The advancement of intelligent vessel technology presents an urgent need to enhance the autonomy and safety of waterway transportation systems. This manuscript proposes a vision-based end-to-end vessel following control method, aiming to explore the potential of directly applying deep learning models to ship motion control. To address the limitations of traditional methods, which rely on precise environmental modeling and inter-vessel communication, this paper uses the lightweight neural network MobileNetV2 as the backbone to construct an end-to-end learning model that integrates heading angle classification and spatial attention supervision. The model directly predicts vessel heading control commands from stereo image inputs, enabling the visual following of a leader vessel. Through the design of a composite loss function, optimizing hyperparameters for attention weights, and conducting validation on a scaled ship model platform, the results show that the proposed model can achieve stable longitudinal following control in non-interactive scenarios, indicating promising potential for practical implementation.

Short-term ship trajectory prediction is highly sensitive to data quality and to the largely unmodeled coupling between maneuvering intention and trajectory evolution. A lightweight Seq2Seq predictor is developed that integrates intention perception with long-short term temporal attention. Intention is inferred through a hybrid module that combines rule reasoning with a neural classifier, enabling interpretable recognition of seven maneuvering intentions across three vessel categories. A long-short term temporal attention fusion module is designed to dynamically balance local details and global trends, while an intention-aware multi-branch decoder is constructed to generate trajectories under explicit maneuvering priors. Experimental results demonstrate that this method goes beyond traditional historical data fitting by establishing a explicit link between human-centered maneuvering intentions and physical trajectory evolution. This framework enhances prediction robustness and interpretability, offering a robust theoretical and technical foundation for autonomous navigation and maritime collision warning systems.

Uncertainty in vessel arrival times can substantially reduce the efficiency of berth planning in port operations. To address this issue, this study proposes a unified, data-driven, predict-then-optimize framework that explicitly links vessel arrival time (VAT) prediction with downstream continuous berth allocation optimization. In the prediction stage, heterogeneous maritime data, including port call records, AIS trajectories, and vessel physical characteristics, are integrated to construct VAT prediction models. In the optimization stage, the predicted VAT is embedded into a continuous berth allocation problem (BAP) model to support berth scheduling decisions. To better reflect real operations, a two-stage evaluation framework is further developed, in which berth plans generated from estimated arrival times (ETAs) or predicted VATs are re-evaluated under realized actual arrival times while preserving the original temporal and spatial service order. Experimental results show that the proposed framework improves VAT prediction accuracy substantially, reducing the MAE and RMSE from 4.795 h and 7.255 h for the vessel-reported ETAs to 2.844 h and 4.934 h, respectively. More importantly, the predicted-VAT-based BAP consistently outperforms the ETA-based benchmark, yielding an overall 35.96% reduction in objective value across tested scenarios. These findings demonstrate that improved VAT prediction can be effectively translated into meaningful operational gains in berth allocation.

To address the strong temporal variability of tidal currents in offshore island-reef waters and their significant influence on the energy consumption of self-propelled buoys during mission execution, this paper develops a time-varying A*-based mission-planning framework for self-propelled buoys operating in dynamic tidal-current environments. First, a gridded navigation environment is constructed based on electronic nautical charts, and a representative time-varying tidal-current field is reconstructed from observation data using the radial basis function interpolation method. On this basis, the conventional two-dimensional spatial state in the A* algorithm is extended into a spatiotemporal state, and a multi-objective cost function integrating energy consumption, navigation time, and obstacle repulsion is established to improve the adaptability of the planning results to dynamic tidal currents and complex obstacle environments. Meanwhile, an in-situ waiting mechanism is introduced, enabling the buoy to avoid high-energy-consumption route segments through temporal scheduling under unfavorable tidal conditions. Simulation results demonstrate that the proposed framework enhances the planner's capability to exploit favorable tidal conditions while simultaneously considering energy efficiency, arrival time, and navigation safety. Further planning experiments under different departure times indicate that the framework can provide support for path selection and mission execution time-window determination in self-propelled buoy mission planning.

