A laden tanker meets a containership head-on at the edge of a traffic separation scheme. Rule 14 is clear — alter to starboard. But the other ship is a hampered vessel, a fishing boat is crossing the TSS astern of her, and "to starboard" now has to be weighed against a dozen unwritten factors a skilled officer computes without thinking: the ordinary practice of seamen.
That seamanship is what COLREGS actually is. The regulations are partly codified — keep clear, stand on, alter to starboard — and partly delegated to judgement, via Rule 2's obligation to use "the ordinary practice of seamen" and to take avoiding action even when doing so departs from the letter of the rules. Established
So when researchers propose putting a large language model — a system famous for fluent, confident, occasionally wrong text — anywhere near that decision loop, the honest question isn't whether LLMs can recite COLREGS. It's whether they can reason under it. The 2025–26 literature has begun answering that question seriously.
Key findings
- Purpose-built LLM agents report ~86% collision-avoidance success and ~90% rule-compliance rates in simulation — strong for the class, nowhere near shipboard-safety numbers. Research demonstration
- LLM decision latency measured at ~11 seconds versus under a second for classical methods — inside the research safety window, but an architectural problem for close-quarters. Research demonstration
- No serious architecture lets the LLM drive the ship. Every credible design sandwiches the language model between tool-use, domain constraints and a deterministic safety layer. Established
- The LLM's demonstrated value is reasoning about weird situations — the encounters no scenario library pre-codes — not executing the rules themselves. Maritime-GPT analysis
- Everything published is simulation or decision-support. Nothing has been tested in real traffic. Established
What the research actually built
The most instructive system so far is Navigation-GPT (IET Intelligent Transport Systems, 2025). Rather than asking one model to do everything, it splits the work: a large LLM core uses ReAct-style reasoning to decompose a navigation task and call external tools — position, CPA, rule lookup — while a smaller, fine-tuned core (LoRA + a rule-controlled RL method the authors call RC-GRPO) produces the COLREGS-compliant manoeuvre. A dynamics model turns the decision into a feasible trajectory; a PID controller tracks it. Research demonstration
The reported numbers are worth quoting precisely, because they describe both the promise and the limit: in the authors' scenario suite the system completed task-to-decision in 11.13 s — against 0.73 s for classical methods — and achieved an 86% collision-avoidance success rate with 90% behavioural-rule compliance, beating its untuned base model (Qwen2.5-7B) by 38 percentage points and exceeding GPT-4o and DeepSeek-R1 baselines on the same tests. A sister architecture, Nav-DLLC, reports comparable figures. Research demonstration
A different research line keeps the LLM entirely advisory: Korean maritime researchers demonstrated a stepwise prompt design that walks GPT through the collision-avoidance process as a decision aid for a manned navigator — situational assessment, route choice, execution — rather than a controller. Research demonstration
Why 86% is simultaneously impressive and disqualifying
For a language model doing multi-vessel reasoning in a simulator, 86–90% is a strong result — these models generalise to scenarios their training never enumerated, which is exactly what classical rule-based systems can't do. Maritime-GPT analysis
For a ship, it is a failing grade by orders of magnitude. A vessel executes thousands of encounters across a working life; a one-in-seven failure rate is not a safety case, it is a casualty rate. Aviation-grade decision systems talk in probabilities per flight hour of 10⁻⁵ and better. The gap between "interesting LLM benchmark" and "safety-critical maritime system" is not a tuning problem — it is a different category of assurance. Maritime-GPT analysis
And there is a subtler problem the success metric hides: the 10% non-compliance may be the right answer. COLREGS itself instructs departure from the rules when collision is imminent. A flat "compliance score" can't distinguish a model that broke a rule because it hallucinated from one that broke it because Rule 2(b) demanded it. Maritime-GPT analysis
The three failure modes that matter at sea
Hallucination is the famous one — and the least interesting, because every serious architecture already wraps the model in tool-calls and constraints precisely so the LLM never reasons over raw guesses. Navigation-GPT's ReAct core exists to force the model to fetch positions and compute CPA rather than confabulate them. Research demonstration
Latency is the quiet one. Eleven seconds is fine for a mid-ocean advisory. It is not fine when a crossing yacht appears from behind a breakwater — and it compounds if the model must re-reason as the picture evolves. Maritime-GPT analysis
Distribution shift is the dangerous one. Language models are trained on the corpus of ordinary seamanship. The situations that kill — degraded sensors, contradictory AIS, a non-COLREGS-obedient target, fog + radar shadows + a third vessel — are precisely the tail the training data under-represents. This is where the technology's core strength (fluent generalisation) becomes its hazard: confident extrapolation into the unknown. Maritime-GPT analysis
Where the sensible architecture is heading
The research consensus — across Navigation-GPT, Nav-DLLC and the prompt-design work — converges on the same shape:
Maritime-GPT synthesis of the architecture pattern common to the reviewed work — the LLM advises within a bounded envelope while deterministic safety systems and the human hold the veto. Maritime-GPT analysis
In that pattern the LLM does the thing it is actually good at — reading an ambiguous situation, gathering context, explaining options in seamen's language — while collision avoidance itself stays with provable algorithms: RL policies validated in simulation, or classical methods whose failures are enumerable. Maritime-GPT analysis
What we know / don't / being tested
What we know
- LLMs can be prompted and fine-tuned toward COLREGS-consistent reasoning, measurably beating general models.
