgets, and geometric verification settings, so differences reflect the matching front end rather than tuning. We find that registration completeness alone overstates reconstruction quality. On the hydrothermal survey RoMa registers all 108 images, yet a cross-method consistency check finds 28% of its adjacent camera pairs disagree with the consensus of the other pipelines by more than 5×, and it returns the highest reprojection error of the nine methods after 9.1 hours of compute. ALIKED registers 99 of 108 with no such disagreement, in 27 minutes. Sustained wide-baseline verification, which RoMa alone maintains beyond a 30◦ baseline, predicts how many images a method attaches, not whether its geometry is sound. Reprojection error ranks the field differently again: DKM attains near-SIFT per-observation accuracy while recovering an order of magnitude more structure. No single pipeline is best on every criterion, and the right choice for a power- and compute-constrained AUV or ROV depends on whether coverage, trajectory reliability, or per-point accuracy is the priority.

Keywords: Autonomous underwater vehicles, Dynamic positioning, Human–ocean interactions and cultural perspectives
Abstract: Robots working alongside humans underwater incorporate a unique set of challenges to the broader fields of human-robot interaction and collaboration, not only related to circumstances of the environment and human safety, but also communication methods, such as capturing attention and conveying information. Prior research in virtual reality (VR) has shown that robots can perform motions as kinemes to express meaning and convey information to their human counterparts. As we move from VR towards real-world deployment, this work aims to address the sim-to-real gap when transferring the virtually tested kinemes onto a physical BlueROV2 in a pool scenario, tackling environmental challenges like fluid dynamics and drift, and mitigating effects without knowledge of absolute pose with respect to world coordinates. In several pool sessions, the scenario mimics the condition in which a diver is completing a manipulation task of closing a valve when the robot needs to interrupt that task and convey a switch in operations. This work expands our understanding of how to design autonomous robot interactions with divers in real underwater environments by enabling dynamic adjustments of robot behavior.

Keywords: Autonomous underwater vehicles, Data fusion, uncertainty quantification, and predictive modeling, Numerical modeling and simulation
Abstract: Hydrodynamic and inertial uncertainty affects underwater-vehicle simulation, prediction, control, and mission analysis, but nonlinear uncertainty propagation is computationally expensive. This paper presents a graphics processing unit (GPU)-accelerated stochastic simulation framework based on the standard marine-craft formulation. Posterior samples of inertia, added mass, damping, and restoring coefficients define deterministic model realizations. Symbolic code generation produces GPU functions for batched dynamics-operator evaluation and complete finite-horizon rollouts. The GPU simulates 65,536 complete vehicle dynamic simulations over a 25 s horizon window in 4.48 s. For these 65, 536 realizations, complete GPU inertia, inverse-inertia, Coriolis, restoring-wrench, and damping evaluations, including data transfer and result retrieval, take 1.38–6.20 ms, compared with 560–890 ms for sequential CPU evaluation, yielding speedups of 91–482 times.

Keywords: Holography and 3D imaging, Imaging and vision, Ocean exploration
Abstract: The underwater environment is challenging for 3-D reconstructions due to light scattering and diffusion from particles suspended in the water; turbidity; wavelength dependent light absorption; and lack of uniformity of illumination. Methods based on Gaussian splatting have generally been developed for conditions that allow good image quality and have primarily been tested using relatively shallow water or highly controlled data sets. There has thus far been little investigation into whether such methods can be used effectively to create high-quality 3-D reconstructions for real world ROV operations in deeper waters.

This paper presents an examination of how well Gaussian splatting-based techniques perform over a variety of publicly available underwater data sets, which represent different degrees of turbidity, loss of illumination through light degradation, and color attenuation. We seek through this study to determine the relative advantages, disadvantages, and limitations of currently available methods so as to provide insight regarding what methods may best be applied to real-world projects.

Keywords: Real-time decision systems for ocean robotics and autonomous operations, Low-cost, scalable oceanographic sensors and platforms, Imaging and vision
Abstract: Marine plastic debris fragments into difficult-to- recover microplastics, so effective mitigation requires localizing macro-debris before it degrades. A spatial-scale gap separates wide-area satellite remote sensing from fixed collection infras- tructure, leaving intermediate-scale localization underserved. We present Ocean Voyager, a low-cost catamaran autonomous sur- face vehicle (ASV) whose central contribution is a staged thermal- to-vision perception architecture for resource-constrained em- bedded hardware. A servo-mounted thermal camera performs a 360◦ sweep that proposes candidate bearings at low cost; only these candidates trigger three-camera RGB inference with multi-view voting for confirmation and false-positive suppression. Because expensive vision inference runs only on proposed can- didates rather than continuously, the architecture is estimated to reduce the vision-inference duty cycle by 20–60× relative to continuous scanning, which is what makes long-duration monitoring more feasible on Raspberry Pi-class hardware. We describe the physical realization of the platform, the rationale for staged perception, and the multi-view voting mechanism, and we report a physical hardware-in-the-loop demonstration in which the complete thermal-to-vision pipeline ran end-to-end on the assembled vehicle. In that run the thermal stage proposed every placed object while the vision stage performed the waste/non- waste decision—correctly rejecting a distractor that in fact pro- duced the strongest thermal signature—illustrating the division of labor the architecture is built on. The detector evaluation and the autonomy controller are reported in companion work; this paper concerns the perception architecture and its system-level integration.

Keywords: Image and video processing, Acoustic telemetry and communication, Remotely operated vehicles, Autonomous underwater vehicles
Abstract: Real-time visual feedback is essential for tetherless control of remotely operated vehicles, particularly during inspection and manipulation tasks. Though acoustic communication is the preferred choice for medium-range communication underwater, its limited bandwidth renders it impractical to transmit images or videos in real-time. To address this, we propose a model-based image compression technique that leverages prior mission information. Our approach employs trained machine-learning based novel view synthesis models to store and render views of scenes that serve as priors, and minimizes the differences between camera images and model-rendered images to improve compressibility. We evaluate our method using datasets collected in an artificial ocean basin and demonstrate significant improvements in compression ratio and image quality compared to existing methods. Moreover, our approach exhibits robustness to the introduction of novel objects within the basin scene and maintains performance under real-world degradations such as backscatter, as demonstrated on an online coral reef dataset. These results highlight the promise of our proposed technique for enabling real-time, high-quality visual feedback over bandwidth-constrained underwater acoustic links.

Keywords: AI/ML for oceanographic and environmental data analysis, Foundation models and large-scale training for marine applications, Sonar imaging
Abstract: Denoising diffusion probabilistic models (DDPMs) show impressive results in producing high-quality synthetic optical images. However, research studies are sparse for generating synthetic aperture sonar (SAS) images, and DDPMs have not been largely studied with large databases of SAS imagery. Generative adversarial networks (GANs) have been used to generate realistic seafloor SAS images, but the generation of realistic object images is limited. Also, in object recognition for SAS imagery, clutter data far outnumbers data on objects-of-interest, and this class imbalance can bias recognition models to have better performance on clutter than objects-of-interest. Toward the end of addressing this imbalance, we compare DDPM methods with a minimal, state-of-the-art GAN (R3GAN) and evaluate both methods' effectiveness in generating accurate and diverse SAS imagery. We demonstrate that DDPMs outperform the GAN on a SAS dataset across Frechet Inception Distance, probabilistic precision, and probabilistic recall metrics. We also find that DDPMs are able to match the sharpness and resolution of the real SAS dataset, suggesting that their generated data can be used to mitigate the clutter-object imbalance.

Keywords: Foundation models and large-scale training for marine applications, Security and resilience of AUV/USV/ROV platforms, Classification and pattern recognition (parametric and non-parametric)
Abstract: Automatic target recognition (ATR) is a critical technology that enables advanced naval capabilities for a variety of underwater and maritime applications using synthetic aperture sonar (SAS). However, most deep learning-based ATR systems remain constrained by the need for human-in-the-loop assessment and extreme class imbalances. Additional operational challenges further exacerbate this bottleneck, such as the extreme scarcity of imaged objects of interest compared to the vast volume of background and clutter imagery. In this work, we present a three-stage framework for contrastive learning with Vision Transformers, adapting the standard ViT foundation model (ViT-7B) pretrained in the DINOv3 framework for underwater SAS ATR. To account for the unique physics of acoustic propagation captured in sonar imagery, we employ a Low-Rank Adaptation (LoRA) workflow in our three-stage framework to enhance ATR performance on at-sea SAS data. While LoRA establishes a robust feature baseline where standard supervision struggles, we also implement a "Refiner" in the second stage, utilizing active learning and Hard Negative Mining to "harden" the decision boundary. Lastly, Supervised Contrastive Learning (SupCon) is performed in the final stage to differentiate "geological decoys" that resemble man-made objects, leveraging principles from instance-level contrastive learning. This approach actively reshapes the feature space, pulling targets into a compact manifold while maximizing their distance from clutter. Our three-stage LoRA SupCon pipeline is evaluated on SAS data collected from an at-sea sensor, showcasing transformative improvements at each stage in the pipeline. The resulting performance is characterized by a peak Matthews Correlation Coefficient (MCC) of 0.8158, a recall of 0.8487, and a precision of 0.7866, establishing a clear pathway toward operational precision in high-clutter acoustic environments.

Keywords: Classification and pattern recognition (parametric and non-parametric), AI/ML for oceanographic and environmental data analysis, Sonar signal processing
Abstract: In passive sonar classification, machine learning algorithms have been used based on the radiated noise targets emit. Although machine learning algorithms for passive sonar classification have advanced significantly, underwater acoustic data collected in real ocean environments often exhibit domain shift due to variations in environmental conditions, which can significantly degrade classification performance. However, few studies have addressed this issue in depth. In this study, we reorganize DeepShip dataset, a publicly available dataset originally composed of four ship-type classes, at the individual-vessel level according to different acquisition conditions to construct a more realistic scenario in which generalization performance under domain shift can be evaluated. To enhance generalization performance in this reorganized setting, we applied a fine-tuning method and compared it with the result of supervised learning.

Keywords: Classification and pattern recognition (parametric and non-parametric), Machine-learning classification of acoustic signals, Surveillance, detection, and threat-assessment techniques
Abstract: Reliable discrimination between rigid targets, such as natural rocks, and elastic targets, including unexploded ordnance (UXO) and sea mines, is a fundamental challenge in active sonar-based survey operations. Elastic objects, such as the metallic shells typical of UXOs and mines, sustain resonant acoustic modes when insonified at frequencies (10–30 kHz in this study) producing wavelengths comparable to the target dimensions, creating backscatter signatures that differ from those of rigid scatterers. This work investigates whether those differences can be captured by time-series features and exploited by machine-learning classifiers, using a publicly available dataset of controlled in-air sonar measurements as a laboratory analog of the underwater problem. From each matched-filtered echo window, features were extracted from the real, imaginary and envelope components using the tsfresh library and reduced via false discovery rate (FDR)-controlled selection; Random Forest and XGBoost classifiers were then evaluated with per-class, imbalance-robust metrics and scan-level bootstrap confidence intervals, across four environments of increasing complexity (free-field, flat interface, and rough interface with proud, i.e. resting above the interface, or partially buried targets). In free-field and flat-interface conditions both classifiers discriminate the classes almost perfectly (accuracy 0.94–0.99, with tight confidence intervals). Under rough interfaces, performance drops to 0.62–0.69 accuracy and the classifiers bias toward the rigid class, so that most elastic (UXO- and mine-like) targets are missed; ROC-AUC and MCC nonetheless indicate that a weak but statistically significant signal survives. As the measurements are in-air (with a much larger target–medium impedance contrast than underwater), these results constitute a controlled proof-of-concept characterizing how surface clutter erodes resonance-based discrimination, rather than a claim of absolute underwater performance.

Keywords: Machine-learning classification of acoustic signals
Abstract: Underwater acoustic target recognition (UATR) with passive sonar is strongly affected by propagation-induced channel variation. Even for the same vessel, the received time-frequency spectrum can change markedly with source-receiver geometry, multipath, and environmental conditions. As a result, a classifier trained on conventional short-time Fourier transform (STFT) inputs may learn recording-specific channel patterns rather than source-intrinsic spectral structure. This issue becomes especially evident under realistic evaluation protocols that separate recordings or measurement locations between training and test data. Therefore, the central problem addressed in this work is how to construct an input representation that is less sensitive to channel dominant distortion while preserving information relevant to target classification.

We propose a dechannelization motivated by the passive sonar observation model R(ω,t)=S(ω,t)H(ω,t)+N(ω,t), where S(ω,t) denotes the source-related spectrum, H(ω,t) the time-varying propagation channel, and N(ω,t) additive noise. Instead of explicitly estimating and inverting H, which is generally impractical in passive sonar operation, the method suppresses the dominant slow channel trend directly in the received spectrogram R(ω,t).

The proposed representation was validated through both physics-based simulation and data-driven classification. In a Bellhop shallow-water simulation with a moving source, the dechannelization map became more similar to the original source spectrum than the raw received spectrum. These results support the physical interpretation that the method suppresses dominant channel distortion and makes source-related structure more explicit. For real-data evaluation, we used a subset of the DeepShip dataset containing four ships observed across multiple measurement locations. Using the same lightweight CNN for all conditions, we compared STFT and dechannelization inputs under four progressively harder split protocols. The proposed representation consistently outperformed the STFT baseline. The results indicate that dechannelization is most beneficial when the evaluation

Keywords: Benchmark datasets, reproducibility, and open-source frameworks, Machine-learning classification of acoustic signals, Passive and active acoustic monitoring
Abstract: Underwater Acoustic Target Recognition (UATR) of ships is well-suited for machine learning, yet its progress is hindered by the lack of large, diverse, and publicly available labeled datasets. In this work, we introduce UniqueShip, a machine learning-ready benchmark dataset for UATR applications sourced from the open Ocean Networks Canada (ONC) repository. Unlike previous datasets, we explicitly control for "data leakage" between the training and evaluation sets to ensure more reliable and generalizable model evaluation that does not encourage the model to memorize individual ships. We demonstrate that typical, random data partitioning in two prominent UATR datasets leads to falsely optimistic test performance, increasing accuracy by 10–48 percentage points compared to our more careful partitioning. Ablations on UniqueShip further show that doubling the number of unique vessels improves accuracy by 2.4–2.6 percentage points, while doubling total audio duration improves only by 0.8–1.3 points, indicating that vessel diversity should drive dataset curation more than total hours. We provide baselines with convolutional and transformer backbones, and analyze how ship metadata correlates with classification performance, finding that individual vessel characteristics predict classification difficulty far better than distance to the hydrophone alone. Overall, UniqueShip contains 2,460 hours of ship-radiated audio from 4,218 unique vessels (3,437 hours including background). We publish the dataset, code, and easy-to-download splits at uniqueshipdata.org to foster further UATR research.