To achieve safe and rules-compliant collision avoidance for Maritime Autonomous Surface Ships (MASS), this paper proposes a Progressive Hybrid Guidance Proximal Policy Optimization (PHG-PPO) method. This method generates expert-guided actions by integrating maneuvering actions compliant with the Convention on the International Regulations for Preventing Collisions at Sea (COLREGs) and minimum-risk actions, thereby alleviating reward sparsity and reducing ineffective exploration in complex encounter scenarios. Furthermore, a dynamic curriculum mechanism is introduced to progressively regulate the balance between expert guidance and policy-autonomous decision-making, thereby improving training stability and mitigating policy oscillation during learning. Based on real-world ship encounter data with radar and Automatic Identification System (AIS), PHG-PPO achieves stable training performance within about 800 episodes and improves overall decision performance by more than 80% compared with three benchmark learning algorithms. In 20 representative Imazu scenarios, PHG-PPO achieves a higher average compliance score than the virtual potential field method, with an improvement of 0.176, and outperforms it in 85% of the scenarios. Moreover, multi-ship simulations and real-navigation-data experiments further verify that the proposed method can generate safe and compliant collision-avoidance trajectories in challenging encounter environments.

ObjectivesThis paper provides a comprehensive review of the current research status of intelligent inland vessels, systematically analyzes the development of autonomous navigation technologies for inland ships, and identifies the key challenges and technical bottlenecks associated with autonomous navigation in critical inland waterways. Methods Based on the multi-space architecture of digital-intelligent mechanics, a novel paradigm for digital-intelligent navigation of autonomous inland vessels is proposed. This paradigm integrates theoretical modeling, physical experimentation, simulation-based computation, big data analytics, and artificial intelligence. Results A digital-intelligent navigation theoretical framework for autonomous ships operating in critical inland waterways is established, revealing the underlying mechanisms of core concepts such as cross-space dynamic modeling, nonlinear intelligent computing, and cross-domain transfer learning. Furthermore, a "ship-shore-cloud" collaborative cloud-control testing platform architecture is developed, providing a theoretical approach to overcoming the limitations of traditional single-physical-space modeling in complex inland environments. Conclusion The development directions of navigation technologies and theoretical frameworks for autonomous inland vessels in the era of intelligent navigation are outlined, providing theoretical foundations and technical support for the continued advancement and engineering implementation of next-generation autonomous navigation systems.

With the rapid advancement of artificial intelligence (AI), maritime autonomous surface ships (MASSs)—also referred to as autonomous vessels—have emerged as an important avenue of research in the maritime industry [...]

Ensuring consistent and reliable collision avoidance under perception uncertainty remains a major challenge for maritime autonomous surface ships (MASS). This paper proposes a lightweight and consistency-aware local path planning framework, called consistent Kalman virtual potential field (CKVPF), to improve decision robustness in dynamic and uncertain maritime environments. The framework combines virtual potential field-based planning with Kalman-filtered state estimation and introduces a historical consistency constraint to reduce decision fluctuations over time. CKVPF is implemented on a 45-meter-long MASS platform and tested through full-scale sea trials, with onboard real-time execution achieved using a Raspberry Pi 5. The experiments cover representative encounter scenarios, including head-on, overtaking, crossing and mixed-traffic, where the target ships exhibit unstable motion patterns. Results show that CKVPF maintains consistent decision-making and safely avoids collisions, achieving a 48.9% improvement in navigation path stability compared to baseline methods, while maintaining an average planning time of 177.8 ms. These findings demonstrate the method’s practical applicability and real-time performance for autonomous navigation under real-world uncertainty.

