(03)

Science

↓ (W21 ‘26 · May 18–24)
01crossref2026-05-21Fitri MaharaniDOI ↗ (opens in a new tab)

Anchoring Legal Certainty upon the Liability Ambiguities of Autonomous Vessels under International Maritime Law (opens in a new tab)

The paper examines legal uncertainties surrounding Maritime Autonomous Surface Ships (MASS) under current international maritime law, which remains human-centered. It highlights gaps in regulations due to AI-driven operations that separate control from the physical vessel. The study proposes redefining the captain’s role, the ship’s legal status, and shifting liability from fault-based to strict or product liability. It also recommends restructuring responsibilities from individuals to systems to improve legal predictability and flag state oversight.

02openalex2026-05-20Hafez Ahmad, Padmanava DashDOI ↗ (opens in a new tab)

Resolving spatiotemporal water quality variability in a river-influenced estuarine system using high-frequency autonomous surface vessel observations (opens in a new tab)

This study analyzed multi-year (2021–2024) high-frequency water quality data from the Western Mississippi Sound using an autonomous surface vessel. Ten parameters, including chlorophyll-a, temperature, and salinity, showed significant seasonal and interannual variability, with summer peaks in temperature and algal indicators. Principal Component Analysis revealed dominant freshwater-marine mixing patterns and carbon-oxygen dynamics, while discharge-water quality relationships varied by flow conditions. Spatial analysis identified critical degradation in one transect due to eutrophication, highlighting the need for targeted coastal management.

03openalex2026-05-20Bingqi Ding, Yingjun Zhang, Lai Wei, Hongrui Lu +2 moreDOI ↗ (opens in a new tab)

A robust stereo-vision system for ship detection and localization on unmanned surface vehicles (opens in a new tab)

This paper presents a stereo-vision system for ship detection and localization on Unmanned Surface Vehicles (USVs). The system uses GS-YOLO, an improved version of YOLO11n with added modules (GLSA, BiFPN, and SIoU loss), to enhance small ship detection accuracy while reducing model parameters by 17.8%. For localization, a cascaded filtering algorithm combining median and Kalman filtering processes depth data to minimize errors caused by dynamic interference. Tests show GS-YOLO achieves 93.5% mean average precision, and the localization method reduces positioning error by up to 89.3% within 50 meters.

04openalex2026-05-18Sangseok Lee, Han Jin Lee, Wonhee LeeDOI ↗ (opens in a new tab)

AIS data-based ship trajectory prediction using a CNN–Transformer–MLP hybrid model with clustering-based feature augmentation (opens in a new tab)

This study introduces a hybrid deep learning model combining CNNs, Transformers, and MLPs to improve ship trajectory prediction using AIS data. The CNN extracts local motion patterns, while the Transformer captures long-range temporal dependencies, and the MLP decodes features for accurate forecasts. K-means clustering enhances model generalization by adding static navigational features. The model outperformed baseline methods, achieving a root mean square error of 1.56 km and a mean absolute error of 1.07 km at a 60-minute prediction horizon.

05openalex2026-05-18CƯỜNG ĐỖ VIỆT, SƠN NGUYỄN THỊ XUÂNDOI ↗ (opens in a new tab)

Tàu biển tự hành và luật biển quốc tế: Khoảng trống pháp ly và kiến nghị hoàn thiện pháp luật Việt Nam (opens in a new tab)

The paper analyzes legal challenges posed by maritime autonomous surface ships (MASS) to international maritime law, including UNCLOS, SOLAS, COLREGs, and STCW, which were not designed for autonomous vessels. It identifies gaps in the current legal framework and evaluates Vietnam’s existing regulations. The authors recommend a multi-tiered approach: creating regulatory sandboxes, updating domestic maritime laws, and engaging in international norm development for MASS. The study aims to improve Vietnam’s legal readiness for autonomous shipping.

06openalex2026-05-24Xinyu Cui, Xuanfa Jin, Xue Yan, Yongcheng Zeng +6 moreDOI ↗ (opens in a new tab)

Sim2Sea: Sim-to-Real Policy Transfer for Maritime Vessel Navigation in Congested Waters (opens in a new tab)

The paper *Sim2Sea* introduces a framework to improve autonomous maritime navigation in congested waters by addressing the gap between simulation and real-world performance. It features a GPU-accelerated simulator, a dual-stream spatiotemporal policy with safety mechanisms, and domain randomization to enhance real-world transfer. Tests show the method achieves faster training and safer navigation than existing approaches. The policy, trained only in simulation, successfully operated a 17-ton unmanned vessel in real congested waters without additional adjustments.