Keywords: Geoacoustic inversion, Sound propagation and scattering, Machine-learning classification of acoustic signals
Abstract: Using different parameterizations impacts the results of ocean sound propagation models. The sound lost when a sound wave and seabed interact is known as bottom loss (BL). It is possible for different seabed types to produce similar BL results, which introduces an ambiguity for backward models that may try to predict a seabed from BL. The field of information geometry has tools that are used to examine this ambiguity, which are applied to find the relative importance of parameters to a model's output. Examples include analysis of the model manifold and Fisher Information Matrix applied to an analytical BL model using automatic differentiation for a multi-layered ocean environment. These examples indicate that, for grazing angles 0-30 degrees and frequencies 500-4000 Hz, as sound speed increases above approximately 1750 m/s, variation in density have little impact on the modeled BL.

Keywords: Ocean noise, Bioacoustics, Passive and active acoustic monitoring
Abstract: Regional and seasonal variability in underwater soundscapes around Japanese coast were examined. Autonomous underwater recorders were deployed at ten coastal sites during summer and winter to characterize ambient sound levels. The results revealed marked seasonal and diel differences: in summer, evening fish choruses were detected in most of Japanese coastal waters except for off Hokkaido and Fukui, whereas in winter, high-intensity, low-frequency broadband noise—primarily driven by strong northwest monsoons—dominated, particularly along the Japan Sea coast. These patterns underscore the dynamic interplay of biophony and geophony in Japanese coastal waters. Even the same source level of anthropogenic noise, the maximum effective range of the noise varies seasonally, becoming longer in summer and shorter in winter due to differences in masking level by the ambient noise. Our findings emphasize the importance of jointly considering source levels and ambient noise levels when evaluating the impacts of underwater noise on aquatic organisms, as ambient sound strongly influences both the audible range of anthropogenic sources and underwater communication of aquatic animals.

Keywords: Systems and observatories, Energy-harvesting and long-duration monitoring solutions, Integration with autonomous platforms and observatories
Abstract: Stationary autonomous seabed nodes are a core element of modern ocean observing systems, supporting long-term environmental monitoring, subsea inspection, and offshore industrial applications such as renewable energy and carbon capture and storage. While sensing technologies are generally mature, platform lifetime and scalability remain constrained by energy availability. In practice, energy requirements are often estimated using simplified assumptions that neglect realistic duty cycles, observing system architectures, communication strategies, and deployment constraints, resulting in premature intervention or conservative over-design. This work presents a system-level analysis of energy requirements for stationary seabed observing platforms under representative field conditions, explicitly linking platform architecture, payload composition, and data communication approach. On the supply side, commonly deployed solutions are examined, including primary and rechargeable battery systems, buoy-mounted solar panels, wind turbines, wave-energy converters, and inductive charging. Typical offshore renewable sources provide on the order of 50-300 W depending on location and environmental conditions. These solutions are assessed in terms of usable energy density, reliability, operational complexity, and compatibility with different observing system layouts. On the demand side, the analysis draws on project and field deployment experience with representative payloads including environmental sensors, ADCPs, imaging sonars, passive acoustic monitoring systems, and optical systems. A medium-scale observing node equipped primarily with passive sensors and an ADCP requires approximately 10-15 kWh for one year of autonomous operation, while additional active acoustic payloads such as echosounders contribute roughly 2 kWh per year. Data handling and transmission are shown to dominate energy demand: transferring 1 MB of data requires approximately 1 mWh using optical links, around 10 Wh using acoustic communication, and up to 100 Wh using low-frequency electromagnetic methods. The results are synthesized into a practical framework for guiding the design of long‑term seabed observing platforms.

Keywords: Integration with autonomous platforms and observatories, Ocean energy, Energy-harvesting and long-duration monitoring solutions
Abstract: The development of a highly capable and versatile autonomous observatory node (AON) is necessary for charging of underwater vehicles and acoustic communication to provide communication and power services to deep ocean. Technical challenges arise when a subsea microgrid is isolated and operates from battery/ fuel cell instead of long cables from surface/ shore. This paper presents the design, simulation and hardware implementation of a DC-DC power converter designed using GaN FETs to provide the required power derived from a 200 V DC input to loads rated at 5 V, 12 V. It enhances the robustness of the system by improving efficiency (>95%), power density (140W⁄〖in〗^3 ), and mitigation of noise and transients (< 2% ripple) in the ocean microgrid

Keywords: Ocean energy, Energy-harvesting and long-duration monitoring solutions, Integration with autonomous platforms and observatories
Abstract: In a conventional Cabled Ocean Observatory (COO) system, electrical power is transmitted from shore through subsea cables and distributed to individual loads using subsea power electronic converters. The converters are typically housed in thick-walled pressure vessels maintained at an internal pressure near 1 bar, which limits heat dissipation and can reduce system reliability. As part of a broader research project, the authors are developing a systematic procedure for designing Pressure-Tolerant Power Electronics (PTPE)-based converters for subsea applications. These converters operate at ambient hydrostatic pressure within thin-walled enclosures filled with dielectric fluid, thereby improving cooling and eliminating the need for thick pressure vessels. This approach enables a shift towards an Autonomous Ocean Observatory Node (AOON) infrastructure. Within this broader effort, this paper focuses on the design and simulation-based validation of a pressure-tolerant Underwater Microgrid (UMG) intended to support an AOON. The microgrid incorporates a Multi-Port Converter (MPC) that autonomously regulates power distribution under varying subsea operating conditions and maintains continuous power delivery to the observatory loads. A system-level testbed is implemented on a Typhoon Hardware-in-the-Loop (HIL) platform to validate the autonomous operation of the proposed microgrid under varying subsea operating conditions.

Keywords: Systems and observatories, Autonomous underwater vehicles, Oceanographic instrumentation and sensors
Abstract: With recent advances in autonomous underwater vehicle (AUV) docking and charging, the development of resident autonomous oceanic systems is a clear next step in advancing AUV technology and the greater blue economy. Transitioning from expeditionary to resident autonomous operations requires multidisciplinary hardware optimization tailored to the specific goals and resources available to each system. We present an open--source hardware optimization tool for microgrid systems composed of a power source, energy storage unit, and power sink(s). The tool simulates time--domain power exchanges between system components, given a time--varying resource input over a configurable duration, to help users determine optimal hardware configurations for their specific applications. To demonstrate its utility, we use the tool to analyze a system composed of a Wave Energy Converter (WEC), central battery, and AUVs using estimated power generation data from a one--meter WEC operating ten miles west of the Oregon coast, coupled with multiple AUV configurations using specifications based on commercially available AUVs. We consider three system classifications based on the ratio of power consumed to power generated, and evaluate the operational characteristics of systems falling into each classification. Finally, we investigate the impacts of central battery capacity on the efficiency and performance of WEC--AUV systems.

Our findings reveal how `power ratio' and central battery capacity shape system performance. Systems with low power ratios maintain constant AUV operation, but underuse available power. Systems with high power ratios fully utilize available power but sacrifice AUV mission frequency to do so. Systems with balanced power ratios occupy the `sweet spot' between over- and under--generation, though they also see the greatest impact from changes in the system's total energy storage. These results demonstrate how the tool can aid in hardware compatibility decisions and performance comparisons across systems with a range of power characteristics. While illustrated here with a WEC--AUV system, the underlying approach generalizes to any microgrid with variable power generation and one or more energy sinks.

Keywords: Offshore structures, Deepwater development technology
Abstract: The deployment of power electronic systems in subsea and offshore environments introduces significant reliability challenges due to sustained exposure to high hydrostatic pressures. Electronic components are susceptible to mechanically induced degradation, which can alter their electrical behavior and reduce operational lifetime. Despite this, controlled experimental platforms capable of systematically evaluating the mechanical robustness of such devices under realistic oceanic pressure conditions are lacking. This work presents the design and implementation of a laboratory-scale hydrostatic loading system capable of replicating deep-sea pressure environments to test the performance of film capacitors. The system employs a pressure intensifier coupled with a pressure cylinder to generate controlled hydrostatic stresses. Preliminary results indicate variations in device performance, demonstrating a coupling between mechanical stress and electrical characteristics. SEM and EDS mapping further help evaluate microscale effects on the film capacitors due to the induced hydrostatic stress. Index Terms —film capacitors, hydrostatic stress, pressure-tolerant electronics

Keywords: Technology for precision fish farming
Abstract: Aquaculture is a rapidly expanding sector in global food production, with increasing emphasis on improving efficiency and sustainability. Monitoring fish growth parameters, such as body length and weight, is crucial for optimizing feeding strategies and harvest timing. However, traditional measurement techniques are labor-intensive, highly invasive, and limited in the variety of data collected. This paper introduces an AI-driven platform that leverages computer vision and machine learning for real-time, less invasive measurement of fish morphological parameters and weight estimation. By leveraging deep learning, the proposed method and platform enhance data accuracy and efficiency while addressing challenges associated with manual measurement. The approach is validated on Rainbow Trout (Oncorhynchus mykiss) in a controlled experimental setup designed to minimize environmental variability. The measurement error for fish morphometrics is approximately 1.23 millimeters, while the error for weight estimation is around 0.179 grams, demonstrating the effectiveness and accuracy of the AI-based system. Importantly, the proposed tool enables real-time measurement, significantly reducing processing time. It is also designed for seamless integration with other experimental setups, requiring only a simple calibration procedure.

Keywords: Instrumentation, sensors and autonomous systems for sea-based aquaculture, Rapid-deployment systems and modular instrumentation, Open-source hardware, software, and community innovation
Abstract: Persistent vertical lines used in conventional trap and pot fisheries create whale-entanglement and lost-gear concerns, prompting the development of on-demand fishing gear that eliminates continuous buoy lines while maintaining trap retrieval capability. This paper presents an intelligent on-demand fishing gear architecture with a conservation observation module. The proposed design separates the physical retrieval pathway from the conservation-record pathway and binds them only through a shared gear identity. The retrieval system employs a piston-based buoyancy-control mechanism that surfaces a recovery line while keeping the primary retrieval line laid along the seabed during deployment, thereby eliminating stored-line handling in the primary recovery pathway. A mechanically distinct backup-float pathway is retained as a contingency recovery method. The conservation observation module consists of a handset-based human-machine interface (HMI) and the EarthRanger platform. The handset records gear identity through Bluetooth Low Energy (BLE), retains deployment records during connectivity outages, and synchronizes them when network access becomes available. Prototype evaluation included component pressure tests, buoyancy and release validation, a 100-cycle endurance experiment, simulation-based validation of the conservation-record data chain, and five at-sea deployment trials. Battery voltage decreased from approximately 16.80 V to 16.15 V during the endurance test. In field trials conducted off Miaoli, Taiwan, ten of ten deployed intelligent on-demand fishing gear prototype units were successfully recovered following soak periods of approximately fifteen days. Data-chain validation against the live EarthRanger sandbox demonstrated successful offline record retention, synchronization after connectivity restoration, and identifier consistency between the handset and platform. The results support the feasibility of the proposed recovery architecture and conservation-record framework; however, the complete integrated workflow combining the recovery hardware and conservation observation module has not yet been validated under operational fishing conditions.

Keywords: Instrumentation, sensors and autonomous systems for sea-based aquaculture, Autonomous underwater vehicles
Abstract: Macroalgae farming is expanding as a component of sustainable blue-food production, yet farm infrastructure and crops remain exposed to environmental hazards that are difficult to observe, particularly in winter. This paper presents a case study of a commercial macroalgae farm on the Swedish Skagerrak coast that lost essentially its entire annual harvest and roughly half of its cultivation infrastructure during the ice winter of 2025–2026, without the damage being detected in time to act. Based on operator data, we quantify the economic impact: a booked infrastructure write-off of approximately EUR 45000, an inventory write-down of EUR 145000 against a foregone sales value of approximately EUR 180 000, clean-up and damage- assessment costs of some EUR 14000, and unquantified losses of research deployments—none of which were recoverable, as the installation was uninsured. From the event chronology and its economics we derive operational requirements for a resident Autonomous Underwater Vehicle (AUV) system with subsea docking that provides persistent, low-cost inspection through winter, including under-ice operation, acoustic perception in low visibility, and event-triggered damage surveys. Inverting the usual comparison, we estimate the maximum cost at which such a system is justified for a small farm—of order EUR 20 000–70 000 with a safety margin, and higher when shared across farms or credited with the routine crop health and growth monitoring that the same platform enables. Persistent condition data could furthermore support insurability of an otherwise poorly insurable risk. The case demonstrates that farm resilience depends not only on structural design, but on the ability to observe, interpret, and act before damage propagates.