To mitigate risk-prone and unstable navigation behaviors caused by perception uncertainties in Target Ships (TSs) from marine radar and Automatic Identification System (AIS) sensors, this study proposes a novel Mixture Density Network (MDN)-based Hybrid-Domain Dynamic Virtual Potential Field (HDVPF) method, termed MDN-HDVPF, to provide the Own Ship (OS) with more consistent and reliable status of TS states. An MDN framework is initially designed by incorporating Gaussian Mixture Model (GMM) and neural network, which can quantify the uncertainties from TSs’ collected speed and course data with feature-based detection, distance-gating, improved loss function and condition-triggered online training. In particular, the generated confidence intervals are used to adjust the trajectories of the TSs as corrected inputs to HDVPF, which can predict the probabilistic trajectories of the TSs and the corresponding minimum potential field sequence. Based on them, the OS’s consistent collision-free trajectory is generated by MDN-HDVPF in combination of hybrid ship domain model and dynamic virtual potential field. Real-world data playback simulations and real-world experiments, comprising four representative cases conducted in the eastern coastal waters of China, together with analyses under a multi-dimensional evaluation framework encompassing safety, consistency, and efficiency, demonstrate that the proposed MDN-HDVPF method improves overall performance by 2.4%, 5.6%, 4.1% and 9.0% on average, respectively, in comparison with Dynamic Virtual Potential Field (DVPF), Rapidly-exploring Random Tree Star (RRT*), Kalman Filter-based Dynamic Virtual Potential Field (KF-DVPF), and Deep Reinforcement Learning (DRL) methods. These results indicate that the proposed method can effectively mitigate the uncertainties in TSs’ perception, thereby enabling safer and more consistent maneuvers of the OS.

To address the issue of autonomous bow-in berthing for inland vessels under environmental influence, a phased manifold-based rapidly-exploring random tree star (PMRRT*) approach is proposed for the trajectory and posture planning (TPP). A systematic framework for bow-in berthing, employing a U-turn or alongside strategy, is developed by partitioning its procedure into collision-free berthing phase and the terminal berthing phase. In the first collision-free berthing phase, the constraint manifold is incorporated to the RRT* algorithm based on water depth, flow, vessel’s dynamics. Meanwhile, the sampling space of RRT* is extended from the configuration space to the state space, and each branch samples the vessel’s position, heading and speed, all constrained by the corresponding dynamic constraint manifolds. The second terminal berthing phase initiates when the vessel is positioned close to the berth, at which point a Bezier curve is employed to design the motion trajectory, considering the orientations of the terminal points. Then, throughout the motion, the vessel’s speed and heading are allocated based on the geometric characteristics of the motion, and an inverse proportional function is applied to regulate the vessel’s speed continuous decay. Simulational experiments with actual water depth and flow data demonstrate that the proposed PMRRT* method achieves higher berthing accuracy compared to traditional RRT*, artificial potential field, and A-star methods, the final heading error remains within 8°, meanwhile achieving the lower steering frequency and reduced heading fluctuations with acceptable computational real-time performance. Moreover, the scale-model berthing experiments in real-world tank further verify the effectiveness of the PMRRT* strategy.

Guidance serves as the basis for the path following of underactuated surface vehicles (USVs). However, the complex maneuverability characteristics, varying features of desired paths, and environmental disturbances pose significant challenges for the path following of underactuated surface vehicle (USV). Therefore, a novel maneuverability-based adaptive line-of-sight (MLOS) guidance control method is proposed in this paper. The proposed MLOS guidance control introducing an adaptive Acceptance Circle radius that adapts to the features of the desired path and USV maneuverability, with its parameters optimized through Simulated Annealing (SA). Finally, simulation tests are carried out in conjunction with a predictive proportional integral derivative (PID)-based USV course-keeping controller to verify the effectiveness of the proposed MLOS guidance control under various scenarios. The results show that it integrally reduces the sailing time and path-following error by average values of 1.75% and 12.63%, respectively, thereby achieving efficient USV path following in various scenarios.

