Two AI traditions are competing for the autonomous ship's helm, and they don't just differ technically — they differ in kind. The reinforcement-learning policy learned collision avoidance by doing it millions of times in simulation; the LLM reads COLREGS and reasons about the situation in language. One is reflex, the other is deliberation — and each one's failure mode is exactly the other's strength.
The LLM-vs-RL question isn't really 'which is better' — it's 'which does what', because the answer is almost certainly a division of labour, not a winner.
Key findings
- RL learns the reflex — collision-avoidance action from state — trained by millions of simulated encounters. Established
- LLMs bring reasoning — COLREGS interpretation, situation description, explainable rationale — but at language-speed, not reflex-speed. Established
- RL's weakness is interpretability and generalization; LLM's weakness is latency and grounding in continuous dynamics. Established
- The hybrid architecture is the emerging answer — LLM as tactical reasoner/COLREGS-checker, RL (or classical) as the reactive controller. Research demonstration
- Neither is certifiably safe alone — RL fails on out-of-distribution encounters, LLMs on confident misreading — so the supervision layer stays. Maritime-GPT analysis
What each one actually computes
The RL policy is a function: ship state in, rudder/throttle action out — trained on simulated encounters until the avoidance behaviour is learned. It's fast, reactive, and opaque — it knows what to do but can't say why. The LLM is a reasoner: it reads the COLREGS rule, interprets the encounter in language, and produces a decision with a rationale attached. It's slow, deliberative, and fluent — it can explain itself but can't match the reflex's millisecond reaction. Established
Where each fails
- <strong>RL fails on the unusual.</strong> An encounter outside its training distribution produces behaviour it never learned — the out-of-distribution problem. Established
- <strong>LLM fails on the dynamics.</strong> Ship manoeuvring is continuous physics; a language model's fluent reasoning may miss the geometry the situation actually requires. Established
- <strong>RL can't explain itself.</strong> The certification problem — a policy that can't say why it turned is hard to trust in a safety case. Established
- <strong>LLM can't react.</strong> In a close-quarters situation, deliberation time is the luxury that isn't available. Established
The division of labour
The emerging architecture is layered: the RL (or classical) layer handles the continuous dynamics — the immediate reflex that keeps the ship safe — while the LLM handles the tactical reasoning — interpreting the COLREGS situation, checking the manoeuvre's legality, explaining the decision to the human oversight. Navigation-GPT and related frameworks point at this structure: the LLM as the 'officer' that thinks in rules, the RL as the 'helmsman' that acts in degrees. Research demonstration
The honest safety verdict
Neither approach is certifiably safe alone — the RL fails on the encounter it never saw, the LLM fails on the fluent misreading of a rule. The MASS-relevant architecture treats both as components under a supervisory layer: the autonomous function bounded by its operating envelope, the human (remote or aboard) retaining authority when the envelope is exceeded. The LLM-vs-RL debate is real, but the safety answer is supervisory — the autonomy works inside its certified envelope, and the human takes the cases it can't. Established
What we know
- RL provides the reflex; the LLM provides the reasoning — complementary capabilities.
- The hybrid layered architecture is the emerging consensus.
- Neither is certifiably safe alone; the supervisory structure is the answer.
What we don't
- Whether an LLM's COLREGS reasoning is reliable enough to certify.
- How the RL/LLM division should be engineered in a real architecture.
- The latency budget that allows LLM reasoning inside the decision loop.
Being tested
- Hybrid LLM-reasoner/RL-controller prototypes.
- LLM-as-COLREGS-checker validation against exam scenarios.
- Supervisory-envelope architectures for bounded autonomy.
Maritime-GPT analysis
The LLM-vs-RL question is a category error wearing a technical question — they don't compete, they layer. The reflex and the reasoner fail in opposite ways: the RL can't explain, the LLM can't react, and neither can be trusted alone. The honest architecture uses each for what it's good at, under a supervisor that knows when neither is. The autonomous ship doesn't need a winner; it needs a division of labour. Maritime-GPT analysis
This is a research feature, not operational guidance. Simulation results are reported as simulation; demonstrations are not deployments; company claims are attributed as such. Verify safety-critical decisions against official nautical publications and flag-state requirements.
Sources & evidence
The architecture survey covering both RL and LLM navigation approaches.
Recent RL evidence the comparison draws from.