Gains:
- Ability to define autonomous and remotely monitored ship (MASS) levels and the role of decision support systems
- Ability to configure collision avoidance (COLREG) logic that combines AIS, radar and ECDIS data with AI
- Ability to verify that ultimate responsibility for AI navigational decisions remains with the captain and COLREG rules
The encounter between two ships at sea is a delicate dance governed by centuries-old rules. One wrong maneuver means conflict, loss of life and major environmental disaster. Ships today; It is equipped with decision support systems that combine radar, AIS (Automatic Identification System; the system where ships broadcast their identity, location and route information), ECDIS (electronic map) and camera data; Autonomous and remotely controlled ships (MASS; Maritime Autonomous Surface Ships) are on the agenda. Artificial intelligence (AI) powers this sensor fusion, situation assessment, and maneuver recommendation. But navigation at sea is still the area where life safety is most acute.
Absolute limit: No maneuver suggested by the AI shall be carried out without confirmation within the framework of collision avoidance rules (COLREG; International Regulations for Preventing Collision at Sea) and by the assessment of the officer/master in charge of the navigation. Whatever the level of autonomy, responsibility for COLREG compliance, good maritime practice and ultimate safety lies irrevocably on the human side (or within the confines of a fully autonomous, certified and class-approved system).
Concepts: COLREG: International rules that determine how ships maneuver relative to each other. CPA/TCPA: Closest approach distance and time remaining (Closest Point of Approach / Time to CPA). Sensor fusion: Combining different sources such as radar, AIS, camera into a single situation image. MASS: Autonomous/remote controlled marine vehicle. Good seamanship: Safe behavior based on experience, beyond rules.
Sensor Fusion and Status Assessment
The first step to navigational safety is to establish an accurate picture of the surrounding traffic. Every sensor has its weaknesses: radar is affected by rain and sea waves, AIS may be off or emit false/manipulated information, camera is weak at night and in fog. By combining these sources, AI calculates location, route, speed, and CPA/TCPA for each destination. But the quality of the fusion is limited by the quality of the inputs: a small fishing boat with its AIS off or a target in the radar shadow may be missing from the system's picture. Therefore, just because the AI says "the environment is clean" does not eliminate its lookout role.
The second step is to evaluate the risk: which target carries the risk of conflict, which encounter geometry is there (head-on encounter, cross-over, catch-up). For each geometry, COLREG defines who is the "give-way" and who is the "stand-on" ship and how to maneuver. AI makes this classification fast; but COLREG requires not just geometry, but “good seamanship practice”: an early, distinct and large maneuver, a clear intention visible on radar. The AI may suggest these, but the officer interprets the spirit of the rule and the local condition.
Caution: While AIS data may appear to be a reliable source of identity, it may be turned off, misconfigured, or deliberately spoofed. It is dangerous to base a maneuver solely on AIS; Radar, visual observation and good maritime practice are used together.
COLREG Compliance and Autonomy Levels
Autonomy is not one thing; It is a ladder from decision support (human decides, system recommends) to remote control and full autonomy. Accountability and verification are different at each step, but the common thread is this: the system must comply with COLREG, and a proposed maneuver must comply with both the letter and spirit of the rule. If a maneuver suggested by the AI suggests an illegal shortcut, such as instead of turning to starboard as the yielding ship, the suggestion will be rejected. Moreover, the early and significant action required by the rules when two ships "see" each other differs from the "minor correction at the last moment" logic.
Critical safety issue: autonomous or decision-aided systems may behave unexpectedly outside the scenarios in which they were trained or tested (e.g. an unusual boat type, unusual traffic, sensor failure). For these “limit cases,” human oversight and takeover capability is critical.
encounter
COLREG role
Expected maneuver
AI's limit
encounter from the beginning
Both sides give way
Both to starboard
Officer comments if geometry is unclear
outlier transition
He who sees the other on his banner gives way
Early, significant return
Priority interpretation lies with humans
catch up
He who grows gives way
open pass
Catching up or divergence, borderline situation
Narrow channel/traffic separation
Local rules
Don't follow the lane
Local knowledge may be missing in AI
Mini Cases
Case 1 — Illegal “efficient” maneuver. A decision support system advises the ship in the yielding position to take a small turn to port as this is "efficient" in terms of fuel/time. However, COLREG expects a clear and distinct starboard maneuver in this geometry; A port turn obscures intentions for the opposing ship and increases the risk of conflict. The officer rejects the suggestion and, as a rule, turns to starboard early and distinctly. Lesson: the maneuver must first comply with COLREG; efficiency is secondary.