To address the challenge of intermittent communication in autonomous surface vessel (ASV) formations under unreliable wireless links, this paper proposes a prediction-based consensus formation control approach. First, a distributed formation framework is established based on graph theory, ensuring that the formation objective is formulated in a consensus form. Then, a control-law-based prediction mechanism is developed to reconstruct the behavior of neighboring vessels during communication loss, using the most recently received states together with the known structure of the formation control law. Based on the predicted neighbor information, a modified consensus control law is constructed to maintain formation cooperation under communication interruptions. Finally, simulation experiments are conducted. The results demonstrate that, compared with conventional linear extrapolation methods, the proposed approach effectively mitigates reference drift and formation distortion during communication blackout periods, and achieves smoother transitions with faster error convergence after communication recovery, validating its effectiveness under intermittent communication constraints.

The Remote Driving and Control (RDC) system constitutes a key component within the technical advancement of intelligent shipping. Its closed-loop operational framework integrates a shore-based monitoring and control platform, many shipboard perception devices, autonomous control units, and high-reliability communication systems. Comprehensive validation of the RDC system in the inland waterway serves as a critical transitional phase from theoretical development to engineering implementation in intelligent shipping. Hence, this study establishes a multi-dimensional performance encompassing key points such as course keeping accuracy, path tracking robustness, collision avoidance efficacy, communication link stability, and human-machine cooperation efficiency. Experimental results demonstrate that systematic testing can be effectively realized within the constructed hardware and software platform, contributing to enhanced operational safety and control performance of RDC systems. These findings offer empirical evidence for the iterative advancement of intelligent shipping technologies. Through the design and execution of multi-scenario experiments and system-level performance assessments, this study identifies critical technical challenges in RDC, including communication latency, bottlenecks in perceptual information fusion, consistency in decision-making, and operator cognitive load under human-in-the-loop conditions.

As ship traffic density increases in port waterways, ships are required to perform low-speed maneuvers more frequently to avoid collisions and grounding. However, reduced crew numbers and limited operational experience have significantly heightened safety risks. Therefore, research on low-speed ship maneuvering performance is essential for ensuring navigational safety. Although various hull force models have been developed, a systematic comparison of their predictive capabilities remains lacking. This study presents a comparative analysis of low-speed hull force models based on Computational Fluid Dynamics (CFD). An efficient CFD simulation method is proposed and validated through static oblique towing tests and rotating arm test. Numerical simulations are conducted under conditions involving large drift angles and high yaw rates. Based on the simulation results, model parameter identification and performance evaluation are performed for several models, including the constant-speed domain and Fourier expansion models. The findings indicate that the constant-speed domain model performs best in predicting longitudinal forces, while the Fourier expansion model demonstrates higher accuracy in capturing lateral forces and yaw moments. Combining the strengths of the constant-speed domain and Fourier expansion models holds promise for the development of a more accurate and comprehensive hydrodynamic modeling framework. This research provides a reliable theoretical foundation for simulating ship maneuvering behavior under low-speed conditions and offers new technical perspectives for maneuverability prediction and the development of automated berthing systems.

In this paper, based on the second-order nonlinear response model of ship maneuvering, a parameter identification method based on CMA-ES evolutionary algorithm was proposed to improve its reliability. The Z-type test was carried out by using the reduced ratio self-propelled ship model, and the improved CMA-ES algorithm and least square method were used to identify the test. After obtaining the parameters, the test forecast was carried out respectively, and the accuracy of the two methods was compared. The results show that the improved CMA-ES algorithm has high precision in parameter identification.

In crowded waters, especially with the presence of group targets such as multiple fishing vessels with unpredictable movements, the vessel navigation often faces challenges on the safe collision avoidance. If a vessel navigates through a fishing vessel group, it may cause high risk and even accident occurrence. To solve this problem, this paper proposes a collision avoidance method for group targets in crowded waters by integrating the Density-based spatial clustering of applications with noise (DBSCAN) Clustering algorithm with the velocity obstacle (VO), namely CVO method. DBSCAN is first employed to identify the groups of multiple target vessels, and these groups are treated as individual targets for the collision avoidance. The Closest Point of Approach (DCPA) and the Time to Closest Point of Approach (TCPA) are calculated for each target. Based on the collision risk assessment, DCPA, and TCPA, high-risk targets are identified and filtered out. Finally, the VO is employed to effectively avoid the collisions with the grouped targets. Simulation experiments demonstrate that the proposed method enhances navigation safety and efficiency, making it valuable for the high-efficiency navigation of autonomous vessel in crowded waters.