Keywords: Acoustic telemetry and communication, Low-cost, scalable oceanographic sensors and platforms
Abstract: Ropeless fishing systems, developed to reduce the risk of whale entanglement in the vertical lines associated with crab and lobster commercial fishing activities, have seen significant advancement in recent years due to federal research and development funding. Grant funding from the National Fish and Wildlife Foundation has supported both the development of an interoperable topside deck box by Teledyne Benthos and enhancements to the Sub Sea Sonics AR4RT acoustic release including the implementation of a ping and ranging response features The Teledyne Benthos Blue Ocean Box (BOB) is a topside acoustic deck box capable of transmitting and receiving acoustic signals from multiple acoustic release manufacturers that operate in the range of 15 to 40 kHz including Teledyne, EdgeTech, and Sub Sea Sonics. The BOB transducer utiliz a wideband ceramic ring and 4 channel ultra short baseline (USBL) receiver allowing for acoustic localization of the subsea releases by measuring both range and bearing. The ability to integrate acoustic protocols from multiple manufacturers benefits regulators, enforcement and fishers interested in the bottom location of their gear, particularly when the acoustic release manufacturer does not offer a topside USBL for localization.

The Sub Sea Sonics AR4RT+ acoustic releases send and receive at 38.4 kHz and 35.7 kHz respectively. Lab testing showed the BOB was capable of successfully transmitting Sub Sea Sonics acoustic protocol at 35.7kHz. The 38.4kHZ return pings from the AR4RT+, however, proved difficult to detect with existing hardware being out of range of the design frequency band. Sub Sea Sonics and Benthos collaborated to develop a novel wake pulse within the SSS existing acoustic protocol that commands the AR4RT+ to switch the transmit frequency to a lower 16 kHz when interrogated by the BOB. This allows the AR4RT+ releases to be backwards compatible with currently deployed SSS topside deckboxes as well as interrogated or acoustically localized using the Benthos BOB system.

Field testing was conducted in the Summer of 2026 in Cape Cod Bay aboard the R/V Ocean Researcher.

Keywords: AI/ML for oceanographic and environmental data analysis, Benchmark datasets, reproducibility, and open-source frameworks
Abstract: Underwater monitoring is essential for biodiversity assessment and conservation, particularly in the Southern Arabian Gulf, where marine ecosystems face extreme environmental conditions and increasing anthropogenic pressure. Baited Remote Underwater Video Systems (BRUVS) provide a non-extractive and repeatable approach for surveying marine biodiversity; however, labor-intensive footage analysis limits their large-scale adoption. To address this challenge, we introduce BRUVFish5, an expert annotated dataset of BRUVS imagery from United Arab Emirates coastal waters covering five fish species selected for their ecological and fisheries relevance, and develop a You Only Look Once (YOLO)-based pipeline for automated multi species detection. Despite the relatively small dataset, our model achieves accurate and stable performance under variable underwater visibility through high-quality annotations and targeted augmentation. This framework demonstrates the potential of regionally trained computer-vision approaches for scalable biodiversity monitoring through accelerated BRUVS analysis.

Keywords: Energy-harvesting and long-duration monitoring solutions, Ocean noise, Ocean energy
Abstract: Long term Internet of Underwater Things (IoUT) deployments are constrained by battery capacity, motivating energy harvesting from ambient underwater acoustic noise. This paper presents a zero-crossing synchronous switch harvesting on inductor (SSHI) for low-frequency underwater acoustic energy harvesting using a piezoelectric hydrophone. The circuit briefly connects an inductor across the transducer capacitance at each piezoelectric-current zero crossing, producing controlled voltage inversion and improving energy transfer to a full-bridge rectifier and resistive load. The switching interval is selected as one-half of the damped resonant period. Output voltage, load power, and conversion efficiency are evaluated using a 1 kΩ reference load showing that the SSHI configuration outperforms the full-bridge rectifier baseline over most operating conditions. The results demonstrate possible low-frequency energy extraction under periodic and nonstationary acoustic excitation for IoUT nodes.

Keywords: Autonomous underwater vehicles, Naval architecture, Ocean energy
Abstract: A highly reliable energy generating device that can harvest wave-energy and attach to an unmanned underwater vehicle would allow the indefinite operation of onboard low-power consuming sensors. We have designed and fabricated an integrated hybrid triboelectric (TEG) and electromagnetic (EMG) generator to meet this need. Our design consists of a tunnel track that has the interior bottom lined with copper tape electrodes and the exterior wrapped with motor winding wire. Within the tunnel, polytetrafluoroethylene (PTFE) balls, to power the TEG, and neodymium balls, to power the EMG, can roll freely in response to vehicle pitch and roll motions. This device has been tested on a tilt table that recreates the vehicle pitch and roll motions caused by interactions with incoming waves. The TEG and EMG are each connected to separate capacitors using simple electrical circuits that include rectifiers. The power generated by the TEG and EMG results in energy being stored on the corresponding capacitor which we determine by measuring the time history of the voltage on each capacitor. The individual performance of the TEG and EMG were determined by operating the system with only a single PTFE ball and then only a single neodymium ball inside. Then, to explore if any interference effects exist when both components are operating, a PTFE and a neodymium ball were placed inside together. Our test results indicate that there is no noticeable interference effect that decreases the performance of either component. We also demonstrate that the addition of a second PTFE ball improves the performance of the TEG by a factor of 5 and also, by a factor of 1.5, the EMG.

Keywords: Ocean energy, Hydrodynamics, Vehicle design
Abstract: Unmanned Underwater Vehicle (UUV) gliders are a cornerstone of long-duration oceanographic observation, relying on high efficiency and low energy consumption to achieve mission lengths of weeks or months. A primary limitation on this endurance is the onboard battery capacity. In-situ energy harvesting offers a compelling solution to extend mission life, with hydrokinetic turbines capable of converting the glider’s forward motion into electrical power to recharge batteries. However, the integration of any such device introduces a critical performance penalty: hydrodynamic drag. This added drag increases the energy required for propulsion, creating a complex optimization problem. The net energy gain must be positive and significant enough to justify the turbine’s presence. This abstract provides a comparative analysis of three prominent axial-flow turbine designs—open rotor, ducted, and open-center evaluating them for this specific application, with a focus on the crucial balance between power generation and induced drag in low-velocity flow regimes typical of UUV gliders (typically < 1.5 m/s).

Keywords: Data fusion, uncertainty quantification, and predictive modeling, Foundation models and large-scale training for marine applications, AI/ML for oceanographic and environmental data analysis
Abstract: Climate control aboard naval vessels is often managed through manual, open-loop damper and valve adjustments that cannot anticipate the rapid, correlated heat-load transients characteristic of shipboard operation leading to chronic comfort failures, wasted energy, and elevated thermal signature. This paper proposes a novel two-timescale supervisory control architecture in which a Deep Deterministic Policy Gradient (DDPG) reinforcement learning agent retunes the gains of conventional inner-loop PID controllers on a slow 300 s cadence, while the fast, bounded PID loops retain direct actuator command authority at 5 s intervals. The plant under control is a coupled three-zone thermal–humidity model of a destroyer chilled-water air-conditioning system, incorporating inter-compartment bulkhead conduction, effectiveness–NTU cooling coils, and seawater-temperature-dependent chiller efficiency, implemented in MATLAB/Simulink. The proposed Supervisory RL-PID controller is benchmarked against not only conventional manual operation, but also a strong Adaptive Gain-Scheduled PID (AGS-PID) baseline that combines Big Bang–Big Crunch offline gain optimisation with online Recursive Least Squares plant identification. In a controlled six-hour coupled simulation, the RL-PID achieves a 10.6% energy reduction while maintaining 98.6% comfort-band occupancy. More significantly, in a 14-day empirical replay test driven by real shipboard disturbance profiles spanning diurnal heat loads, weather conditions, and chiller temperature variation, the RL-PID demonstrates 1.4% lower energy consumption than AGS-PID while critically reducing actuator command variation by approximately 60%, without any online plant identification. These results indicate that judicious application of reinforcement learning in a supervisory capacity restricting the RL agent to bounded, infrequent PID gain adjustments while a conventional controller retains direct regulatory authority can substantially reduce mechanical wear while maintaining competitive efficiency and acceptable comfort under the evaluated conditions: a trade-off well suited to the low-maintenance, high-endurance demands of naval HVAC applications.

Keywords: Vehicle performance, Autonomous underwater vehicles, Ocean economic potential
Abstract: Autonomous deep-sea resource collection introduces uncertainty due to coupled robotic, offshore, and subsea operations. This paper presents a techno-economic simulation framework to optimize and evaluate system scalability. The approach integrates physics-informed models with cost structures to produce probabilistic forecasts of key performance indicators, including cost per tonne and return on invested capital. A Monte Carlo methodology models uncertainty across parameters: vehicle productivity, reliability, environmental conditions, and economic inputs. Parameters are represented by probability distributions derived from experimental data and benchmarks. Correlations between parameters are incorporated to capture system-level risk accurately. A Genetic algorithm determines optimal operating and design parameters, such as speeds and component sizing (e.g., batteries and thrusters), accounting for cascading and recursive system impacts. The simulation executes thousands of stochastic realizations to generate distributions for key outputs. This identifies expected performance, downside risks, and tail events. Sensitivity analysis quantifies the contribution of individual uncertainties to economic outcomes, guiding engineering priorities. Results show that the Monte Carlo approach provides a more robust basis for decision-making than deterministic models. The framework supports iterative refinement as technologies transition from prototype to commercial deployment. This work contributes a structured, transparent, and adaptable methodology for techno-economic assessment in complex ocean engineering systems, with applicability extending beyond deep-sea resource collection to broader autonomous offshore operations.

Keywords: Acoustic telemetry and communication, Autonomous underwater vehicles, Protection of offshore energy and critical maritime infrastructure
Abstract: Underwater acoustic communication is essential for autonomous underwater vehicle (AUV) operations used in inspection and survey tasks around subsea structures. In these environments, acoustic signal propagation is often obstructed by surrounding structures, making it difficult to maintain reliable communication links. A common approach to addressing this issue is multi‑hop communication using relay nodes to bypass obstacles. Many existing MAC protocols rely on specialized acoustic modems, which limits their practicality due to cost and implementation constraints. This study proposes a packet-driven relay scheduling method that operates with commercial off-the-shelf acoustic modems. The proposed method embeds relay transmission delay information directly into packets, enabling each node to determine its transmission timing solely based on received data. This allows temporal separation of transmissions without global time synchronization, thereby reducing packet collisions. The protocol also supports dynamic adjustment of relay order and transmission delays, making it suitable for AUV communication networks in changing environments. In the proposed protocol, the source node, which is either an AUV or a surface vessel, generates a relay schedule using its estimated position and known relay-node positions. The schedule specifies the relay order and is embedded in the packet along with a relay index indicating the current transmission stage. Upon reception, each relay determines its forwarding timing based on the relay index and its relative position. Relay nodes thus forward packets sequentially without simultaneous transmissions. The destination node can receive packets from any relay within communication range, enabling reliable communication even when obstructions exist or the destination position is unknown. Simulation results demonstrate the feasibility of the proposed approach, and ongoing field experiments aim to validate its performance in real-world environments. The method is expected to be effective for AUV operations in structurally complex underwater settings.

Keywords: Acoustic telemetry and communication, Optical communication, Systems and observatories
Abstract: Persistent subsea sensing is increasingly required for applications such as environmental monitoring, offshore infrastructure decommissioning, and carbon capture and storage (CCS) monitoring. Deploying permanent cabled observatories is costly and geographically restrictive, motivating subsea data muling, where Autonomous Underwater Vehicles (AUVs) or Remotely Operated Vehicles (ROVs) periodically retrieve data wirelessly from seabed systems. Underwater communication technologies offer complementary capabilities: acoustic links provide reliable long-range connectivity but limited data rates, while optical links enable high-bandwidth transfer over short distances with alignment constraints.

The primary challenge is therefore not the availability of communication technologies, but their integration into a unified operational system that can efficiently operate across links with widely differing bandwidth, latency, and reliability characteristics. This work presents a field-validated hybrid communication architecture, dubbed OASIS (Optical and Acoustic Subsea Integrated System), that enables adaptive data transfer across acoustic and optical links through coordinated transport, routing, and link-layer behavior. In the demonstrated use case, a seabed lander equipped with a CO2 instrument continuously acquires data that is wirelessly retrieved by a visiting vehicle in an offshore environment.

Keywords: Optical communication, Marine optics technology and instrumentation, Beam propagation
Abstract: This paper presents the design and experimental validation of a 3-degree-of-freedom (DoF) underwater simultaneous lightwave information and power transfer (U-SLIPT) system for contactless subsea intervention. The transmitter combines a 455 nm laser diode, a four-lens array whose spacing is adjusted by a linear actuator, and a two-axis gimbal. The receiver uses a series-connected photovoltaic (PV) array and a supercapacitor bank. A reduced-order spatial-overlap model represents the combined effects of beam coverage and irradiance density on the PV output, while a grid-based overlap representation retains the square receiver geometry. In a water-tank testbed, the measured maximum power point (MPP) was 18.20 V, 0.081 A, and 1.469 W, compared with a predicted 17.77 V and 1.469 W. The supercapacitor bank reached 100 J in 130 s, with an energy-curve root mean squared error (RMSE) of 5.138 J. The spatial power-map error remained below 0.05 W near the aligned region and increased to approximately 0.136 W near the PV edges.

Keywords: Vehicle navigation, Maritime domain awareness and autonomous monitoring, Sonar and transducers
Abstract: The VHF Data Exchange System (VDES), an extension of the Automatic Identification System (AIS), is being standardized as a next-generation maritime communication system for e-Navigation. In its terrestrial component, VDE-TER, shore stations manage communication resources within service areas generally designed using link budgets and theoretical propagation models such as the two-ray model. However, the relationship between theoretical and actual communication ranges in maritime environments has not been sufficiently validated for bidirectional communication. This study quantitatively evaluates VDE-TER communication range, focusing on uplink (ship-to-shore) and downlink (shore-to-ship) asymmetry. VDE-TER Link ID 19 is the primary evaluation target, with ASM Link IDs 3 and 7 evaluated for comparison. Laboratory receiver sensitivity and PER versus SNR tests, two-ray propagation analysis, and field tests in coastal waters were conducted. Received power, noise level, packet reception, and vessel speed were measured in the field. For Link ID 19, the theoretical communication range was approximately 20 km, whereas the measured ranges were approximately 8 km on the uplink and 4 km on the downlink. Similar asymmetry was observed for Link IDs 3 and 7. Although measured received power generally agreed with the two-ray model, no uplink-downlink difference in propagation loss large enough to explain the range asymmetry was observed. The noise level at the ship station was 4–6 dB higher than that at the shore station, resulting in a lower SNR on the downlink. Communication quality also improved at lower vessel speeds, suggesting that fading associated with vessel motion may affect link performance. These results indicate that VDE-TER service area design should account for theoretical communication range, noise level, fading associated with vessel motion, and bidirectional communication quality.