Case 2 — Target with AIS turned off. A system says "no AIS targets in the vicinity, navigation is clear". On radar and through binoculars, the lookout notices a small fishing boat with its AIS turned off. A dangerous approach could have occurred if AIS had been relied upon alone. Lesson: sensor fusion works with incomplete data; Lookout and radar observation are indispensable.
Case 3 — Borderline takeover. During the autonomous test run, a boat is encountered performing an unusual maneuver that the system has not seen before. The system produces a hesitant and late recommendation. The guard on the bridge notices the situation early, takes over and performs the safe maneuver manually. Lesson: human supervision and rapid takeover capability are essential for the limit state behavior of autonomous systems.
Copiable Prompt Templates
Template 1 — Encounter evaluation (COLREG framework):
Role: You are a navigational safety decision support consultant. Context (representative, for training purposes): my own ship course [x], speed [y];a target course [a], speed [b], bearing bearing [c].Task:1) Determine the type of encounter (head/out/overtaking).2) Who gives way, who keeps track, according to COLREG?3) Describe a legal, early and obvious maneuver.Constraint: This is a training analysis; The actual maneuver decision is made by the officer in charge of the navigation, COLREG, and good seamanship. Do not rely on AIS alone; Also emphasize radar/visual observation.
Template 2 — Sensor fusion gap analysis:
Evaluate a navigational situation picture (representative):radar, AIS and camera data summary [...].1) Which targets are seen with only one sensor (risk)?2) How should I handle the possibility of AIS off/suspicious targets?3) List situations where this picture may be missing (radar shadow, fog, small boat).Constraint: Never take the "surroundings clear" result as definitive; Lookout is required.
Template 3 — Autonomy takeover scenario:
Produce a draft takeover plan for an autonomous/decision-assisted navigation system:1) In what situations should the human take over (limit cases, sensor failure, unusual traffic)?2) Warning and time required for takeover.3) What is the safe default behavior when the system is unreliable?Constraint: Emphasize that ultimate safety responsibility remains with the human/certified system.
Template 4 — Maneuver recommendation check:
Check the following maneuver proposal for COLREG:[suggested maneuver and geometry].1) Does this maneuver follow the letter of the rule?2) Is it early, distinct and large (good seamanship)?3) Is the intention clear for the opposing ship?4) Does it deviate from the rule for the sake of efficiency?Mark any points that are illegal or unclear.
Weak prompt / Strong prompt
Weak prompt:
These two ships are approaching, what should I do? Tell me the shortest route.
Powerful prompt:
Role: You are a navigational safety training consultant.Context (representation): encounter geometry of my own ship and a target[description]; radar+AIS+visual data summary is attached.Task:1) Determine the encounter type and COLREG roles.2) Recommend legal, early and obvious maneuver.3) How do I verify if AIS is missing/suspicious?Constraint: Recommend the maneuver appropriate to COLREG and good seamanship, not the "shortest path"; State that the final decision lies with the navigation officer.
The weak prompt only asks for "shortest path" and ignores the rule; Powerful prompt COLREG centers on roles, authentication and human authority.
Common mistakes
- Putting efficiency ahead of COLREG. The maneuver must first comply with the rules and good seamanship; fuel/time is secondary.
- Relying solely on AIS. AIS may be down, faulty or spoofed; Radar and visual observation are essential.
- Assuming that the "environment is clean" outcome is certain. Sensor fusion works with incomplete data; Lookout is indispensable.
- Ignoring borderline situations. The autonomous system may be unreliable in the scenario it does not see; A takeover plan is required.
- Making the maneuver late and small. COLREG requires early, significant and major maneuvering; The last moment correction hides the intention.
In summary
Navigational safety and collision avoidance is the area where AI is supported by sensor fusion and situational assessment, but life safety is the sharpest. Each maneuver recommendation must comply with the letter and spirit of COLREG, AIS should not be relied upon alone, the “environment is clear” output should not eliminate lookout, and the human takeover capability for autonomous systems should be preserved. The ultimate safety responsibility lies with the navigation officer and the captain.
Application task
Create a representative encounter scenario (your ship and a target, geometry and sensor data). Have the AI apply the “encounter evaluation” template and specify the encounter type and COLREG roles. Test the proposed maneuver for compliance with the "maneuver recommendation check" template. Add the possibility of an AIS-covered target and discuss the gaps of sensor fusion. Write down who has the final decision and the terms of the takeover.
checklist
- [ ] I first evaluated the maneuver based on COLREG and good seamanship.
- [ ] I did not consider AIS the only source; I included radar and visual observation.
- [ ] I backed up the "Environment is clean" output by looking for it.
- [ ] I thought about missing data cases of sensor fusion (small boat, fog).
- [ ] I defined the takeover conditions for the autonomous/decision-supported system.
- [ ] I left the final navigation/maneuver decision to the officer in charge and the captain.