This paper presents the development and implementation of a test platform specifically designed for autonomous ship models operating in inland waterways. By integrating advanced remote driving control technologies, the platform aims to enhance the efficiency of testing and controlling intelligent ships within these environments. The system comprises two main components: the shipborne terminal and the shore-based terminal. It features a physical ship model, a sophisticated navigation system, and robust communication modules, enabling multi-scenario simulations and real-time data transmission. The platform is equipped with a range of functionalities, including heading maintenance, line-of-sight (LOS) tracking, Zigzag tests, turning tests, and a one-key return feature. Leveraging remote driving control technology, operators can maneuver and monitor the ship model in a safe and cost-effective setting, thereby significantly reducing the risks and expenses associated with traditional field trials. Our study evaluates the platform's stability, precision, and controllability through a series of tests conducted at 2400 rpm, including turning tests at angles of 35° and 25°, as well as Zigzag tests at angles of 30°/30° and 20°/20°. The results affirm the system's feasibility and lay the groundwork for future advancements in intelligent inland shipping.

In this study, we predicate the estimated arrival time of container vessels at Hong Kong Port in 2023 by leveraging a comprehensive dataset that integrates vessel arrival data, physical characteristic information, and real-time in-port vessels flow records. The dataset, comprising over 640,000 samples with 24 features, is meticulously constructed through rigorous data preprocessing, including duplicate removal, cross-source matching using unique identifiers, and extensive feature engineering with various encoding techniques. To mitigate data leakage and enhance model generalizability, we employ two partitioning strategies: a time-based split, which organizes data chronologically, and a trip-based split, which groups records by complete voyages. Based on these datasets, we implement four predictive models—eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Categorical Boosting (CatBoost), and the deep learning-based Mamba tabular model (Mambular)—optimizing the first three’s hyperparameters via five-fold cross-validation and grid search. The models are evaluated using Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE), with performance assessed by comparing the predicted remaining voyage time to the actual arrival time. Our results indicate that XGBoost achieve the best performance on the time-series dataset, while CatBoost and LightGBM excel on the voyage-based dataset. Although the deep learning model do not yield superior results in this instance, its application to tabular data represents a promising avenue for future research. These findings underscore the potential of advanced machine learning algorithms, combined with tailored data partitioning strategies, to enhance predictive accuracy in vessel arrival time forecasting and offer valuable insights for optimizing port operations and logistics planning.

Ship automatic collision avoidance is an emerging core technology for achieving assisted and even autonomous navigation with human-like decisions. Ship collision avoidance maneuvers must be implemented within their lifecycle, between the earliest and latest effective points. Investigating the latest point is crucial for clarifying the mechanism of ship collision avoidance, helping to reduce unnecessary early maneuvers. To address this issue, this study proposes a novel approach to study ship collision avoidance behaviors mechanism by introducing the concept of ship arena. We define the arena as the latest boundary for collision avoidance actions to prevent the infringement on the ship domain. By integrating the ship domain with the Velocity Obstacle (VO) method, we propose a collision risk detection method anchored in the concept of a collision zone. This method enables the calculation of the latest collision avoidance distances, i.e., the arena, for ships within this area. We develop a ship arena generation method for specific encounter scenarios and validate our proposed approaches and strategies through simulations and real Automatic Identification System (AIS) data verification. Our results show that the proposed arena method provides a robust framework for assessing collision avoidance behaviors, determining maneuver timing, and supporting decision-making processes in maritime navigation.