Keywords: Marine safety and security, Surveillance, detection, and threat-assessment techniques, Search-and-rescue technologies and emergency response
Abstract: GMDSS is designed primarily for distress alerting and response, whereas future maritime operations increasingly require communication before a situation becomes critical. This paper investigates whether proactive collision-risk assessment must rely on continuous high-rate state reporting, or whether receiver-side prediction can reduce the communication demand. A common two-vessel encounter is evaluated in four simulations. First, event-triggered reporting exposes the communication--prediction trade-off. Second, a risk-adaptive policy allocates approximately the same communication budget more effectively than periodic or residual-only reporting. Third, a primary-link outage shows that receiver-side prediction provides temporal resilience, while heterogeneous terrestrial and satellite bearers provide communication diversity. Fourth, a spoofing-like navigation drift demonstrates that prediction cannot compensate for corrupted input; an independent terrestrial A-PNT observation detects and bounds the error. The results support a goal-oriented architecture in which communication depends on prediction error, operational risk, link availability, and information integrity rather than on a fixed reporting rate.

Keywords: Modeling and dynamic analysis, Automatic control, Motion planning, Unmanned aerial system, Autonomous landing
Abstract: Unmanned aerial systems (UASs) are most often used in the marine environment for the purpose of aerial reconnaissance. This work expands the use of UASs in the marine environment by introducing an approach to enable safe and reliable autonomous landing of a UAS on the water's surface in an unknown sea state. It is assumed the UAS has very little prior information of the local wave characteristics in the landing area. Therefore, algorithms were developed to enable the UAS to determine the dominant wave period, significant wave height, and mean wave propagation direction in situ using range measurements with onboard range sensors. With the desired ocean wave characteristics identified, a method was devised to identify more persistent low wave-height segments in the wave field using onboard range sensor measurements. These more persistent low wave-height segments provide safe landing windows for the UAS and for the UAS to make in-water measurements. A staged landing algorithm was developed to ensure safe and timely UAS autonomous landing on the water surface. This study was performed using simulations and, for the simulation conditions used in this research, the algorithms were successfully tested and verified in sea states 3 and 5.

Keywords: Low-cost, scalable oceanographic sensors and platforms, Vehicle design, Autonomous underwater vehicles
Abstract: Bio-inspired, fish-like robots are promising platforms for collecting data within sensitive environments, due to the relatively less disruptive nature of oscillatory flapping propulsion. However, current fish-inspired robots often cannot match the maneuverability of their biological counterparts. This study aims to close the gap by expanding on our previous work BlueKoi, a koi- and tuna- inspired robot that captures the benefits of body flexibility with only two actuators. We introduce a rapid maneuver for angular displacements, named the unpowered turn, which utilizes a change of the robot's morphology to turn without active tail flapping. We perform a parametric analysis investigating the impact of other control variables on the turning performance, including the the time spent coasting in the unpowered turn, the speed during the initial forward motion phase, and the angular speed in which the head turns in the turn. We explore how combinations of maneuvering strategies can achieve more complex trajectories that would be relevant for obstacle avoidance within a prescribed larger area in a real-world environment. Furthermore, we extend our simplified, flat-plate model to the unpowered turn, demonstrating that the maneuver can be described through hydrodynamic equations.

Keywords: Vehicle design, Vehicle navigation, Surveillance, detection, and threat-assessment techniques
Abstract: In the scope of military and civilian applications, rapid environmental assessment (REA) and payload delivery are important deployment scenarios. However, conventional uncrewed surface vehicles (USVs) have inherent disadvantages with respect to these time-critical missions due to their generally slow speeds, limited range, and/or logistical requirements. Hydrofoiling USVs with autonomous capabilities can overcome these limitations by using lift from a submerged wing to lift the hull above the water surface. In this paper, we present Evolo, a hydrofoiling USV developed at KTH, Stockholm, Sweden, and show results from the past two years of field testing. The Evolo platform has been upgraded continuously with advanced sensors and control systems for autonomous operation. Maritime autonomy is achieved through a combination of lidar and camera for the detection and classification of obstacles. Despite hydrodynamic challenges, sonars have been integrated for seafloor imaging and underwater communication.

Keywords: Sound propagation and scattering, Ocean noise
Abstract: Cavitator geometry selection for supercavitating vehicles is commonly driven by drag performance, cavity stability, and ventilation demand. This research develops an analytic hydroacoustic evaluation of the compact loading dipole for disc, conical, and hemispherical cavitators under a fixed maximum-cavity-radius constraint. Combining cavity scaling, a normalized one-sided force spectrum, and Curle radiation gives separate closed-form figures of merit for peak pressure spectral density and integrated mean-square pressure. Under this constraint, mean cavitator drag is identical across geometries; force unsteadiness and characteristic frequency determine the acoustic ordering. An illustrative coefficient set places the disc below the 45° cone, 30° cone, and hemisphere by 1.88, 3.65, and 1.19 dB in peak level. At equal fluctuation-to-mean-drag ratio, the two cones instead fall 0.26 and 0.79 dB below the disc. Closed-form crossover laws show that the alternative geometries match the disc at relative force coefficients between 0.40 and 0.65, depending on geometry and acoustic objective. The framework enables acoustic-aware cavitator screening before CFD and defines direct force-spectrum measurement targets.

Keywords: Real-time decision systems for ocean robotics and autonomous operations, Maritime domain awareness and autonomous monitoring, Surveillance, detection, and threat-assessment techniques
Abstract: Perception systems for maritime objects are a crucial component of unmanned surface vehicles. Recent years have seen significant growth in computer vision applications for maritime environments, which have the potential to play key roles in surveillance, conservation, and safety operations. However, compared to land-based applications, several difficulties make computer vision tasks more challenging for maritime environments, including high camera motion, large variance in object distances, and frequent object occlusion. One task that is particularly difficult is Multi-Object Tracking (MOT), which is the process of detecting objects and assigning consistent object IDs throughout a sequence of images. A typical approach is tracking by detection method, in which a detection model predicts bounding boxes for each frame and a separate tracking module associates the bounding boxes between frames based on location and appearance. The goal of this study is to develop a new computer vision-based maritime object detection and tracking method with emphasis placed on (1) detection of objects of all scales at distinctly different ranges, including small objects, and (2) enhanced re-identification for salient object tracking in instances of occlusion. These goals have been achieved by first training object detection models using an in-house collected maritime dataset, then adapting the StrongSORT algorithm to better suit the maritime environment. In particular, two innovations are made on the StrongSORT algorithm: (1) the re-identification (Re-ID) model is finetuned on a dataset of maritime objects, and (2) a tracking parameter optimization approach is developed. Finetuning the Re-ID model allows the tracker to better distinguish between maritime objects with similar appearances. The tracking parameter optimization scheme then determines the best parameters such as weights and thresholds that will leverage appearance and motion costs to improve the tracking performance.

Keywords: Real-time decision systems for ocean robotics and autonomous operations, Maritime domain awareness and autonomous monitoring, Surveillance, detection, and threat-assessment techniques
Abstract: Autonomous maritime perception requires not only reliable object detection and tracking, but also accurate semantic classification across multiple levels of object specificity. While significant progress has been made in maritime object detection and tracking, fine grained vessel and non-vessel classification remains challenging due to class imbalance, visual similarity between classes, object scales, and image degradation caused by camera motion and environmental conditions. This work presents a solution with hierarchical convolutional neural network (CNN)-based maritime object classification framework designed to operate within an existing detection and tracking pipeline. The approach relies on multiple CNN classifiers with varying levels of descriptiveness. Confidence thresholds are utilized such that objects that are easy to classify will receive a more detailed classification, while objects that are less clear will receive more general classes.

Keywords: Real-time decision systems for ocean robotics and autonomous operations, Data fusion, uncertainty quantification, and predictive modeling
Abstract: Autonomous landing of an uncrewed aerial system (UAS) on a moving vessel (MV) in open waters is a challenging problem due to environmental disturbances and unreliable communication between the platforms. In this work, a vision-based approach is proposed to estimate and track the landing point on the MV. The method leverages the visual features of the MV superstructures to estimate relative pose, eliminating the need for communication or fiducial markers. A suitable kinematic motion model is used to predict the motion of the vessel, and the landing point estimates are continuously corrected by using visual observations in a filtering framework.

Keywords: AI/ML for oceanographic and environmental data analysis, Maritime domain awareness and autonomous monitoring, Real-time decision systems for ocean robotics and autonomous operations
Abstract: This paper presents a sim-to-real study of marker buoy recognition using synthetic training data generated in Unreal Engine, with the goal of identifying which visual features are most informative for real-world classification. The study considers nine buoy classes: north, east, south, west, port, starboard, safe water, isolated danger, and special. To support controlled ablation, synthetic datasets were generated using fixed random seeds such that scene background, buoy position, orientation, and viewing conditions remained constant across datasets; only the presence of top marks and color information was varied. The proposed pipeline generated 18,000 semantically labeled synthetic images with polygon annotations in under 3 hours for each study condition, including full-feature, no-top- mark, no-color, and grayscale variants. Models were then trained and compared across these conditions. Evaluation on real buoy imagery shows that synthetic data can transfer effectively to real-world recognition, but that performance depends strongly on the visual representation used during training and evalua- tion. In particular, the grayscale condition achieved the highest matched-condition performance, while grayscale-trained models performed poorly when evaluated on color imagery, indicating strong specialization to the training representation. Removal of top marks generally reduced performance, indicating that top-mark information remains an important discriminative cue. For the top-mark ablation on real imagery, top marks were removed by generative inpainting to approximate feature removal where naturally matched data were not available. The isolated danger class was comparatively underrepresented due to limited mark availability and this is reflected in class-level performance. The results suggest that robust sim-to-real buoy classification depends on preserving discriminative top-mark information and maintaining consistency between the visual representations used for synthetic training and real-world evaluation.

Keywords: Vehicle navigation, Imaging and vision, Real-time decision systems for ocean robotics and autonomous operations
Abstract: Autonomous maritime navigation requires reliable knowledge of buoy positions, yet chart-based systems fail when buoys are uncharted, temporary, or locally deployed, as in harbors, inland waterways, and competitive ASV courses. Real-time visual localization is essential.

We present a vision-based system for localizing floating navigation markers from a single monocular camera aboard an autonomous surface vehicle (ASV). The system combines fine-tuned YOLOv8 object detection with geometric triangulation and clustering to produce georeferenced buoy position estimates without requiring pre-loaded nautical charts or specialized ranging sensors such as LiDAR, radar, or stereo cameras.

A fine-tuned YOLOv8n detector converts each image into bearing angles to visible buoys. These are combined with GPS position and compass heading to generate georeferenced lines-of-sight. As the vessel moves, lines-of-sight from multiple viewpoints intersect near the true buoy location, and intersections are clustered to yield final position estimates with an empirical per-buoy uncertainty bound.

The detector was fine-tuned on a self-collected dataset of 5,443 annotated maritime images spanning lakes in Germany and Finland, the Rhine River, and Mediterranean and Baltic coastal waters, merged with a publicly available buoy dataset for 9,340 training images. YOLOv8n achieved 0.933 mAP@50 with 0.946 precision on the held-out test set and runs in real time on edge hardware comparable to the Raspberry Pi 4.

We systematically compare three clustering algorithms for aggregating noisy intersection points: K-means, DBSCAN, and HC-DBSCAN (a constrained DBSCAN variant with Bayesian hyperparameter optimization). Evaluated on 24 real-world sequences, all three achieve mean localization errors within a 3 m spread (10.2–13.2 m), indicating that accuracy is determined primarily by upstream geometric factors (e.g. heading accuracy), rather than by the clustering choice. HC-DBSCAN yields the best precision (10.2 m) but at 16× runtime, limiting it to offline use. DBSCAN is recommended as the default for real-time deployment where buoy count is unknown, combining competitive accuracy (11.6 m error) with real-time processing on small ASVs.

Keywords: Real-time decision systems for ocean robotics and autonomous operations, Classification and pattern recognition (parametric and non-parametric), Imaging and vision
Abstract: Reliable real-time pose estimation of underwater vehicles—both autonomous (AUVs) and remotely operated (ROVs)—is essential for multi-robot coordination, docking, and close-proximity underwater operations, yet remains challenging due to the scarcity of annotated underwater imagery and the strict latency constraints of onboard processing. We present an end-to-end pipeline that addresses both challenges through purely synthetic training data and efficient edge deployment on an NVIDIA Jetson platform. Rather than rendering complete scenes, we composite textured vehicle models into real 360° panoramas captured in the Baltic Sea, so that the water retains the appearance of genuine water while the vehicle can be placed at any pose. Bounding boxes, semantic keypoints and per-keypoint occlusion labels are recovered directly from a colour-coded multi-pass render, so no frame is annotated by hand and adding a vehicle costs little beyond a textured model. Per-class YOLO-NAS Pose models are trained at 640 × 480 and exported to TensorRT, covering three hull geometries: a torpedo-shaped AUV, an open-frame ROV (BlueROV2), and a shell-type ROV (Fifish V6). The complete pipeline runs onboard at 82 fps while drawing 9.1 W, with no surface link. We validate on Baltic footage recorded independently of the training data, at a different site and season, and report the gap that proved most consequential: real water absorbs far more than the rendered fog, so confidence falls off with range even where the rendered size range covers the target, and a per-channel contrast stretch at inference recovers most of the loss without retraining.