- Tool-use + constrained cores substantially reduce — not eliminate — hallucination in these agents.
- Every peer-reviewed deployment-ready design is human-supervised or safety-gated.
What we don't know
- Whether simulation metrics transfer to real sensors, real targets and real non-compliance.
- How an LLM agent behaves in multi-ship encounters with three-plus actors — barely tested.
- How to certify a system whose reasoning core is non-deterministic under IMO/class survey regimes.
Being tested
- LLM-as-advisor patterns for manned bridges (the stepwise-prompt line of work).
- Rule-controlled RL fine-tuning to bind language reasoning to formal rule sets.
- Hybrid stacks: LLM for intent/insight, classical or RL planner for the manoeuvre.
Maritime-GPT analysis
The right question isn't "LLM vs. rules" — it's which layer the LLM occupies. Today's evidence says: the language model belongs at the sense-making and communication layer — explaining the traffic picture, surfacing the applicable rule, flagging what a junior officer might miss — while the provable system keeps the wheel. The research programmes that respect this layering will age well; the ones chasing end-to-end LLM autonomy are building benchmark champions, not ships' systems. Maritime-GPT analysis
The more interesting 2026–28 frontier is explainability as a safety feature: a collision-avoidance action that can explain itself in COLREGS terms ("stand-on vessel failed to alter; Rule 17(c) action taken") is worth more to a master than a marginally better path that can't. LLMs may end up earning their place aboard not as decision-makers but as the first navigation systems that can say why. Maritime-GPT analysis
Open research problems
- Sim-to-sea validation: no published study has run these agents against real tracked encounters from AIS archives at scale — the obvious next benchmark.
- Multi-agent conflict: interactions between two AI-assisted ships simultaneously re-reasoning about each other are essentially unstudied.
- Certification path: how a flag State or class society verifies a probabilistic reasoning component remains an open regulatory-engineering question — the MASS Code's risk-based framework is where it will be tried.
- Graceful degradation: documented behaviour when the LLM layer disagrees with the safety gate, or when connectivity/tools drop mid-encounter.
Evidence scorecard
| Dimension | Rating | Basis |
|---|---|---|
| Evidence strength | Moderate | Multiple peer-reviewed simulation studies with convergent architectures; no at-sea validation. |
| Research maturity | Experimental | Prototype agents in simulated scenarios; decision-support prototypes for crewed navigation. |
| Regulatory maturity | Established rule, unclear application | COLREGS fixed since 1972; MASS Code (2026) provides the framework where AI conformance will be tested. |
| Maritime deployment | None verified | No published operational deployment of LLM-based collision avoidance aboard a commercial vessel. |
| Confidence in conclusions | High | All performance figures quoted directly from peer-reviewed sources; uncertainty explicitly bounded. |
Related tools & reading
- The MASS Code — the regulatory frame this work lands in
- Edge AI — where onboard reasoning actually runs
- AI weather routing vs. classical optimization
- Maritime glossary — COLREGS, CPA, MASS and more
Sources
- Navigation-GPT: A Robust and Adaptive Framework Utilizing Large Language Models for Navigation Applications — Yang, Liu, Chen et al.Peer-reviewed (IET Intelligent Transport Systems)
Dual-core ReAct + LoRA/RC-GRPO framework; reports 86% collision-avoidance success, 90% behaviour compliance, 11.13 s decision latency vs 0.73 s classical. Simulation study. - Path planning and collision avoidance technologies for maritime autonomous surface ships: a review of COLREGs compliance, algorithmic trends and the Navigation-GPT framework — Journal of Navigation (Cambridge University Press)Peer-reviewed review (2025)
Organises the field into classical, reactive and learning-based families; evaluates COLREGS-compliance gaps across methods. - A Stepwise Prompt Design for GPT-based Real-Time Collision Avoidance Decision Support for Manned Ship Navigators — Namgung & Choi (Mokpo National Maritime University)Peer-reviewed (2025)
GPT as navigator decision-support rather than controller — formalised prompts for the collision-avoidance process, aimed at crewed ships. - Multi-ship collision avoidance research (2025 collection) — Frontiers in Marine SciencePeer-reviewed (2025)
State of multi-agent reinforcement-learning approaches moving toward realistic multi-vessel encounters. - Dual large language model core-driven adaptive framework for ship navigation agents (Nav-DLLC) — Chinese Journal of Ship Research (SciOpen)Peer-reviewed
Independent replication of the dual-core LLM navigation-agent pattern; comparable reported metrics. - Convention on the International Regulations for Preventing Collisions at Sea (COLREGs), 1972 — International Maritime OrganizationPrimary — regulation
The rules being automated — including Rule 2's reliance on 'the ordinary practice of seamen'.
Fact-check: every performance figure is taken verbatim from the cited paper's reported results; all describe simulation, not sea trials. Regulatory statements verified against IMO instruments (accessed 5 October 2026). Editorial analysis only — not navigation, safety or compliance advice.