07openalex2026-05-20Marlon Müller, Florian Finkeldei, Hanna Krasowski, Murat Arcak +1 moreDOI ↗ (opens in a new tab)

Falsification-driven reinforcement learning for maritime motion planning (opens in a new tab)

This paper introduces a falsification-driven reinforcement learning (RL) method to improve rule compliance in autonomous maritime vessels. The approach generates adversarial training scenarios where vessels violate maritime traffic rules, expressed as signal temporal logic specifications. Experiments on open-sea navigation with two vessels show the method produces more relevant training scenarios and enhances consistent adherence to traffic rules. The study highlights the use of temporal logic robustness measures to extend maritime traffic rules.

09crossref2026-05-18David Ede, Adam Pekarski, Moritz Ziegler, Elisabeth ClausenDOI ↗ (opens in a new tab)

Safety Requirements for Mixed-Traffic Operation of Autonomous Haulage Systems in Surface Mining (opens in a new tab)

Researchers at RWTH Aachen University are developing a structured safety concept for integrating autonomous machinery in surface mining, addressing the lack of standardization in mixed-traffic operations. A key component is an integrated safety control system relying on perception sensors, though their performance under harsh mining conditions remains poorly understood. To study this, a long-term test stand was built at a quartz sand open-pit mine to evaluate sensor reliability across varying environmental influences. The work aims to improve safety and efficiency in autonomous mining operations.

10crossref2026-05-18Runze LiDOI ↗ (opens in a new tab)

WorldSense: Enabling Safe Autonomous Navigation Under Rare Event Scenarios (opens in a new tab)

WorldSense is a framework designed to improve autonomous vehicle safety in rare or high-risk driving scenarios. It combines predictive world modeling, anomaly detection, and safety-aware planning by encoding the environment with a CNN-based system and predicting future scenes using a recurrent memory module. The system detects anomalies in real time and adjusts trajectories to avoid collisions. Tests on CARLA and nuScenes show a 34.2% reduction in collisions and a 14.6 percentage point increase in route completion compared to transformer-based baselines.

11openalex2026-05-22Aud M. Wahl, Marte F. Giskeødegård, Bjarne G. Pareliussen, G LamvikDOI ↗ (opens in a new tab)

Maintenance management in shipping: the intricate interplay between organizations, technology, and humans (opens in a new tab)

This study examines how digital technology impacts maintenance management in shipping, based on input from 14 Norwegian companies. It identifies benefits like improved efficiency, data integration, and real-time collaboration but also challenges such as system compatibility, data ownership, and workforce resistance. Concerns include shifting power dynamics, reduced crew autonomy, and skill gaps in digital competencies. The authors stress the need for holistic planning, change management, and cultural adaptability to balance technology with organizational and human factors.

12crossref2026-05-18Marwan Zaghloul, Mohammed Ibrahim Awad, Omar Shehata, Diaa EmadDOI ↗ (opens in a new tab)

Reactive Local Navigation for Autonomous Exploration Using Riemannian Motion Policies and Hierarchical Wavemaps (opens in a new tab)

This paper introduces RMP-Nav, a reactive local planner for differential-drive robots that uses Riemannian Motion Policies and hierarchical wavemaps for autonomous exploration. The method reduces local-planning runtime to under 0.1 ms per call, significantly faster than the optimization-based FALCO planner, while maintaining comparable waypoint-completion performance in open environments like Campus and Forest. It also improves obstacle clearance and reduces jerk but shows moderate performance drops in constrained spaces like garages. The findings suggest reactive control is effective in geometry-heavy settings, though optimization-based approaches may still excel in topologically complex areas.

13openalex2026-05-24Alexandre Foucher, Cédric Seguin, Dominique Heller, Johann LaurentDOI ↗ (opens in a new tab)

Horizon Based Cross-View Geolocalization in Maritime Environment (opens in a new tab)

This paper presents a new method for visual geolocation in maritime environments using a 50° camera to match ground images with simulated horizons from a Digital Elevation Model. The approach employs a siamese network and compact visual descriptors to correlate images without relying on GNSS or radar. Tested over a 25 km² area, it achieves a mean localization accuracy of 38 meters and heading accuracy of up to 8°. The method enables GNSS-free navigation for autonomous maritime systems in unfamiliar or hostile environments.