Keywords: Model-based signal processing techniques, Sonar signal processing, Geoacoustic inversion
Abstract: Dispersion curve detection using U-Net provides the locations of observed curves, but it does not provide the absolute mode number of each curve or the full set of normal mode dispersion curves including unobserved modes. In this study, the U-Net score map is matched with KRAKEN curves assigned mode numbers to construct the full set of normal mode curves. When some modes are not visible, multiple absolute mode number sequences can explain the same observation. Therefore, rather than selecting a single sequence, the method returns a small number of candidates with similar scores together with the full curve set for each candidate. The method was tested on 120 samples with the environment and propagation range placed between grid points. For all 115 samples in which the number of observed curves was estimated correctly, the mode number sequence used to generate the data was included among the final candidates. For the candidate corresponding to the generation sequence, all suppressed curves with mode numbers within the observed mode range or one mode below the lowest observed mode satisfied the completion criterion, whereas some suppressed curves two or more modes below the lowest observed mode did not. Range and time alignment after candidate selection further reduced the curve position error.

Keywords: Sonar signal processing, Array signal processing and array design
Abstract: Adaptive beamforming can potentially improve detection performance over conventional beamforming by (indirect) interference suppression towards the directions of interference signals. In the sonar literature, Capon (a.k.a MVDR) beamforming with spatial smoothing (here referred to as SS-Capon for short) has been used as a form of adaptive beamforming. In this solution, the same set of samples used for data-detection (the primary data set, PDS) is also used for the covariance matrix estimation. Another solution, is to use a secondary data set (SDS) which is assumed to share the noise covariance with the PDS, in the estimation of the noise covariance matrix. The two algorithms, adaptive-matched-filter (AMF) and generalized-likelihood-ratio-test (GLRT) which are investigated in this paper, use this principle. We also suggest a scheme to set the detection threshold of SS-Capon based on the SDS. In this paper we compare the performance of SS-Capon with AMF, GLRT and a robustified version of GLRT. The comparison is based on an elaborate ray-tracing based bi-static sonar simulation model. The results show that AMF and GLRT outperform SS-Capon if the detection threshold is ``genie aided'' to obtain a false alarm rate of 1%. However, when the detection thresholds are directly based the SDS, the false alarm rate of AMF and GLRT far exceed 1%.

Keywords: Sonar signal processing, Model-based signal processing techniques
Abstract: We consider the problem of detection and positioning of unknown underwater vehicles based on tone detections from multiple distributed hydrophones. The positioning is done by matching the tone strengths and a transmission loss map (TLM). We estimate the a posteriori distribution of the position and present it on a map. Large uncertainties in the TLM are handled by viewing the transmission loss errors as stochastic variables. An additional random variable of great significance is the power of the transmitted tone. The TLM is calculated based on the bottom type (e.g. sediment layers), the sound speed profile (i.e. the sound speed as a function of depth) and the bathymetry using acoustic propagation modeling. Experimental results are presented.

Keywords: Sonar signal processing, Classification and pattern recognition (parametric and non-parametric), AI/ML for oceanographic and environmental data analysis
Abstract: Most works on direction of arrival estimation using neural networks focus on architecture design and data representations. The impact of specific activation functions remains less explored. Motivated by the sinusoidal structure of the underlying signal model, this paper investigates whether physics-inspired periodic activation functions improve performance in DOA estimation tasks. A convolutional neural network (CNN) is trained using various activation functions, including ReLU, Leaky ReLU, ELU, tanh, Swish, GELU, sine, and cosine, while keeping all other hyperparameters fixed. Extensive simulations across different training sample sizes and training epochs show that modern smooth activations such as GELU and Swish consistently outperform both traditional and physics-inspired periodic activations. This paper studied the impact of activation functions in CNN-based DOA estimation. Contrary to intuition, physics-inspired periodic activations do not improve performance. Instead, modern smooth activations such as GELU and Swish provide superior accuracy and stability across different training conditions. The results suggest that matching activation functions to the forward signal model is not sufficient, and that optimization and generalization properties play a more critical role.

Keywords: Sonar signal processing, Surveillance, detection, and threat-assessment techniques, Sonar and transducers
Abstract: Estimating ship speed and closest-point-of-approach (CPA) range, defined as the minimum source-receiver distance, from a single hydrophone is important for passive maritime surveillance and target tracking. In shallow water, the broadband component of ship-radiated noise forms interference patterns as the source range changes, and these patterns can be described by the waveguide invariant (WI). WI-based processing can narrow the set of possible motion parameters, but ship speed and CPA range remain coupled, making it difficult to determine a unique solution. This study uses Doppler shifts observed in tonal components to distinguish the candidates remaining after WI-based processing. WI and Doppler scores are calculated for candidate speed and CPA-range pairs and combined to jointly estimate the two parameters. The method is evaluated using a 61-min single-hydrophone recording from the Shallow-water Acoustic Variability EXperiment 2015 (SAVEX15). The WI-only score forms a narrow ambiguity ridge rather than a unique speed-CPA-range solution. Doppler information helps reduce the ambiguity along this ridge, and the fused estimate agrees closely with the GPS reference, with errors of 0.45% in speed and 2.15% in CPA range.

Keywords: Buoy Technology, Oceanographic instrumentation and sensors, Current measurement technology
Abstract: From 2018 to date, the National Oceanic and Atmospheric Administration’s (NOAA) Center for Operational Oceanographic Products and Services (CO-OPS) has been pursuing an effort to design, develop, test, and transition to operations a real-time oceanographic/meteorological buoy system for use in shallow water, coastal environments. The initial version buoy system, hereinafter referred to as the CURBY (originating from CURents BuoY), was developed to support emerging measurement needs for the tidal current surveys conducted under CO-OPS’ National Currents Observing Program (NCOP). Following initial successes during tidal current survey operations in the Delaware Bay, Savannah River, and Columbia River, the CURBY system design has evolved to include new capabilities, and it has also been used to support a broader range of NOAA applications.

Updates on the CURBY system, ongoing design enhancements, and operational use have been presented in several recent articles and conference presentations. The focus of this paper is on results from the latest system field testing.

All operational deployments of the CURBY system conducted by CO-OPS to date have been in enclosed estuarine and tributary environments with limited fetch and relatively low wave energy. To assess system performance capabilities and limitations at an open ocean coast, high wave energy environment, CO-OPS partnered with the Eastern Carolina University Coastal Studies Institute (ECUCSI) to conduct a series of two CURBY test deployments at a location approximately 0.5 nautical miles offshore of Nags Head, NC.

The paper will include a detailed review of the CURBY performance throughout periods of varying wind, wave and along shore current conditions. Highlights will include mooring system performance, conditions of all hardware compliments, and an investigation of data quality in relation to dynamic conditions and buoy motion. Plans for continued testing, ongoing design enhancements and plans for future operational use will also be discussed.

Keywords: Oceanographic instrumentation and sensors, Low-cost, scalable oceanographic sensors and platforms, Oceanography: physical, geological, chemical, biological
Abstract: In 2015, NOAA’s Pacific Marine Environmental Laboratory (PMEL) developed Pop‑up floats to expand ocean observing capacity in the U.S. Arctic. Since then a total of >75 pop-up floats have been deployed, from nine research vessels over ~17 research cruises across the Bering and Chukchi Seas.

Designed as a low‑cost, lightweight (<35 kg), moored float with a modular architecture, it collects physical and biogeochemical data in four deployment phases: seafloor‑based bottom measurements, vertical water column profiles during ascent, under‑ice observations (when ice is present), and surface measurements. The light-weight modular design allows deployment by a wide range of operators and the addition or removal of sensors to meet specific mission requirements. Sensors may include photosynthetically active radiation (PAR), fluorometer, and camera systems, but all floats include two temperature sensors (one at the top and one at the bottom of the float) and one pressure sensor.

Initial field tests in 2015 characterized seasonal stratification in the Chukchi Sea. The floats now support two main purposes: first, under-ice measurements and second, mapping the cold pool that forms annually on the SE Bering Sea shelf. These observations provide a better understanding of ecosystem dynamics, and food web processes that affect fisheries dynamics. Development, deployments, and success rely on robust partnerships, such as with the U.S. Coast Guard, commercial fishing vessels, and coastal community members. Pop‑up float data contribute to broader observing initiatives and have potential integration into regional operational observing systems such as the Backyard Buoy program.

The evolving role of the pop‑up float has been an essential part of the EcoFOCI program’s observing assets. Recent updates include reduced fabrication complexity, improved release mechanisms, and increased system reliability for scalable deployments. This synthesis serves as an updated technical paper documenting design evolution, engineering refinements, and system performance since Langis et al., (2018) capturing the current state of the pop-up float platform.

Keywords: Low-cost, scalable oceanographic sensors and platforms, Systems and observatories, Citizen-science networks and participatory ocean observations
Abstract: Subsurface coastal and shelf seas remain persistently undersampled despite their dynamism and their ecological, maritime, and economic importance. Commercial fishing vessels operate continuously in exactly these waters, and fishing gear already performs a vertical profile on every set and haul. We report three regional implementations of fishing-vessel-based ocean observing operated by Ocean Data Network with regional partners, within the Fishing Vessel Ocean Observing Network (FVON), an Emerging Ocean Observing Network of the Global Ocean Observing System (GOOS): the North Atlantic Arc across Iceland, the Faroe Islands, and Scotland; Oceanography on Deck in the Bering Sea; and an early-stage pilot in California. Sensors mounted on fishing gear transfer data automatically to a deck unit and onward to shore, where quality control is applied and data are released, with a target of fifteen minutes from surfacing to public availability, including onto the WMO Global Telecommunication System. Using NOAA's Observing System Monitoring Center archive of GTS-delivered profiles for 1 May to 31 July 2026, fishing vessels delivered more near-real-time subsurface temperature profiles inside the U.S. Exclusive Economic Zone portion of the Bering Sea than all Argo floats, gliders, and profiling moorings operating there combined, by a factor of thirty-three; the same holds in Icelandic and Faroese waters. Instrumentation and data-pipeline costs correspond to roughly USD 200–310 per thousand square kilometers of shelf observed per month. We describe the operational architecture, the quality-control system, the data-sharing arrangements that in practice govern how far these data travel, and we invite modelers and applied users to test them.

Keywords: Low-cost, scalable oceanographic sensors and platforms, Vehicle performance, AI/ML for oceanographic and environmental data analysis
Abstract: Harmful Algal Blooms (HABs) pose significant environmental, economic, and public-health challenges in coastal and inland water bodies, creating a need for scalable methods to collect spatially distributed water-quality data. This paper presents an autonomous surface vessel-based sensing system for in-situ water-quality monitoring related to harmful algal bloom detection. The system integrates a BlueBoat Autonomous Surface Vessel (ASV), a motorized winch for depth-controlled sonde deployment, a multiparameter water-quality sonde, side-scan sonar, and an event-camera payload. Autonomous waypoint missions were used to collect measurements across selected sampling regions and multiple depths, enabling both horizontal and vertical characterization of environmental conditions. Field deployments were conducted in Lake Pontchartrain, Louisiana, with additional measurements collected in False River, Louisiana to increase dataset diversity. The collected measurements included algal pigment indicators, nutrient concentration, turbidity, temperature, and other water-quality parameters relevant to bloom formation. In addition, complementary sonar and event-camera data were collected during vehicle operation to provide contextual environmental information. A machine-learning classification approach was also evaluated using the collected water-quality measurements to support HAB and non-HAB classification. The results demonstrate the feasibility of using a modular ASV platform for repeatable, multi-modal water-quality data collection and support future development of autonomous monitoring systems for HAB applications.

Keywords: Oceanography: physical, geological, chemical, biological, Buoy Technology, Low-cost, scalable oceanographic sensors and platforms
Abstract: Understanding the conditions that give rise to extreme surface waves requires observations that extend below the air–sea interface. High-frequency subsurface measurements have primarily been restricted to fixed, moored platforms, limiting observations in regions where moorings are impractical or impossible. We present the Spotter Velella, a free-drifting measurement system that enables simultaneous surface wave and subsurface observations from a compact, satellite-connected platform, allowing for more robust data collection in hard-to-reach regions.

In 2025 and 2026, two fleets of 12/13 Velella units equipped with subsurface current meters were deployed in the Agulhas Current to investigate how subsurface currents and horizontal shear influence extreme wave events. Thus far, this experiment provided 135 days of wave-resolving observations from 13 of the deployed free-drifting Spotter Velellas, comprising 313 million displacement samples and 18.4 million individual waves with colocated near-surface current measurements. The observed crest-to-trough rogue-wave rate, defined by (H>2H_s), is 1.15 per 10,000 waves, approximately one third of the narrow-band Rayleigh prediction. Among 9.6 million waves with matched current observations, exceedance is independent of total current speed but increases by approximately 34% across the observed range of one-hour current-vector change. This enhancement cannot be attributed to wave blocking: opposing, converging currents show no elevated exceedance, and no observed current reaches the deep-water blocking velocity. These results demonstrate that crest and crest-to-trough rogue-wave criteria are not interchangeable and highlight the importance of wave-resolving, colocated wave and current observations for evaluating extreme-wave mechanisms.

