Why the Collision Regulations are Challenging for Marine Autonomous Systems reliant on Space Based PNT

This paper will examine the application of PNT based solutions, including GNSS and LEO-PNT, when proposed as part of the framework to resolve the risk of collision between marine autonomous ships (MAS) and other vessels when navigating in open ocean and then extend that consideration into other operating areas.
It will examine the International Regulations for Preventing Collisions at Sea (COLREGs) (which are widely known to the developers of MAS) and explain how their application rests within the wider duty for “the observance of good seamanship” (which may be less well understood). It will explain how the combination of these two factors drives the action required by vessels when manoeuvring to avoid collision and will discuss the uncertainty inherent in COLREG application.
It will then examine the aspects of the COLREGs which concern navigation within Traffic Separation Schemes (TSS) and Narrow Channels, two areas in which both the PNT solution available to the MAS and any manoeuvre required by the COLREGs to avoid collision are both vital to the safety of the MAS and those vessels around it.
It will discuss the challenge between the navigation of a vessel in absolute terms, required to safely navigate within the TSS or narrow channel (reliant on sources of PNT) and the requirement for the MAS to manoeuvre within the limits prescribed by the COLREGs in relative terms when taking action to avoid collision with another vessel.
It will conclude by examining the case for a PNT based anti-collision framework based on S-421 (part of the International Hydrographic Office’s S-100 standard set for marine navigation) under the broader IMO Navigation drive. S-421 supports a full spectrum of route exchange—from onboard systems to shore-based applications and Vessel Traffic Services (VTS), with the development of Very High Frequency Data Exchange (VDES) able to support high bandwidth data exchange as a facilitator.
Finally, the paper will consider the opportunities and risks which a PNT based anti-collision system offers in the context of the COLREGs, as the rules by which collisions are avoided and blame apportioned following a collision.

According to the EMSA 2024 report on marine casualties and incidents, most accidents occur in ports and coastal areas. These regions present significant navigational challenges due to high traffic density, limited visibility, and other environmental complexities. Despite these challenges, ports and coastal zones are critical for automation and remote operations, as every vessel’s journey begins and ends in such areas. The goal of automation is not only to increase productivity and reduce resource use, but also to minimize operational errors, thereby reducing incidents and accidents. This is also a key requirement for maritime automation, as automated systems must be at least as safe as, or safer than those operated by humans. The current doctrine of exclusive vessel-based sensor systems is insufficient in addressing safety in these complex environments. The effectiveness of on-board sensors is constrained by limited sensor ranges and line-of-sight occlusions arising in congested traffic conditions, which collectively undermine navigational safety. To achieve the required level of safety, novel shore-based sensor systems are a promising extension.
To address this gap, this paper proposes the concept of "Islands of Automation". This concept involves deploying shore-side sensor units across a specific area of interest. The sensor configuration is tailored to the operational needs of the location. While weather sensors, lidar, and radar are typical components, the system can integrate any sensor type required for a given scenario. These sensors transmit collected data to a central node via communication links, where the data is fused and shared with connected vessels operating within the area, enhancing navigational safety. This creates a localized, well-monitored zone, referred to as an "Island of Automation", where safe automated operations are possible. Although the concept is applicable to various maritime environments, this paper focuses specifically on port operations.
As a proof of concept, the paper presents example data from a German port where sensors are installed as a backup to global navigation satellite systems (GNSS). These sensors detect target vessels using camera images, radar, and the vessel’s radar electromagnetic signature. The collected information is used to identify and locate vessels, then relay their positions to a remote operations center or automation system. This ensures safe operations even during GNSS spoofing or jamming events, demonstrating the potential and benefits of the Island of Automation concept.

Magnetic anomaly–based navigation is emerging as a promising alternative to GNSS for maritime applications. However, its deployment is constrained by the coarse spatial resolution of global magnetic models, which provide anomaly data at multi-kilometre grid spacing. To address this limitation, various enhancement approaches have been proposed, including geostatistical interpolation methods such as kriging and minimum-curvature gridding [1], neural-network-based methods such as BPNN–triangle matching [2], GAN-based magnetic map super-resolution [3], and SLAM-based map updating using magnetic measurements [4]. Most existing methods generate deterministic magnetic predictions without uncertainty information. The absence of uncertainties limits the effectiveness of navigation filters, as they cannot distinguish reliable map regions from areas affected by sparse sampling or weak anomaly contrast. This can lead to overweighted measurement updates, filter inconsistency, or divergence.

To address this challenge, this study proposes a magnetic anomaly–based navigation framework based on Gaussian Process (GP) regression, enabling high-resolution magnetic maps with quantified uncertainty. In the proposed approach, the crustal magnetic anomaly is modelled as a spatially correlated stochastic field. Sparse magnetic observations are used to train a GP regression model that outputs both a predictive mean, representing the enhanced magnetic map, and a predictive variance, providing a location-dependent measure of map uncertainty.

The system architecture consists of two components: offline GP-based magnetic map enhancement and online navigation using a GP-adaptive EKF. A Unity-based simulation environment is used to generate vessel trajectories and assess navigation performance. Unity serves as a simulation platform for scenario generation and system evaluation.

In the offline stage, sparse magnetic measurements are used to train the GP model. Results show that the GP predictive mean closely reconstructs the underlying magnetic anomaly field, while the associated ±2σ uncertainty bands expand in regions where the EMAG2v3 map is sparse or exhibits weak anomaly gradients, explicitly indicating spatial reliability.

In the online stage, vessel trajectories are simulated in Unity, providing inertial measurement unit (IMU) data and ground-truth positions. IMU data are processed by the EKF prediction step to estimate vessel position and velocity. At each update, the predicted position is queried against the GP model to obtain the expected magnetic anomaly and its associated variance. The GP variance is used to adapt the magnetic measurement noise covariance, ensuring that magnetic updates are downweighted in regions of high uncertainty.

Simulation results in an EMAG2v3 maritime region show that the proposed GP-based approach reduces interpolation error and outperforms BPNN–triangle matching. The results demonstrate the value of uncertainty-aware magnetic map enhancement for resilient GNSS-denied maritime navigation. This work was conducted under the ESA NAVISP Element 1 MANAA project (NAVISP-EL1-078).


References
[1] Kay, M., Dimitrakopoulos, R., “Integrated Interpolation Methods for Geophysical Data: Applications to Mineral Exploration,” Natural Resources Research, 2000.
[2] Wang, Q., Zhou, J., “Triangle Matching Method for the Sparse Environment of Geomagnetic Information,” Optik, 2019.
[3] Cuenca, A., Moncayo, H., Gavilanez, G., “Artificial-Intelligence-Assisted Geomagnetic Navigation Framework,” IEEE TAES, 2025.
[4] Wang, X. et al., “Exponentially Weighted Particle Filter for Simultaneous Localization and Mapping Based on Magnetic Field Measurements,” IEEE TIM, 2017.