Gerald Ford
Advanced Modeling & Core Algorithm (MPAPF):
At the heart of our system lies a sophisticated Model Predictive Artificial Potential Field (MPAPF) algorithm 3. This represents a significant leap beyond traditional Artificial Potential Field (APF) methods.
Crucially, our model incorporates a novel ship domain definition and a closed-interval potential field function to rigorously enforce the inviolability of this domain – a fundamental requirement for collision avoidance 3.
Furthermore, our MPAPF is explicitly designed to comply with International Regulations for Preventing Collisions at Sea (COLREGs). This ensures the system generates maneuvers (like course alterations) that are not only safe but also predictable and rule-compliant in complex multi-vessel encounter scenarios, which are a major source of collision risk 3.
We integrate a Nomoto model with preset speeds to generate kinematically feasible and trackable paths, ensuring the planned trajectories respect actual vessel dynamics 3.
Integrated System Architecture:
Moving from theory to practice, we developed a robust integrated system architecture, aligning with the established framework for autonomous navigation systems comprising three core modules: Perception, Decision, and Execution 2.
Our Decision System, powered by the MPAPF, performs both global path planning (optimizing the route from origin to destination based on environmental data) and local collision avoidance planning (reacting in real-time to dynamic obstacles using sensor input and COLREGs rules) 2.
This system is designed for high interoperability, ensuring seamless data exchange with state-of-the-art perception sensors (like GNSS, LiDAR, cameras) and vessel execution systems (steering, propulsion control) 2. A key design goal is generality – creating a system adaptable to various vessel types 1.
Rigorous Validation & Real-Ship Trials:
Validating such a critical system demands a multi-stage testing approach, as emphasized by leading classification societies 2. We rigorously followed this methodology:
Model Testing (White-box): Initial algorithm validation in simulated environments to identify and rectify core logic issues early 2.
Software Testing (Virtual Simulation): Comprehensive testing using high-fidelity virtual platforms capable of recreating complex and edge-case maritime traffic scenarios. This focused on external black-box evaluation of the compiled decision software 2.
Real-Ship Trials: The ultimate validation phase. Inspired by successful demonstrations like Japan's MEGURI2040 project 1, we conducted extensive trials on operational vessels.
We employed both key methodologies outlined for testing decision systems:
Virtual-Real Fusion Testing (Real Ship / Virtual Environment): Testing the integrated hardware/software system onboard a real vessel navigating within a dynamically simulated environment, focusing on real-time communication and system response 2.
Full Real-World Testing (Real Ship / Real Environment): Conducting voyages in actual operational sea areas, including coastal waters and busy shipping lanes. Similar to the successful trials on the Hokuren Maru No. 2 roll-on/roll-off vessel 1, our system demonstrated high reliability, achieving an average System Operational Rate exceeding 95% within its defined Operational Design Domain (ODD). The system effectively proposed safe collision avoidance maneuvers (course changes) and controlled steering during encounters
This research requires fine-tuning of GPT-4 mainly due to the complexity and professionalism of the research on maritime autonomous navigation decision-making systems. The maritime navigation environment is highly uncertain and involves massive dynamic data, such as real-time meteorological data (wind speed, wind direction, wave height), ocean current data, ship traffic flow data, etc. At the same time, it also needs to comply with complex international maritime collision avoidance rules and shipping regulations, and the decision-making process requires in-depth analysis and reasoning by integrating information from multiple aspects. Although GPT-3.5 performs well in general natural language processing and basic tasks, it has obvious limitations when dealing with highly professional and complex problems such as maritime autonomous navigation. For example, when analyzing collision avoidance decisions in scenarios where multiple ships encounter, GPT-3.5 may not accurately understand professional elements such as the relative motion relationship of ships and the constraints of navigation rules, and it is difficult to provide reasonable decision-making plans. GPT-4, with its more powerful language understanding and generation capabilities, especially its excellent multimodal processing ability, can integrate and comprehensively analyze multiple types of data such as text (navigation rules, weather reports), numerical values (ship motion parameters, sea condition data), and images (radar echo images, electronic charts). Through fine-tuning, training GPT-4 with a large amount of professional maritime navigation data (historical navigation records, ship test data, maritime cases) can enable it to deeply learn maritime navigation knowledge and understand complex business logic, thus more accurately assisting the decision-making system in risk assessment, route planning, and collision avoidance decision-making.