Keywords: Oceanography: physical, geological, chemical, biological, Numerical modeling and simulation, Systems and observatories
Abstract: The Australian Bureau of Meteorology (ABoM) Ocean Model Analysis and Prediction System, OceanMAPS (OMAPS), based on the Modular Ocean Model version 5 (MOM5), is the first global ocean forecast system to operationalise an Ensemble Kalman Filter (EnKF) for data assimilation (DA). This innovation represents a significant advancement in operational ocean forecasting, delivering world-leading forecast skill and enhancing user confidence in OMAPS products for environmental monitoring, planning, and decision-making. In this study, we assess the forecast skill of the OMAPS system against a diverse suite of independent observations including global drifter buoys, fish-aggregating-devices (FADs), Surface Water and Ocean Topography (SWOT) satellite mission, and Acoustic Doppler Current Profiler (ADCP) measurements. These datasets enable evaluation of model performance across a range of spatial scales and from the ocean surface to the subsurface. Results demonstrate strong agreement between OMAPS and observed surface currents derived from both global drifters and FADs. Comparisons with SWOT-derived geostrophic currents show correlations exceeding 70% in energetic western boundary current regions, with reduced skill in less energetic areas. Subsurface validation against ADCP measurements also indicates good agreement. These results demonstrate robust, depth-dependent forecast skill, extending from the near-surface to depths of approximately 200 m, and highlight the capability of OMAPS to provide reliable ocean current forecasts across a broad range of applications.

Keywords: AI/ML for oceanographic and environmental data analysis, Foundation models and large-scale training for marine applications, Data fusion, uncertainty quantification, and predictive modeling
Abstract: Offline signal processing can insert future information into a long-lead El Niño-Southern Oscillation forecast even when the neural network is temporally ordered. This research audits and rebuilds a compact, observation-only forecasting pipeline around a strict issue-date boundary. The resulting system maps 12 monthly ERSSTv5 anomaly fields to raw monthly Niño 3.4 at 1-, 6-, 12-, and 18-month leads. Its 49,604-parameter CNN-ConvLSTM combines spatial attention, persistence-residual regression, and an optional Cahn-Hilliard-type latent update. Training ends in 1983, validation spans 1984-2003, and all 267 targets from January 2004 through March 2026 form the test series. In the archived pipeline, target-time trend restoration raised the saved 18-month correlation from 0.433 to 0.579, while two-sided splitwise wavelets formed a second future-data path. Under causal evaluation, the CH-type model records 0.448 ± 0.040 correlation and 0.851 ± 0.050 °C RMSE at six months across five fixed seeds. Direct AR(12) records 0.456 correlation and 0.841 °C RMSE on the same targets and retains the highest correlation at 12 and 18 months. Hence, the compact latent update improves six-month persistence, while the issue-date boundary separates forecast evidence from future-data leakage.

Keywords: Data fusion, uncertainty quantification, and predictive modeling, AI/ML for oceanographic and environmental data analysis, Benchmark datasets, reproducibility, and open-source frameworks
Abstract: Great Lakes water levels affect infrastructure, navigation, and shoreline management, all of which benefit from planning in advance. Deep learning (DL) models have been used to predict Great Lakes water levels, but long-horizon forecasting remains challenging because hydroclimatic processes accumulate over time. This work systematically evaluated DL models for Lake Superior water-level forecasting across input representations, historical context lengths, forecast horizons ranging from 30 to 180 days ahead, and sequence-to-one predictions versus sequence-to-sequence predictions. Hyperparameter optimization (HPO) was employed within each condition, and configurations were selected using validation RMSE. Experiments showed that the validation-selected input representation and context length differed across sequence-to-one architectures, while the overall validation-selected 180-day configuration achieved a test RMSE of 12.2 cm. The sequence-to-sequence model was comparable in overall performance to multiple sequence-to-one models, with a mean horizon-wise relative RMSE difference of 0.82% in favor of the sequence-to-sequence model. The sequence-to-sequence model had a 15.66% higher test RMSE than the sequence-to-one model at the 30-day horizon, while the sequence-to-sequence model had lower test RMSE values across longer horizons from 90 to 180 days ahead. Selected models were refit across ten independent seeds and evaluated across threshold-based water-level regimes to assess robustness. The sequence-to-sequence model achieved lower test RMSE for all water-level regimes at longer horizons (120--180 days). In conclusion, in this Lake Superior evaluation, long-horizon forecasting performance depends not only on architecture but the entirety of the forecasting problem formulation, and multi-horizon forecasting models show comparable performance to multiple sequence-to-one models, which could reduce the number of experimental conditions in future works.

Keywords: Data fusion, uncertainty quantification, and predictive modeling, Data assimilation
Abstract: Under the ongoing Pacific Infrastructure Supporting Continuous Engineering and Science (PISCES) project, near-real-time data streams from multiple mobile and fixed platforms, such as sensors from unmanned underwater vehicles (UUVs) and autonomous surface vehicles (ASVs), including Vessel-mounted acoustic Doppler current profilers (VMADCPs), are being tested to develop an integrated data network system to enhance operational model forecasts with data assimilation. VMADCPs provide high-resolution ocean current data for operational coastal ocean models across large spatial distances. When these measurements are available in near-real time, assimilation enhances the fidelity of tidal current predictions that may benefit path planning of UUVs and ASVs, improving efficiency and safety in challenging coastal and inland marine environments.

As part of PISCES activities, a field survey was conducted with a VMADCP system to measure currents in areas of the Puget Sound spanning from Hood Canal to Sequim Bay, WA covering a range of tidal currents that can reach up to 3.5 m/s. The data were quality controlled with appropriate bias corrections and then assimilated into forecasts from the regional operational ocean model Salish Sea and Columbia River Operational Forecasting System (SSCOFS)—in virtual time scale. SSCOFS is based on the finite volume community ocean model (FVCOM), developed on a high-resolution unstructured grid framework. Despite the robustness of the SSCOFS forecasts, bathymetry smoothing and limited vertical resolution to capture local features mean that current predictions from the model may not be sufficiently accurate at spatial scales of the order of autonomous sensing platforms. To address this, we developed a data assimilation (DA) framework for SSCOFS to assimilate measured current data in near-real-time by coupling to the Data Assimilation Research Testbed (DART). DART-embedded ensemble adjusted Kalman filters (EnAKF) were utilized to enhance the model’s current forecast in 1-hour assimilation windows. Although data assimilation was not operational in real time in sync with the survey, we have demonstrated that the model forecasts can be noticeably improved with EnAKF-based DA in virtual time, and

Keywords: AI/ML for oceanographic and environmental data analysis, Data fusion, uncertainty quantification, and predictive modeling, Foundation models and large-scale training for marine applications
Abstract: Large-eddy simulation (LES) reduces the cost of turbulent-flow prediction by resolving energetic motions while modeling the effects of unresolved scales through a subgrid-scale (SGS) closure. Its accuracy therefore depends on how well the closure represents fine-scale effects on the resolved flow, a particular challenge in nonstationary obstacle flows where separation and wake dynamics are time-dependent and the filtering operation interacts with immersed boundaries. To address this challenge, we develop a generalized data-driven SGS closure for two-dimensional incompressible flow past obstacles with dynamic inflow velocity fields. The closure forcing is modeled with a compact convolutional neural network (CNN), trained on filtered direct numerical simulation (fDNS) fields and conditioned on the instantaneous Reynolds number and its recent time history, while a sparse variational Gaussian process (SVGP) head models the CNN residual and provides uncertainty estimates. Training spans stationary and nonstationary flows past bluff and streamlined bodies, with generalization evaluated across unseen geometries, Reynolds-number regimes, and forcing histories. We find that a priori predictions remain strongly correlated with the reference closure forcing, while a posteriori LES rollouts remain stable and recover resolved dynamics, integral quantities, and turbulent statistics across unseen flow configurations. After the LES and fDNS trajectories decorrelate, the learned closure continues to reproduce the statistical behavior of the fDNS reference.

Keywords: Dynamic positioning, Automatic control, AI/ML for oceanographic and environmental data analysis
Abstract: Recently implemented engineering tools have improved the navigational capabilities of autonomous ocean-profiling floats. By altering dive profiles to take advantage of favorable ocean currents, tools like FloatCast and FlowPilot help maximize the value of regional ocean samples from deployed floats. This work presents a complimentary predictive tool that can inform decisions regarding fleet-level profiling strategy based on subsurface flow data from EM-APEX floats. This predictive method uses recurrent neural networks to forecast subsurface velocity profiles of EM-APEX floats, which are then integrated to estimate trajectories. The near-continuous stream of subsurface velocity training data from each EM-APEX float is discretized in a physically meaningful way and used to train the neural networks. We outline our full prediction methodology and present promising results from these subsurface velocity forecasting and integration techniques. We then isolate the contributions to the total prediction error from both the subsurface velocity forecast and the trajectory integration tools. These predictive tools provide valuable insights to operators making sampling strategy decisions.

Keywords: Naval architecture, Hydrodynamics, Vehicle performance
Abstract: Methods to provide fast and accurate predictions of hydrodynamic forces on vessels are constantly being explored. Strip theory is a method that can provide fast and accurate predictions of hydrodynamic forces vessels experience in a seaway. The work done by previous authors eliminating the ubiquitous symmetry assumption helps to align strip theory closer to reality. Strip theory is typically used to calculate the frequency dependent hydrodynamic forces and moments for each strip and then integrated along the length for the vessels total hydrodynamic forces and moments. To solve for the vessel’s motions, the equations of motion are typically also solved in the frequency domain which is acceptable for periodic and linear dynamics. However, ship motions are highly non-linear and a time domain approach is a more appropriate analysis. The basis for time domain transformations are the famous papers by Cummins and Oglivie. These methods, however, use constant hydrodynamic added mass and damping at infinity frequency A(∞) and B(∞). Due to the dynamic nature of the changing submerged geometry, this my not be sufficient when heel angles are present. Work done by Marlantes suggests that hydrodynamic predictions begin to break down at heel angles over three degrees for symmetric codes .

Keywords: Vehicle design, Naval architecture, Vehicle performance
Abstract: Early-stage ship design couples hull form, propulsion, and maneuvering, but these disciplines are commonly evaluated in separate loops, which limits the use of gradient-based optimization when the design variables change the hull surface itself. This paper describes a differentiable multidisciplinary design optimization framework in which a parametric hull, a thin-ship wave-resistance model, a Wageningen B-series propulsor with hull--propeller interaction, and a three-degree-of-freedom maneuvering simulation are assembled into a single computational graph with analytic derivatives throughout. Maneuvering criteria are event-driven: the advance and tactical diameter are defined at the instants when the ship reaches prescribed heading changes. These metrics are made compatible with gradient-based optimization by normalizing the simulation time, solving for the event time with a root-finder, and differentiating through the resulting implicit relation. The framework is demonstrated on the DTMB 5415 hull, where a joint hull and propeller optimization subject to hull-validity, propulsor-feasibility, and turning-circle constraints converges in a handful of analyses and returns a substantially lower predicted delivered power at fixed speed and near-constant displacement. This outcome supports the practical argument for treating maneuvering as a design constraint rather than a downstream check: when the criteria that a ship must ultimately satisfy are differentiable and inside the optimization, the design that emerges is one that already meets them.

Keywords: Vehicle performance, Vehicle design, Remotely operated vehicles
Abstract: Experimental evaluation of High Speed Craft (HSC) is traditionally performed using towing tanks. These types of facilities are large, expensive to run, limited in access, and may have limitations on achievable top speed. An alternative option to this is open-water testing. This offers lower testing costs and provides real-world conditions such as random wind, wave and currents, unrestricted 6-Degree of Freedom (DOF) motion, and propeller effects. In 2023, F ¨urth Laboratory built an 8-ft HSC model with onboard propulsion and data acquisition systems. Since then, work has been focused on improving the open-water testing scheme of this model. This paper focuses on the design of the new propulsion system for the model, detailing several methods for predicting the resistance of the vessel by utilizing empirical and numerical simulation methods. These methods are validated against recorded data from towing tank tests. The calculated resistance is then used to determine the powering requirements to achieve planing speed. By determining the total shaft power requirement for each motor, and ultimately providing a quantitative basis for redesigning the model’s drive train and propulsion system for future open-water testing.

Keywords: Vehicle performance, Vehicle design, Remotely operated vehicles
Abstract: Planing surface vessels operating at high speeds exhibit complex dynamics that can compromise safety and structural integrity. This work presents a low-cost, instrumented, scaled high-speed vessel (HSV) used to experimentally characterize these dynamics in free-running conditions. Time- and frequency-domain analyses reveal clear signatures of key instabilities, including porpoising and chine-walking. These results indicate that the scaled platform captures representative dynamic behavior of full-scale vessels, supporting its use as a cost-effective testbed.

Keywords: Maritime decarbonization and eco-efficient operations
Abstract: Wind-assisted propulsion systems (WAPS) have attracted increasing attention as an effective approach to reducing fuel consumption and carbon emissions in the shipping industry. Among WAPS technologies, the Flettner rotor provides auxiliary thrust by utilising the Magnus effect, while its energy saving performance is strongly influenced by seasonal wind conditions. This study develops an integrated ship energy system model incorporating the main engine, diesel generator sets, and Flettner rotors to evaluate the seasonal energy saving potential of WAPS. Representative meteorological data from February, May, August, and November are analysed to investigate the effects of wind characteristics on propulsion power demand, fuel consumption, and the performance of different rotor speed ratios. The results indicate that the Flettner rotor can reduce overall ship fuel consumption by approximately 4.17%–13.13%, with the highest savings achieved in February under more favourable wind conditions. Although November exhibits relatively high average wind speeds, less favourable wind directions significantly reduce the achievable fuel savings, demonstrating that wind direction is as important as wind speed in determining WAPS performance. In addition, the best performing rotor speed ratio varies with seasonal conditions. Among the investigated values, a rotor speed ratio of 4 provides the greatest fuel savings in May and August, whereas a lower ratio performs better in February, and November. The proposed model provides a practical framework for evaluating the seasonal performance of wind assisted propulsion systems and offers guidance for rotor operation and energy efficient ship operation.

