(03)

Science

↓ (W25 ‘26 · Jun 15–21)
01crossref2026-06-19Tymoteusz Miller, Irmina DurlikDOI ↗ (opens in a new tab)

How Far from the Shore? Federated Maritime Intelligence for Autonomous Ship and Harbor Maneuvering (opens in a new tab)

This paper reviews research on autonomous ship and harbor maneuvering from 2015 to 2026, covering control systems, AI, sensor fusion, digital twins, and logistics. It introduces a seven-layer cyber-physical taxonomy and a maturity model to assess integration levels, from isolated vessel autonomy to fully federated systems. While progress has been made in individual areas like control and perception, gaps remain in cross-layer coordination, cybersecurity, and regulatory compliance. The authors propose a Federated Maritime Cyber-Physical Architecture to improve interoperability and resilience in autonomous harbor operations.

02openalex2026-06-20Vishnu RavendranathanDOI ↗ (opens in a new tab)

VesselNav-Bench: A Transparent, Reproducible Benchmark for Comparing Autonomous Ship Navigation Policies (opens in a new tab)

VesselNav-Bench is an open benchmark for comparing autonomous ship navigation policies, using a shared 3-DOF simulator to ensure transparent and reproducible evaluations. Seven agents, including rule-based, optimization-based, and learning-based methods, were tested across nine scenarios under calm and disturbed conditions, with 360 episodes per agent. Classical and MPC approaches performed best in calm conditions, while MPC excelled under disturbances; safety shielding improved RL policy safety but revealed persistent generalization gaps. COLREGS Rule 17 compliance was limited to 0.76 across all agents, highlighting challenges for ASV certification. The benchmark, including code and logs, is publicly available.

03openalex2026-06-16Jiale Xiang, Eduardo Blanco‐Davis, Xuri Xin, Jin Wang +1 moreDOI ↗ (opens in a new tab)

Large language model-empowered technology evolution analysis of remote operation centres of maritime autonomous surface ships (opens in a new tab)

This study introduces an LLM-based framework to analyze Remote Operation Centres (ROCs) for Maritime Autonomous Surface Ships (MASS) using patent data. It identifies 22 key technology domains, with maritime data communication, remote control, and teleoperation as the most interconnected clusters. The analysis of 973 patents shows rapid growth in ROC technologies post-2018, driven by advancements in sensing, communication, and control systems. Findings highlight the need for standardized safety measures and targeted investments to support reliable MASS operations.

04crossref2026-06-18Guoquan Xiao, Ruijie Rao, Yuanming Chen, Xiaobin HongDOI ↗ (opens in a new tab)

LDA-D3QN-Based Autonomous Navigation for Unmanned Surface Vehicles in Complex Obstacle Scenarios (opens in a new tab)

This paper introduces LDA-D3QN, an improved deep reinforcement learning method for autonomous navigation of unmanned surface vehicles (USVs) in complex obstacle scenarios. The method reduces redundant environmental data by combining target-related and local obstacle features, improving decision stability and training efficiency. Tested on a simulation platform, LDA-D3QN achieved a 91.4% success rate, outperforming other reinforcement learning algorithms like PPO and Dueling DQN. Preliminary real-world sensor data validation supports its feasibility for future autonomous navigation applications.

05crossref2026-06-15Fulko van WestrenenDOI ↗ (opens in a new tab)

On the shape, size and realisation of ship safety domains on the North Sea (opens in a new tab)

This study analyzes ship safety domains in the North Sea using AIS data from 2014, focusing on vessels over 45 meters long traveling at 5–25 knots. Results confirm that safety domains are elliptical, with the major axis expanding at higher speeds while the minor axis remains constant, and ship size does not significantly affect domain dimensions. Avoidance maneuvers rely on CPA and TCPA calculations, typically initiated 600–1800 seconds before the closest point of approach, with modest course changes to maintain the elliptical domain. The findings suggest elliptical domains could improve conflict-detection algorithms and navigation systems for both crewed and autonomous ships.

06crossref2026-06-15Dr. Vishwanath Koti, Yashwanth AradhyaDOI ↗ (opens in a new tab)

Autonomous Navigation AGV Using Vision-Language Models for Natural Language Guided Indoor Navigation (opens in a new tab)

This paper introduces an Autonomous Guided Vehicle (AGV) that uses Vision-Language Models to navigate indoor spaces based on natural language commands. The system, built on a Raspberry Pi 5 with a camera, ultrasonic sensors, and motors, interprets spoken instructions and combines visual and language data to identify targets and avoid obstacles. Tests in various indoor settings showed high accuracy in command recognition and reliable navigation performance. The approach improves flexibility over traditional fixed-route AGVs in dynamic environments.

07openalex2026-06-17Minmin Zheng, Shuwu Wang, Haiyang Tu, Qianyu LuDOI ↗ (opens in a new tab)

A path planning approach for ship collision avoidance integrating BRB reasoning and velocity obstacle algorithm (opens in a new tab)

This study combines Belief-Rule-Based (BRB) reasoning and the Velocity Obstacle (VO) algorithm to improve ship collision avoidance in dense maritime traffic. BRB reasoning groups ships by movement trends and proximity, simplifying decision-making, while the VO algorithm generates safe avoidance paths. Safety domain radii for ships and groups are adjusted based on ship length. Testing in two scenarios demonstrated enhanced navigational safety and efficiency in congested waters.