14crossref2026-05-23emanuele colomboDOI ↗ (opens in a new tab)

GOVERNED AUTONOMOUS INFRASTRUCTURE (GAI) A New Paradigm for Autonomous Intelligence, Runtime Governance, and Controlled Self-Improvement (opens in a new tab)

The paper presents Governed Autonomous Infrastructure (GAI) as a new framework for managing advanced autonomous AI systems. GAI consists of four layers: Sentinel System for oversight, Adaptive Research Layer for self-improvement, Synapsis Lab for testing, and AEGIS for operational control. It emphasizes infrastructure-based governance to ensure persistent oversight, safety, and controlled adaptation. The approach aims to support future AI development in areas like multi-agent coordination and robotics.

15openalex2026-05-19William H. BoothbyDOI ↗ (opens in a new tab)

Methods of warfare and brain technologies (opens in a new tab)

The paper examines how existing laws on warfare methods apply to brain technologies. It analyzes specific tactics like blockades, unmanned systems, civilian participation, human shields, and deception operations. The study also covers space, cyber, maritime, and autonomous operations, assessing their legal implications when combined with brain technologies. Hypothetical examples illustrate the legal and practical challenges in each case.

16openalex2026-05-19Adarsh KeshriDOI ↗ (opens in a new tab)

AASTF: Agentic AI Security Testing Framework (opens in a new tab)

The AASTF is an open-source Python framework designed for automated security testing of autonomous AI agent systems. It monitors agent execution graphs in real time, introduces adversarial payloads, and assesses vulnerabilities using the OWASP Top 10 for Agentic Applications (ASI) taxonomy. The framework includes 50 pre-built attack scenarios and generates reports in SARIF, JSON, or HTML formats. It classifies outcomes into three categories: VULNERABLE, REFUSAL_ECHO, or SAFE.

18openalex2026-05-19Sardar Muhammad Ali, Ke DuDOI ↗ (opens in a new tab)

Development and Preliminary Evaluation of Low-Cost VOC Sensors for USV-Based Oil Spill Detection (opens in a new tab)

This study developed a low-cost, modular VOC sensor system for unmanned surface vehicles (USVs) to detect oil spills in marine environments. The system was tested in lab and outdoor settings, showing reliable detection of oil-contaminated conditions despite variability from airflow and ambient factors. Results confirmed the sensor’s ability to distinguish oil spills from background environments. The findings suggest potential for scalable, data-driven oil spill monitoring using USVs.

19crossref2026-05-21Yawen Zhang, Wenkui Li, Chenyang Shan, Haoyu Bu +1 moreDOI ↗ (opens in a new tab)

Task-Heterogeneous Formation Planning and Control for Unmanned Surface Vehicles Based on Hybrid Deep Reinforcement Learning (opens in a new tab)

This paper introduces a hybrid deep reinforcement learning (HDRL) framework to manage task-heterogeneous formation control for unmanned surface vehicles (USVs). The approach splits tasks into leader path planning (using single-agent DRL) and follower formation tracking (using multi-agent DRL), incorporating real USV physical constraints. A dynamic gating mechanism and mirror mapping improve control performance and policy sharing. Simulations show a 44.59% increase in planning success rate over single-agent DRL and a 21.79–39.64% improvement in tracking performance over multi-agent DRL.

22crossref2026-05-19Yichun Zeng, Mingshan XieDOI ↗ (opens in a new tab)

Goal-guided greedy experience replay-enhanced reinforcement learning for efficient autonomous navigation (opens in a new tab)

The paper by Zeng and Xie (2026) addresses inefficiencies in deep reinforcement learning (DRL) for mapless navigation caused by uniform experience sampling. Their Goal-guided Greedy Experience Replay (GER-RL) method prioritizes important experiences using non-uniform sampling to improve data utilization. Simulation experiments show GER-RL enhances navigation performance, increasing average reward, success rate, and reducing collision rate, time, and path length. The approach outperforms existing DRL-based methods in key metrics.

23openalex2026-05-18A. DhyaniDOI ↗ (opens in a new tab)

Safe Navigation of Autonomous Vessels in Inland Waterways under Uncertain and Abnormal Operational Conditions (opens in a new tab)

This paper by A. Dhyani (2026) presents a framework for safe autonomous vessel navigation in inland waterways under uncertain and abnormal conditions. It introduces a robust system identification method for vessel maneuvering models, a Nonlinear Model Predictive Control (NMPC) system for constraint-compliant navigation, and a sensor Fault Detection and Isolation (FDI) scheme to monitor and diagnose sensor faults. The study also proposes a risk mitigation approach using Bayesian Belief Networks and Partially Observable Markov Decision Processes to handle failures like collisions or grounding. Validation includes full-scale trials, simulations, and case studies demonstrating improved prediction accuracy, control performance, and fault resilience.

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