Keywords: Autonomous underwater vehicles
Abstract: We developed a method of using two long-range AUVs (LRAUVs) with different payloads for collaborative searching and sampling of phytoplankton patches. An imaging vehicle, equipped with an in situ microscopy system, searches for phytoplankton patches on a yo-yo transect along a series of waypoints. Another vehicle follows on the same yo-yo transect carrying a water sampler. When the imaging vehicle detects a phytoplankton patch based on imaging signals, it switches from transect mode to spiral mode and sends acoustic messages about the peak signal level and the corresponding depth to the sampling vehicle. After receiving these messages, the sampling vehicle switches from transect mode to sampling mode. It acoustically tracks the imaging vehicle and triggers water sampling at the peak signal depth. Meanwhile, the imaging vehicle takes microscopic images of the surrounding water column to provide contextual optical information. In May 2026, this method was tested in Monterey Bay.

Keywords: Foundation models and large-scale training for marine applications, Vehicle navigation, Vehicle performance
Abstract: Autonomous underwater and surface vehicle teams can accomplish complex ocean tasks when their decision-making is optimized in simulation prior to deployment. However, surface vehicles primarily contribute to fleet performance by supporting underwater counterparts, making their impact difficult to measure and optimize. Prior work uses indirect reward structures to link supportive actions to downstream task success, but these methods struggle when multiple surface vehicles provide support at different points in time toward a single underwater task. In this work, we introduce Adaptive Indirect Rewards, a method that extends credit assignment across multiple supporting vehicles. Our key insight is that the value of a surface vehicle’s contribution may only be realized through subsequent support from others, requiring credit to propagate across vehicles and time. We demonstrate that our approach improves multi-vehicle coordination in a simulated task where surface vehicles assist an underwater vehicle in reaching target locations. Our contribution is a principled extension of indirect rewards for multi-vehicle, temporally distributed support, with discussion toward scaling to higher-fidelity simulation and hardware.

Keywords: Automatic control, Systems and observatories, Marine geodetic information systems
Abstract: To conduct detailed observations of the deep seafloor for wide area by a single research vessel, it is an effective approach to deploy the multiple tether-less robots simultaneously. By combining hovering autonomous underwater vehicles (AUVs) that approach within a few meters of the seafloor to take wide-area moving photographs, with landing robots on the seafloor and conduct long-term, fixed-point observations, a deep-sea environmental monitoring system can be established. However, underwater robots that reach the deep seafloor will have position estimation errors, both in inertial navigation system (INS) data and acoustic positioning data of the ultra-short baseline (USBL) from the research vessel, which affects the positional accuracy of seafloor visual observation data. In the case of multiple underwater robots operation, the positional errors of each robot will be clearly reflected in the visual observation data. As a method to correct this, the visual data can be used when the multiple underwater robots approach each other. In this study, we first used the positioning data of the research vessel's USBL positioning system for estimating two AUVs Hobalin, Hobalin2 [1] and one of three landers Edokko Mark-I type-T2, -T3, -HSG [2] near the seafloor. The used USBL positioning system is Exail GAPS [3], and its positioning accuracy is listed as 0.2% in the catalog. Based on the slant range in 1,500m depth, the position accuracy of each robot is estimated to be less than 10m. In the next step, the AUV near the seafloor performed an acoustic approach to the target lander using autonomous acoustic positioning with its onboard USBL positioning device (SeaTrac X150 [4]). The shortest target distance was set at 4m. Since the positioning accuracy is listed as 2% in the catalog, the positioning accuracy from the AUV to the lander at the closest approach is estimated to be around 0.1m. In practice, when approaching within 10m, the AUV's onboard camera can capture images of the lander, and wireless LAN communication between the AUV and the lander becomes possible using an underwater optical wireless communication device (Shimadzu MC100 [5]). In this study, we succeeded in capturing images of the lander with the AUV's onbo

Keywords: Autonomous underwater vehicles, Vehicle navigation, Automatic control
Abstract: This paper presents a cooperative localization framework that extends the operational range of visual markerbased underwater positioning through coordinated operation between an Autonomous Surface Vehicle (ASV) and an Autonomous Underwater Vehicle (AUV). The proposed mobile active marker method employs an ASV equipped with active ArUco markers that autonomously follows the AUV, thereby providing continuous visual references beyond the inherent range limitations of fixed marker installations. The AUV estimates its pose relative to the ASV using an Unscented Kalman Filter on Manifolds (UKF-M) operating on SE2(3), fusing detections of active ArUco markers illuminated by 475 nm LED strobes. The AUV’s estimated pose is communicated to the ASV, enabling autonomous tracking that keeps the markers within the AUV’s visual field; when detection fails due to depth or turbidity, the system automatically switches to USBL acoustic positioning. Water-tank experiments demonstrated sub-decimeter prediction accuracy with stable ASV tracking, targeting infrastructure inspection in structured environments without fixed underwater marker infrastructure while maintaining bounded localization error.

Keywords: Security and resilience of AUV/USV/ROV platforms, Automatic control, Integration with autonomous platforms and observatories
Abstract: Unmanned surface vehicles (USVs) and other autonomous vehicles are increasingly being deployed in maritime environments to support surveillance, reconnaissance, and monitoring missions. In many of these operations, multiple vehicles must coordinate their actions dynamically, adapting to events such as evolving objectives, resource limitations, or partial system failures. Blockchain based smart contracts (SCs) have emerged as a promising mechanism for encoding coordination rules and shared state in a decentralized and tamper-resistant manner, eliminating the need for a single trusted coordinator. However, current SC deployments typically assume relatively static logic, which contrasts with the dynamic requirements of long duration missions involving autonomous systems. This work conducts a systematic comparative study of three mechanisms for introducing dynamism into SCs used for autonomous vehicle coordination: SCs whose behavior evolves through parameter changes and finite state machines; SCs that support structural evolution via upgradeability patterns; and a hybrid approach combining both. The comparison is structured around the same mission scenarios and evaluation metrics, enabling a direct assessment of which form of dynamism best fits each mission profile. The results are intended to provide actionable guidance for the design of coordination systems in USV based and other autonomous multi-vehicle maritime applications.

Keywords: Dynamic positioning, Real-time decision systems for ocean robotics and autonomous operations, Remotely operated vehicles
Abstract: When sub-sea field work requires tight synchronization with human operators, such as in deep-sea sampling or offshore inspection, large ROVs are deployed from topside support vessels with long steel tether cables. These tethers not only serve as the ROV’s chief recovery mechanism and power source, but also provide topside human operators with high-speed video and control. The cable’s length necessitates seamless coordination with the vessel crew to ensure surface motion does not affect the ROV’s movement. If the tether is stretched too tight, the heave of the vessel can affect the ROV; if the tether is too slack, the cable may tangle or “hockle”. Applying proactive autonomous control to this multi-vehicle system will reduce human error, improve adaptability, and scale for more complex mission profiles. In this project we present a new autonomous tether behavior along with two simulation environments for evaluating vehicle movement and at-sea dynamics of vehicle pairs. While designing multi-vehicle systems is well-studied in the literature, even in heterogeneous contexts, the development of human-robot collaboration at this level requires strict and methodical care.

Keywords: Real-time decision systems for ocean robotics and autonomous operations, Vehicle performance, Automatic control
Abstract: Design of the low-level speed and heading loops of a small uncrewed surface vehicle (USV) is improved by a dynamic model of the vessel, and it is rarely clear how much model fidelity is sufficient. For vessels under seven meters with a single thruster and rudder, tow-tank testing is impractical and parameters must be identified from limited field trials. We treat sufficiency as a property of a stated design job, taking that job to be the proportional-integral plus feedforward stabilization loops of a autopilot, and identify three model levels against that target: first-order linear SISO transfer functions, a nonlinear state-space form with quadratic drag, and a NARX multilayer perceptron. First-order models reach R^2=0.93 in surge and R^2=0.83 in yaw rate, enough to fix a PI+FF compensator design. The robustness margin of a compensator is a budget for model error, and the Level-1 to Level-2 discrepancy measured against that budget bounds the range over which the simpler model suffices. A speed controller designed from it behaves as intended only between 0.35 and 1.28 m/s: below that band the discrepancy exceeds the stability margin, and above it the loop stays stable but loses more than half its intended closed-loop bandwidth. The NARX model simulates the comparison trial at R^2=0.47 in speed and 0.85 in yaw rate, settling at 1.30~m/s where the vessel does 1.22, but it plateaus near 1.3~m/s beyond the speeds its corpus covers.

Keywords: Data fusion, uncertainty quantification, and predictive modeling, Real-time decision systems for ocean robotics and autonomous operations, Numerical modeling and simulation
Abstract: Forecasting the roll and pitch motion of a planing craft over long horizons is difficult because its future motion depends not only on the current vessel state, but also on changes in speed, actuation, and environmental conditions. This study uses a controlled Neural Ordinary Differential Equation (Neural ODE) to test whether selected future signals can extend the useful forecast horizon.

The approach is evaluated in two stages. First, a synthetic benchmark with known dynamics and forcing isolates the value of future forcing information. Supplying the exact forcing path reduces 20-s roll and pitch root-mean-square error by 89.0% and 92.2%, respectively, compared with the same architecture without future forcing. Second, experiments using free-running Generic Prismatic Planing Hull data compare several candidate signals during uninterrupted 100-s rollouts: no future signal, body-frame specific force, body-frame velocity, and their combination. Future body-frame velocity is the most useful tested signal, reducing roll and pitch error by 55.3% and 53.9% relative to a persistence forecast. Delaying or replacing the velocity path reduces accuracy, while specific-force-conditioned models are less stable across model initializations.

These results show that long-horizon roll and pitch forecasting depends on both the learned vessel-response model and the information supplied about future operating conditions. Measured future velocity is not itself a deployable input; rather, it identifies information that practical systems may need to estimate from planned control actions and environmental observations.

Keywords: Data fusion, uncertainty quantification, and predictive modeling, Real-time decision systems for ocean robotics and autonomous operations, Naval architecture
Abstract: Data-driven methods to learn and understand the dynamic behavior of a ship [1-2] in real-time are becoming increasingly essential for safe operations, especially when operating conditions such as speed, cargo loading, or structural degradation push the vessel toward instability. While low sea state, straight line or gently turning maneuvers produce quasiperiodic roll, pitch, yaw, and heave signals that are readily captured by conventional time domain analysis, subtle precursors of instability often manifest as complex, nonlinear deviations that are difficult to detect with handcrafted features. In this work we propose an unsupervised deep learning framework that learns a compact latent representation of ship motion time series using an autoencoder trained on a large corpus of nominal maneuvers from the Generic Prismatic Planing Hull model test data sets [3] and a free running high speed planning hull [4]. By projecting new motion sequences into this learned embedding space and tracking their trajectories, we aim to distinguish normal operating regimes from regions associated with incipient instability, thereby providing a data-driven early warning mechanism for ship stability monitoring. This is a fundamentally different approach from the LSTM based method presented by the authors in [5] to perform fault detection of an autonomous undersea vehicle.

Keywords: Numerical modeling and simulation, Open-source hardware, software, and community innovation
Abstract: Autonomy development for marine autonomous sur- face vehicles (ASVs) remains challenging because field testing is costly and difficult to reproduce. Existing marine simulation tools provide mature capabilities in vessel dynamics, robotics integration, and visual rendering, but these strengths are empha- sized differently across platforms. This paper presents ASTRA, an ASV simulation framework built in NVIDIA Isaac Sim that combines marine vehicle dynamics, photorealistic perception, and Robot Operating System (ROS 2) autonomy interfaces within a common environment. ASTRA incorporates a Fossen-style ma- rine dynamics model together with geometry-based buoyancy and programmable spectral wave disturbances, enabling non-planar hydrostatic and wave-induced motion. The simulator reproduces the sensor configuration of a physical BlueBoat testbed, including GPS, IMU, LiDAR, and RGB-D sensing exposed through ROS 2. This allows localization, perception, planning, and control software to be developed against interfaces consistent with the physical platform. The BlueBoat’s existing ROS 2 localization and perception modules were ported directly into ASTRA, while targeted zero-propulsion experiments were used to verify the implemented wave-body coupling. The simulated roll response was consistent with the expected non-oscillatory hydrostatic behavior, and heave remained highly coherent with the generated wave field, with approximately unity gain at the dominant wave frequency in the tested sea states. ASTRA is intended as an open- source testbed for marine autonomy development, simulation- based benchmarking, and sim-to-real experimentation.

Keywords: AI/ML for oceanographic and environmental data analysis, Foundation models and large-scale training for marine applications, Real-time decision systems for ocean robotics and autonomous operations
Abstract: OceanWatch-AI introduces, to the author's knowledge, the first proposed multi-agent large language model (LLM) architecture specifically designed for autonomous marine water quality monitoring, anomaly detection, probabilistic forecasting, and regulatory advisory production. Five role-differentiated CrewAI agents collaboratively analyze real-time telemetry from NOAA CO-OPS buoys, Copernicus Sentinel-3 satellite imagery, and IoT coastal sensors. The central architectural innovation, the Oceanographic Shared Memory Protocol (OSMP), is a cross-agent temporal memory framework using a shared Redis repository with Pydantic v2 schema validation, enabling agents to reason jointly over synchronized, multi-source oceanographic data without context saturation. Target performance benchmarks include anomaly detection F1 > 0.90, dissolved oxygen MAE < 0.40 mg/L at 24 hours, HAB risk AUC > 0.90, and expert report acceptance > 90%, with end-to-end advisory latency under 30 minutes. Evaluation is planned at three U.S. coastal sites covering coral reef, hypoxic zone, and HAB environments.

Keywords: AI/ML for oceanographic and environmental data analysis, Imaging and vision, Marine safety and security
Abstract: Underwater gas blowouts produce bubble plumes in which individual bubbles overlap, deform, coalesce, and break up, making optical bubble-size measurements challenging. This study presents a prompt-guided workflow that uses the GPT Image 2 model to annotate visible bubbles without task-specific model training. The generated contours are converted into instance-level bubble masks through image processing and subsequently used for bubble segmentation and size measurements. Evaluation using synthetic and experimental images showed that the predicted bubble-size distributions generally followed the known reference distributions, with the absolute relative difference in median equivalent spherical diameter remaining below 24%. Repeated-generation tests demonstrated consistent results for most outputs, although occasional image resizing and translation were observed. Application to laboratory gas-blowout images further demonstrated that the workflow can extract bubble-size distributions from dense, highly overlapping bubble images. These results demonstrate the potential of generative image models for training-free bubble annotation and quantitative analysis, while indicating that improved output validation and contour extraction are still needed for routine measurements.

