Abstracts:Knowledge distillation (KD) is a predominant technique to streamline deep-learning-based recognition models for practical underwater deployments. However, existing KD methods for underwater acoustic target recognition face two problems: 1) the knowledge learning paradigm is not very consistent with the characteristics of underwater acoustics and 2) the complexity of acoustic signals in ocean environments leads to different prediction capacities in teacher and student models. This induces feature misalignment in the knowledge transfer, rendering suboptimal results. To address these problems, we propose a new distillation paradigm, i.e., channel–spatial aligned global knowledge distillation (CSGKD). Considering that the channel features (indicating the loudness of signals) and spatial features (indicating the propagation patterns of signals) in Mel spectrograms are discriminative for acoustic signal recognition, we design the knowledge-transferring scheme from “channel–spatial” aspects for effective feature extraction. Furthermore, CSGKD introduces a global multilayer alignment strategy, where all student layers collectively correspond to a single teacher layer. This allows the student model to dissect acoustic signals at a granular level, thereby capturing intricate patterns and nuances. CSGKD achieves a seamless blend of richness and efficiency, ensuring swift processing while being detail oriented. Extensive experiments on two real-world oceanic data sets confirm the superior performance of CSGKD compared to existing KD methods, i.e., achieving an accuracy (ACC) of 82.37% ($\uparrow$ 2.49% versus 79.88%). Notably, CSGKD showcases an 8.87% improvement in the ACC of the lightweight student model.

Abstracts:This article describes a compact full-waveform pulsed laser line scanner (PLLS-360°) mountable on small autonomous underwater vehicles (AUVs). The PLLS-360° is an omnidirectional imager suitable for applications, such as seafloor mapping, underwater asset inspection, object detection, ice-sheet mapping, and construction progress monitoring due to its 360° field of view (FOV). This article provides an overview of the PLLS-360°. It further discusses the system’s performance and describes how the data for this unique 360° FOV architecture is processed. The processing of the measurements involves correcting for the system receiver characteristics, geometric spreading of the returned signals, light attenuation in water, and the AUV’s motion and position to enhance the visualization of underwater 3-D scenes. This article further highlights some of the applications of the LiDAR system in ocean mapping.

Abstracts:Hourly forecasting of ocean currents is crucial for planning and control of underwater gliders and other small, underactuated underwater robots. It is a common practice to use the velocity of the depth-averaged current for estimating the ocean current surrounding the underwater robot. However, this method does not consider the spatial–temporal correlations of ocean currents, which can result in large errors. In this paper, we propose a novel deep learning fusion model, based on fused spatial–temporal graph convolutional networks (FSTGCN), which incorporates the physical properties and online measurements of ocean currents that can be modeled as Gaussian processes. The correlations between different sampling points of the 2-D ocean current data in 2-D space is constructed exploiting the incompressibility of ocean currents. Then, spatial–temporal graph convolutional networks that fuse the properties of the currents are proposed to extract their spatial–temporal dependencies. Based on measurements of underwater gliders, we further improve the proposed model to fuse model forecasts (weak constraints) and online measurements (strong constraints), which are evaluated using three different datasets and field trial data. The qualitative and quantitative results demonstrate that our fusion model outperforms competing methods and can achieve sufficient prediction accuracy in real ocean scenarios. In brief, the proposed model provides more accurate perception of ocean currents and helps underwater gliders achieve more efficient observation of ocean phenomena than other existing methods.

Abstracts:The ability of fish to turn sharply is one of the key skills for surviving in water, allowing them to more effectively tackle multiple survival challenges, including evading predators and pursuing prey. Imitating the sharp turning maneuvers of fish to improve the maneuverability of robotic fish is extremely challenging, but important for practical applications. This article proposes a motion control strategy for sharp turning maneuvers of robotic fish, which combines fish-like C-turn behavior and a variable stiffness mechanism. Based on the developed dynamic model, with the optimization goal of minimizing turning time, the parameters of C-turn motion and fishtail stiffness are optimized to improve the turning efficiency of robotic fish. Simulation and experimental results demonstrate that the proposed turning strategy for robotic fish performs well in shortening turning time and improving turning efficiency. Compared with the strategies for the case of no C-turn and fixed stiffness, the case of no C-turn and adjustable stiffness, and the case of C-turn and fixed stiffness, the turning time for the proposed turning strategy is decreased by 53.2%, 41.1%, and 27.9%, respectively.

Abstracts:This article addresses the challenges of actuator constraints in underactuated autonomous underwater vehicles (AUVs) during the design of time-varying formation controllers. It focuses specifically on actuator input saturation, actuator faults, and frequent actuator use. A novel distributed saturation and fault-tolerant time-varying formation consensus (TVFC) control system, incorporating a dynamic event-triggering mechanism, is proposed. First, to address the TVFC, a reference trajectory is generated for each AUV using neighbor information and formation information. Error tracking is performed using the dynamic equation in a noninertial polar coordinate system, with the AUV as the origin. A saturation function is employed in controller design to ensure that the actuator input remains within its upper bounds. A constrained Lyapunov function is constructed to prove the stability of the system. Subsequently, a saturated observer is used to detect disturbances and faults, compensating for actuator input errors caused by faults. The observer also estimates the severity of actuator faults based on triggered sampling outcomes. Finally, a dynamic event-triggered mechanism is implemented to reduce the operating frequency of the actuators. Simulation results demonstrate the effectiveness of the proposed control algorithm in achieving TVFC control and fault tolerance while respecting actuator input saturation constraints. Additionally, the algorithm enables effective monitoring of actuator health status and reduces actuator execution frequency. Field experiments with two underactuated AUVs partially verify the feasibility of the proposed method.

