A DRL collision-avoidance policy is a pile of weights. It turns to starboard — correct, probably — but the officer watching it has no idea whether it saw the crossing vessel, weighed the shoal to port, or is simply doing what worked in training. In a domain where a wrong manoeuvre sinks ships, "probably correct" without a reason isn't a safety case. Explainability isn't a nice-to-have for maritime AI; it's the bridge between a prediction and a decision a human can stand behind. Maritime-GPT analysis
Key findings
- DRL avoiders can be made readable — sub-task critics, Q-value deltas and attention maps render what the AI perceived as dangerous and what it was optimising for. Research demonstration
- Inherently interpretable architectures beat bolted-on explanation — logic-programmed COLREGS gives traceable reasoning for free, no post-hoc guesswork. Research demonstration
- Contrastive explanations work — but cost attention: marine officers understood objectives better, at higher cognitive load. Research demonstration
- The winning pattern is demand-driven explanation: answers when asked, not a continuous narration. Research demonstration
- Explanation and performance trade off — the most readable systems are rarely the most capable, forcing a real engineering choice. Maritime-GPT analysis
Three ways to make a decision legible
Post-hoc explanation — open the black box after the fact. The Applied Ocean Research work is the template: decompose the DRL critic into sub-task networks so you can see which concern (which ship, which risk) drove the action, surface Q-value changes as intent, and use attention to show what the agent was "looking at." It genuinely renders the reasoning — the AI's perceived threat map becomes readable. Research demonstration
Interpretable-by-construction — don't build a black box at all. The IXDSS work encodes COLREGS directly in Prolog: the system reasons over the rules symbolically, so every recommendation carries its own deduction path — which rule, which classification, which conclusion — plus interactive what-if ("what if the target turns to port?"). No interpretation needed because there's nothing opaque to interpret. Research demonstration
Contrastive explanation — explain by comparison. The officer doesn't need the model's internals; they need "why turn to starboard and not slow down." The 2026 study tested exactly this with serving marine officers: contrastive explanations clarified the system's objectives — but added cognitive load, meaning the display itself becomes a burden under pressure. Research demonstration
The trade the papers don't hide
Read the methods honestly and a hard trade emerges: the deepest networks are the least explainable, the cleanest explanations come from simpler systems. A Prolog reasoner explains itself perfectly and handles a narrow slice of encounters; a deep RL policy handles more of the sea and explains itself only approximately. Maritime-GPT analysis
Post-hoc explanation papers over this — attention maps and saliency areplausible readings of a network, not guaranteed-accurate ones; a confident wrong explanation can be worse than none, because it manufactures trust. The honest framing the field is arriving at: explanation is a safety feature with its own failure modes, not free transparency. Maritime-GPT analysis
What we know / don't / being tested
What we know
- Decision processes in DRL avoiders can be made human-legible.
- Rule-encoded systems give correct explanations by construction.
- Explanation load is a real cost — officers work harder with more of it.
What we don't know
- Whether post-hoc explanations are faithful or just persuasive.
- The right explanation for each role — master vs remote operator vs investigator.
- Whether explanation aids or erodes calibrated trust over long exposure.
Being tested
- Sub-task critics + attention for intent visualisation.
- Contrastive "why this and not that" interfaces for supervisors.
- Demand-driven disclosure — explain on request, stay quiet otherwise.
Maritime-GPT analysis
The LLM may be the missing explanation layer. The unsolved part of maritime XAI isn't producing a legible internal state — it's translating it into seamen's language. A system that can say "stand-on vessel failed to keep clear; Rule 17(c) action taken" is worth more than a marginally better path that can't. That is precisely what language models do, and it's the strongest case for an LLM aboard that doesn't involve trusting it with the helm — the model explains a decision it did not make. Maritime-GPT analysis
Open research problems
- Faithfulness metrics — verifying an explanation reflects the real decision process, not a plausible story.
- Role-adaptive disclosure — different explanation depth for the OOW, the ROC and the accident investigator.
- Explanation under time pressure — what a supervisor can actually absorb in a developing encounter.
- Calibrated trust — explanations that support appropriate reliance, not blanket over-trust.
Evidence scorecard
| Dimension | Rating | Basis |
|---|---|---|
| Evidence strength | Moderate | Working methods plus early human-factors data; thin at scale. |
| Research maturity | Experimental | Demonstrated techniques; no fielded explainable avoider. |
| Operational significance | High | Explainability is a prerequisite for supervised autonomy under MASS. |
| Confidence in conclusions | High | Methods and limits drawn from named studies. |
Related tools & reading
Sources
- Explainable AI for ship collision avoidance: decoding decision-making processes and behavioral intentions — Applied Ocean Research 156 (2025), art. 104471Peer-reviewed — also arXiv 2405.09081
Sub-task critic networks + Q-value analysis + attention to render a DRL avoider's reasoning and perceived threat readable to humans. - Logic programming for an interactive explainable naval decision support system (IXDSS) — FAIA / IOS Press (2026)Peer-reviewed
COLREGS encoded in Prolog — inherently traceable decision paths plus interactive what-if analysis; explanation comes free with the architecture. - MARCOL: a maritime collision-avoidance decision-making testbed — AAAI — MARCOLPeer-reviewed
An explainable multi-objective framework explicitly built around COLREGS ambiguity ('ordinary practice of seamen') and repeatable testing. - "Why this avoidance maneuver?" Contrastive explanations in human-supervised maritime autonomous navigation — arXiv 2604.08032 (2026)Preprint — user study
Contrastive explanations tested with serving marine officers: they aid understanding of system objectives but increase cognitive load — demand-driven explanation wins.
Methods and user-study findings from the cited works; accessed 5 October 2026. Editorial analysis only.