Keywords: AI/ML for oceanographic and environmental data analysis, Passive observing sensors, Sensor synergy
Abstract: Deep learning models have achieved high accuracy in marine species identification from Baited Underwater Video (BUV), yet their integration into fisheries science workflows remains limited due to insufficient interpretability. This study introduces a comprehensive post-hoc explainable AI (XAI) pipeline that provides transparent, multi-level explanations for predictions generated by a CDAS v3 ensemble combining ResNet18 and Vision Transformer (ViT) architectures.

The proposed framework consists of four complementary components: (1) Grad-CAM and ViT attention rollout to capture and compare morphological feature attribution across architectures; (2) Out-of-Distribution (OOD) saliency maps that localize pixel-level evidence driving anomaly detection; (3) a Class-conditional Homogeneity Index (CHI) that quantifies feature-space cohesion as a reliability measure; and (4) a domain-grounded large language model (LLM) that translates XAI outputs into structured, plain-language reports for fisheries scientists.

Evaluation on 3,444 validation samples across nine Northwest Atlantic species reveals critical gaps between classification accuracy and model trustworthiness. Notably, halibut/flounder achieves 100% accuracy yet exhibits a 60.6% OOD flag rate and low CHI, indicating weak feature clustering and high uncertainty. Cross-model analysis shows limited correlation between Grad-CAM and ViT attention focus, demonstrating that each architecture captures distinct morphological cues and supporting their complementary use in an ensemble.

By linking visual explanations, uncertainty quantification, and domain-grounded natural language reporting, this work advances interpretable and trustworthy AI for marine ecological monitoring and provides actionable guidance for improving dataset representation and deployment readiness in fisheries science

Keywords: AI/ML for oceanographic and environmental data analysis, Foundation models and large-scale training for marine applications
Abstract: Visual sea-state classification gives autonomous marine platforms a direct measure of surface conditions, yet its terminal decision must remain compact and deterministic. This research introduces ObserverNet, a 24,984-parameter convolutional encoder followed by a parameter-free, fixed-attractor Allen-Cahn layer. Across five architectures and ten runs per architecture, the reaction-diffusion update raises mean accuracy from 78.84% to 84.89% and reduces the mean train-test gap from 20.51% to 10.91%. State-dependent perturbations produce the smallest gap, 9.47%, while the best run reaches 87.88%. Identity attractors further reduce deployment to a closed-form rule: four-channel spatial argmax followed by a 49-site vote. Per-image symmetric INT8 quantization gives 4,273 correct predictions on the official 4,800-image MU-SSiD test set, reaching 89.02% and exceeding MUSeNet’s 88.7%. INT8 retains the float hard-projection accuracy exactly while changing only two labels. SimAVR executes the ATmega328P decision head in 3,716 cycles, equivalent to 232.25 μs at 16 MHz. The compiled image occupies 1,112 bytes of Flash, including a 784-byte test tensor. ObserverNet therefore forms a direct pathway from physics-regularized training to deterministic 8-bit decision logic.

Keywords: Data fusion, uncertainty quantification, and predictive modeling, Data visualization, Biologically inspired processing
Abstract: Confidence in classification and visualization of objects surrounding marine Uncrewed Vehicles (UxVs) is crucial for efficient autonomous UxV operation. By utilizing bio-inspired techniques, development of robotic perception based on the sensory systems found in marine fauna, specifically odontocetes (toothed whales), creates a higher-performance methodology with which UxVs perceive their world-view. For enhanced marine perception, the authors implement improvements on their previously presented algorithm, the Weighted Heuristic Algorithm for Localized Entity Evaluation (WHALEE). The results from the WHALEE updates were found to align better to real-world dolphin perception. To make full use of WHALEE, the authors also present a web-based visualization method (post-processed offline) to easily translate robotic perception, displaying the UxV’s perceived environment to the end user (i.e., the UxV operator). Initial results of the visual implementation are presented, along with areas for future work for further improvement and implementation.

Keywords: Sonar imaging, Sonar signal processing, Classification and pattern recognition (parametric and non-parametric)
Abstract: Forward-looking sonar (FLS) enables long-range underwater 3-D mapping, but reconstructing 3-D structure from FLS images is ill-posed because each return loses its elevation angle. Over a steeply rising seabed, this ambiguity accumulates across frames into an inclined surface that does not physically exist -- a false slope artifact -- distorting the geometry and causing object misclassification. Existing methods suppress it with multi-pass or multi-sensor strategies that require repeated observations or extra sensors, often impractical underwater. We instead propose an ambiguity-aware selective correction framework that uses single-pass FLS data. From the temporal seabed trajectory, it detects the regions dominated by this elevation ambiguity. Within these regions, it separates valid slopes from false ones by comparing the range gradient against a simulation reference, then corrects only the affected geometry while preserving reliable measurements. In water-tank experiments, the method substantially reduces false slope artifacts and recovers a misclassified object as correctly recognized, all without additional sensing. This enables accurate 3-D landmark estimation for AUV mapping and navigation.

Keywords: Sonar imaging, Surveillance, detection, and threat-assessment techniques, Real-time processing and machine learning for high-density sensing
Abstract: Close-range identification of mine-like contacts is the decisive stage of a mine countermeasure survey, and forward- looking sonar is the sensor of choice for it because it remains usable in turbid water. A single sonar frame does not resolve elevation, so target geometry must be recovered by fusing many viewpoints, and which viewpoints are attainable is limited in the field by current rather than by the sensor. We present a controlled synthetic benchmark that compares backprojection, an alternating direction method of multipliers (ADMM) volumetric- albedo formulation and a neural implicit signed-distance method on identical footing, at the scale of ordnance rather than the metre-scale objects for which they have previously been evaluated, and across angular coverage from sixty degrees to a full orbit. The three methods prove close in overall quality; what separates them is the character of their error, the neural method returning a complete but less precise surface, ADMM a sparse but highly precise one, and backprojection an intermediate result. Reconstruction of the observed surface is insensitive to angular coverage over the range tested, while computational cost differs by more than three orders of magnitude. For an onboard identification pass under constrained coverage, the classical method is therefore the rational default.

Keywords: Sonar imaging, Holography and 3D imaging, Data visualization
Abstract: Acoustic cameras enable underwater perception under conditions of extremely limited optical visibility by emitting multiple narrow beams and forming two-dimensional fan-shaped images from the returning echoes. However, a single acoustic image encodes only range and bearing information relative to the sensor, while elevation information is lost during projection onto the two-dimensional image plane. Building on prior work that recovers this lost elevation by rolling the acoustic camera and applying filtered back-projection (FBP) to the resulting sinogram, this paper addresses a subsequent limitation of the reconstructed point cloud: because each acoustic beam integrates returns over a finite angular aperture, certain surface regions of the target remain unsampled across all roll angles, producing discontinuous gaps whose severity increases with range due to beam divergence. Densifying the raw acquisition to close these gaps directly increases acquisition time and reconstruction cost, motivating this study of lightweight, non-learning gap-filling methods applied as a post-processing stage rather than through denser raw sampling. Three classical geometric completion strategies -- voxel-based morphological processing, Moving Least Squares surface fitting with edge augmentation, and alpha-shape surface reconstruction were applied without modification to FBP-reconstructed acoustic point clouds obtained from five underwater targets of distinct geometry, following a common 3D Total-Variation denoising stage. Results show that the three methods exhibit clearly complementary behavior. These consistent cross-target trends indicate that classical hole-filling techniques remain applicable to acoustic point clouds despite their pronounced speckle noise and non-uniform density, with method suitability depending on the intended application rather than on a single accuracy metric.

Keywords: Model-based signal processing techniques, Hydrography / Seafloor mapping / geodesy, Surveillance, detection, and threat-assessment techniques
Abstract: Gaining a comprehensive understanding of the underwater environment requires detailed three-dimensional information. Sidescan sonar is a cost-efficient acoustic survey technology for imaging large underwater areas at high resolution. However, the large elevation beamwidth responsible for large area cover-age rates also leads to a loss of explicit height information. In this work we show, that three-dimensional objects can be re-constructed from sidescan sonar images. Our proposed method integrates a differentiable sidescan sonar simulator based on optical path tracing into a gradient-based optimization routine. This optimization framework is used to reconstruct a generic three-dimensional object from simulated sidescan sonar data, achieving accuracies of approximately 5 cm corresponding to the along-track resolution of the simulated system. The results are a proof of concept that three-dimensional information can be inferred from two-dimensional sidescan sonar images with high precision. This study also serves as a starting point for investigating which parameters of underwater imaging scenar-ios can be analysed using differentiable path tracing inside a gradient-based optimization scheme.

Keywords: Sonar imaging, Remotely operated vehicles, Marine GIS and data fusion
Abstract: Sonars generate a significant amount of noise. With the advent of new technology capable of producing full 3D point clouds, the noise is amplified in sparse point clouds, making it challenging to recognize features for navigation, recognition, or reconstruction. To address this challenge, we propose using two different sonar modalities: one that produces a 2D intensity image and another that generates a 3D point cloud. By implementing auto-calibration, we can filter out noisy features between the modalities to enhance feature extraction. Experiments demonstrate that auto-calibration improves performance over manual calibration by 5% and that filtering enhances feature extraction by more than 40% relative to the raw point cloud. Code and datasets are given at https://theaprilab.org/fls-3d-calibrator.

Keywords: Marine education and outreach, Low-cost, scalable oceanographic sensors and platforms, Citizen-science networks and participatory ocean observations
Abstract: Educational robotics currently possesses a gap between introductory programs and advanced autonomous competitions, creating both an educational and financial jump for students. To address these barriers, this paper presents SeaBridge: a modular marine robotics platform that provides a scalable pathway to robotic sensing, navigation, and autonomous operation. Its hardware and software architecture supports incremental upgrades of the same base platform, rather than requiring a transition between robotics as students’ skills advance. A module incorporating scientific sampling payloads also allows students to apply their robotics skills to environmental data collection. SeaBridge provides an accessible framework for developing increasingly advanced robotics skills.

Keywords: Marine education and outreach, Ocean literacy, outreach, and education initiatives, Citizen science and community involvement
Abstract: The rapid growth of marine robotics has created a demand for new methods of attracting people to ocean engineering careers. At the intersection of design, programming, systems, and environmental science, marine robotics is an increasingly valuable way to introduce students to STEM at a young age. However, despite the growing availability of entry- and higher-level marine robotics educational opportunities, there exists a lack of accessible programming and curriculum to support the transition from the guided, step-by-step construction of a basic program, such as the SeaPerch, to the independent construction and engineering of a sophisticated autonomous vessel, as represented by the RoboBoat program. This paper details the need for an intermediate marine robotics educational program, the development process of an educational unmanned surface vehicle (USV) program to bridge this gap, as well as the constraints necessary to ensure the program is accessible to the largest possible population of students.

Keywords: Marine education and outreach, Ocean literacy, outreach, and education initiatives
Abstract: Accessibility remains a major challenge in underwater robotics, as existing platforms are typically complex, costly, and difficult to operate for non-expert users, limiting their adoption beyond professional contexts and reducing opportunities to engage younger generations with marine technologies. This paper presents a serious game (SG) built around SciURo, a bio-inspired robotic fish, designed to provide an accessible and interactive entry point into underwater robotics. A preliminary evaluation involving 62 high school students across four workshop sessions showed significant improvements in self-efficacy, perceived usefulness, and perceived ease of use, confirming the effectiveness of the approach in enhancing accessibility and engagement. Knowledge outcomes were more mixed: no item reached statistical significance, and the most conceptually demanding item remained at a low correct-response rate at both time points, suggesting that implicit gameplay alone is insufficient to convey mechanistically complex physical principles without explicit instructional support. These findings highlight both the potential and the limitations of SGs as tools for underwater robotics education, underscoring the need to balance intuitive interaction with accurate and explicit representation of real-world dynamics.

Keywords: Marine education and outreach, Vehicle design
Abstract: Student robotics competitions provide a structured way to connect engineering coursework with open-ended, field-tested systems development. This paper presents the integration of the Artificial Intelligence Maritime Maneuver Indiana Collegiate Challenge (AIMM ICC) into Indiana University’s Intelligent Systems Engineering capstone sequence. AIMM ICC requires student teams to design, build, integrate, and validate autonomy-enabled unmanned surface vehicles (USVs) under realistic outdoor conditions. The capstone maps these requirements into a two-semester design-build-test process spanning requirements definition, subsystem development, hardware selection, system integration, testing, and on-water validation. Across three cohorts, student designs progressed from an exploratory adaptation of a competition-provided vessel, to a semi-submersible marsupial architecture, to a multi-agent system strategy. Competition performance improved alongside this curricular integration, culminating in an overall championship and sweep of all AIMM ICC categories. Initial student outcomes indicate that participation supported workforce placement, continued graduate study, and returning-student knowledge transfer. These results suggest that AIMM ICC can serve as a repeatable capstone framework for maritime autonomy education and field-tested autonomous systems design.

Keywords: Calibration of acoustic systems and metrology, Sonar and transducers, Sound propagation and scattering
Abstract: With the renewed growing interest in underwater acoustics, many educational institutions have introduced related courses, creating a need for improved teaching materials and accessible laboratory tools to replace traditionally expensive and bulky sonar systems. This paper presents a low-cost experiment that is being used to teach underwater acoustics to students of the Portuguese Naval Academy, as part of the Underwater Acoustics course. The experiment consists of a water tank with two hydrophones, a signal generator and an oscilloscope for data acquisition. The full experimental laboratory guide is presented, together with the results that the students obtained. The learning outcomes successfully met the requirement of the course, giving the students an additional learning method based on hands-on experience.