Abstracts:COLREG sets maritime navigation safety rules that are defined to be human-interpretable. With the advent of autonomous surface vehicles (ASVs), the level of detail of the rules and completeness to ensure trustworthy navigation decisions by AI algorithms have become a subject of doubt. Proper interpretation of the rules presumes professional knowledge on maritime theory and practice. In particular, it concerns situations where the interpretation and application of rules is hardened due to ambiguous contextual conditions. In this article, we study the consistency and sufficiency of a subset of COLREG rules by applying constraint logic programming and formal verification techniques to identify corner cases that do not directly contradict COLREG, however, could lead to anomalous behavior of ASVs and induce safety hazards in uncommon but still possible navigation situations. As the main contribution, we expose the conditions under which ASVs could exhibit potentially deadlock and livelock-like behaviors. Namely, we exemplify the situations where uniform control decisions under the conditions of positional and movement symmetry of the vessels could lead to either infeasible trajectories or even to dangerous approaches. The identified cases portrayed with witnessing trajectory diagrams indicate potential ambiguities in the rules and could guide rules update in the future.

Abstracts:Autonomously controlling the position of remotely operated underwater vehicles (ROVs) is of crucial importance for a wide range of underwater engineering applications, such as in the inspection and maintenance of underwater industrial structures. Consequently, studying vision-based underwater robot navigation and control has recently gained increasing attention to counter the numerous challenges faced in underwater conditions, such as lighting variability, turbidity, camera image distortions (due to bubbles), and ROV positional disturbances (due to underwater currents). In this article, we propose (to the best of the authors’ knowledge) a first rigorous unified benchmarking of more than seven machine learning-based one-shot object tracking algorithms for vision-based position locking of ROV platforms. A position-locking system is proposed which processes images of an object of interest in front of which the ROV must be kept stable. Then, our proposed system uses the output result of different object tracking algorithms to automatically correct the position of the ROV against external disturbances. Numerous real-world experiments are conducted using a BlueROV2 platform within an indoor pool in order to provide clear demonstrations of the strengths and weaknesses of each tracking approach. Finally, to help alleviate the scarcity of underwater ROV data, we release our acquired data base as open-source with the hope of benefiting future research.

Abstracts:High-accuracy localization is essential for underwater reinspection missions, which often require revisiting sites with complex structures for inspection and maintenance. Traditional localization systems—such as inertial navigation, Doppler velocity logs, and acoustic positioning—frequently fall short in accuracy or cost-effectiveness for these tasks. Machine learning-based visual relocalization presents a promising alternative, estimating poses from monocular images captured by onboard cameras using models trained on data from previous deployments. In this work, we evaluate the performance of such learning-based estimators in both clear and turbid water environments, examining the effects of color information, model architecture, and training data diversity. We further propose a novel view synthesis-based strategy to augment training data, enhancing pose estimation at previously unseen viewpoints. Finally, we improve localization robustness by fusing pose estimates with additional sensor inputs via an extended Kalman filter, resulting in smoother and more accurate trajectories.

Abstracts:The diverse structural designs of bionic autonomous underwater vehicles (AUVs) offer substantial advantages for ocean exploration, ecological and environmental monitoring, etc. A reasonable structure is crucial for the stable and precise operation of an AUV. In this article, the overall performance of the proposed ellipsoidal underwater robot (EUR), characterized by strong adaptability, flexible movement, and high antipressure ability, is evaluated by combining computational fluid dynamics (CFD) and real-world movement experiments. First, the structural design and motion of the EUR are optimized by CFD simulations, which provide favorable parameter data. Subsequently, the EUR propulsion system is analyzed by calculating velocity, thrust, and antipressure indexes of different models under same conditions. Next, dynamic performance evaluation yields optimized angles, enhancing the flexibility and efficiency of the EUR’s movements. Finally, a number of experiments, including linear, turning, and diving motions, are performed to verify the overall performance of the proposed EUR. Simultaneously, comparative experiments also show that the maneuverability of the EUR is improved. The evaluation methodology employed in this article provides favorable gradient-based parameters for robot control and state estimation, serving as a technical reference for the design of new bionic underwater robots.

Abstracts:Marine snow causes light spots and veil-like blurs in underwater images due to scattering and degrades their quality. Therefore, removing this type of noise from underwater images is necessary. However, the lack of valid and diverse marine snow image data sets has impeded the development of machine learning-based algorithms. To tackle this issue, we propose a pipeline named OmniMSI for generating marine snow images. This pipeline takes into account the diverse morphologies of marine snow particles and their natural distribution patterns. Specifically, OmniMSI initiates from collecting large-scale marine snow particle images in natural environments, classifying them by morphology, analyzing their distribution patterns, and thus creating diverse marine snow masks. The pipeline then overlays the masks on real-world underwater images without marine snow and employs the designed post-process to synthesize marine snow images. Leveraging this pipeline, we create a Multi-morphology and Multidistribution Marine Snow image Dataset (MM-MSD), including 6000 pairs of marine snow images and corresponding reference images. We also present a novel marine snow removal method—marine snow-aware diffusion, which can accurately estimate the location of marine snow in images and remove it progressively. Based on MM-MSD, we conduct a comprehensive study of the state-of-the-art algorithms qualitatively and quantitatively, developing a new benchmark for marine snow removal. The benchmark evaluations demonstrate the performance and shortcomings of existing algorithms, paving the way for further research in marine snow removal. The data set and code will be released.