09openalex2026-06-16Dr. B. Rajasekhar ReddyDOI ↗ (opens in a new tab)

Runtime-Slayers/POSEIDON-AI-Autonomous-Submarine-Control-Brain: POSEIDON-AI-Autonomous-Submarine-Control-Brain (opens in a new tab)

The paper introduces a deep convolutional neural network (CNN) for automated target classification in anti-submarine warfare (ASW) using passive sonar spectrograms. Trained on 120,000 simulated spectrograms across 8 target classes, the model (ResNet-18-based) achieves 94.7% accuracy, with submarine detection sensitivity at 96.2% and a 2.1% false alarm rate. It outperforms traditional DEMON detectors and identifies known and previously undocumented spectral features. The results rely on simulated data, requiring real-ocean validation before operational use.

10crossref2026-06-16Deepanshu ParnamiDOI ↗ (opens in a new tab)

Generative AI–Driven Predictive Models for Advancing Autonomous Robotic Surgery (opens in a new tab)

This paper explores how generative and predictive AI models enhance autonomous robotic surgery. It outlines the evolution of AI-driven surgical systems, from perception to control, and examines key principles like real-time prediction, multimodal fusion, and simulation-to-real transfer. The study highlights applications in tasks like suturing, dissection, and navigation, demonstrating potential for semi- and fully autonomous operations. Future research directions include surgical foundation models and hybrid physics-AI systems.

11crossref2026-06-16Arun Agrawal, Bathula Prasanna Kumar, Kiran Onapakala, Sagarika Malla +4 moreDOI ↗ (opens in a new tab)

Advancing Toward Next-Generation Autonomous Surgical Robotics (opens in a new tab)

This paper examines the progression of surgical robotics from basic mechanical systems in the 1980s to current platforms capable of interpreting tissue, regulating force, and performing tasks with minimal human input. It highlights key advancements like computer vision, force sensing, and machine learning that enable autonomy in procedures such as soft tissue manipulation and suturing. The authors address challenges like regulatory approval, ethical concerns, and training data quality as barriers to wider adoption. Future directions and their impact on surgical training, healthcare systems, and patient outcomes are also discussed.

12openalex2026-06-18Piotr Ściegienka, Łukasz Wróbel, Daniel Dąbrowski, Marcin Michalak +4 moreDOI ↗ (opens in a new tab)

Image-Based Classification of Ship Hull Cleanliness Based on Transfer Learning (opens in a new tab)

This study applies transfer learning to classify ship hull cleanliness using images, based on the ISO 8501-4 standard. Pre-trained convolutional neural networks (ResNet50, EfficientNetB0, MobileNetV2) and hybrid models with XGBoost were tested on binary (cleaned vs. uncleaned) and multi-class (WA1, WA2, WA2.5) tasks. Binary classification achieved high performance despite limited data, while multi-class classification proved difficult due to subtle class differences. The method supports automated inspection for underwater robotic cleaning systems.

13crossref2026-06-17Ivan Karin, Ivana Golub Medvešek, Joško Šoda, Oleksiy MelnykDOI ↗ (opens in a new tab)

Foundations of an integrated decision support framework for maritime aids to navigation (opens in a new tab)

This paper analyzes current methods for planning maritime Aids to Navigation (AtoNs), finding them fragmented across scientific studies, IALA guidelines, and national laws. Scientific tools are partial, IALA offers broad standards but lacks site-specific procedures, and legal frameworks define roles without clear, reproducible placement methods. The authors propose an integrated framework covering spatial, operational, and technological domains, demonstrated through a case study of a 2024 Adriatic Sea grounding. The framework aims to improve risk-informed decision-making but remains conceptual, requiring further development and validation.

14openalex2026-06-15Jeroen Pijpker, Marten Struijk, Fadi MohsenDOI ↗ (opens in a new tab)

Exposing Vulnerabilities in NMEA Gateways: Insights from Shodan and Honeypot Experiments (opens in a new tab)

This study examines vulnerabilities in NMEA gateways, critical components in maritime networks, by analyzing internet-exposed devices using Shodan. Researchers identified 4,305 unique endpoints transmitting NMEA messages, with 1,542 analyzed in detail to assess risks like GPS spoofing and autopilot manipulation. A honeypot experiment detected no direct attacks on NMEA gateways but recorded automated scanning and reconnaissance activity. Findings suggest these gateways may become prime targets if left unsecured.

16crossref2026-06-15Chijioke Leonard Nkwocha, Abhilash Kumar ChandelDOI ↗ (opens in a new tab)

Development of a low-cost UGV for autonomous crop monitoring: navigation optimization using boustrophedon path planning and pure pursuit tracking algorithms (opens in a new tab)

No abstract or body text was available, so this entry is reported from metadata only.

18crossref2026-06-18Nanxi Wang, Chenyu Liu, Xinliang Zhou, Jiaping Xiao +2 moreDOI ↗ (opens in a new tab)

Deep learning-based ship collision risk assessment under limited data: a Variational Autoencoder with Multilayer Perceptron and convolutional neural network approach (opens in a new tab)

No abstract or body text was available, so this entry is reported from metadata only.

© 2026 Ship Autonomy · Summaries are AI-generated; articles belong to their publishers · Privacy · Contact