The collision-avoidance literature is full of impressive policies — and most of them were trained on ships that don't exist. A point-mass model that turns on command, accelerates instantly, and obeys clean kinematics produces a policy that performs beautifully in simulation and dangerously at sea: a real 200m bulk carrier has inertia, sway, a seconds-long yaw lag, and a response that depends on loading, speed, and sea state.
The dynamics gap — between the simulated ship and the real one — is where collision-avoidance research meets its engineering limit, and why 'works in sim' and 'works at sea' are different claims.
Key findings
- Most collision-avoidance research uses kinematic point models — ships without inertia, sway, or lag. Established
- A real vessel's response is lagged and load-dependent — the rudder doesn't turn the ship; it starts a process that takes seconds. Established
- The sim-to-real gap is the transfer problem — a policy learned on simple dynamics can fail on the real ship's response. Established
- Higher-fidelity dynamics in training measurably improve transfer — the twin-based and 3-DOF/6-DOF models close the gap. Research demonstration
- The safety margin must absorb dynamics uncertainty — the collision-avoidance envelope must include the ship's response lag. Maritime-GPT analysis
The point-mass fiction
The kinematic model — position, velocity, heading, nothing else — treats the ship as a particle that turns instantly. A real ship is a dynamical system: the rudder creates a yaw moment, the hull responds with lag, the ship carries sway through the turn, and the whole response depends on how it's loaded and how the sea is running. A policy trained on the fiction has never felt the delay between 'order rudder' and 'ship turns' — which is exactly the delay that matters in close quarters. Established
What the real dynamics add
- <strong>Inertia.</strong> A loaded vessel carries momentum that resists the manoeuvre the policy ordered. Established
- <strong>Yaw lag.</strong> The seconds between helm order and heading change — the lag that makes close-quarters timing critical. Established
- <strong>Sway and drift.</strong> The ship doesn't move along its heading — it slides, and the model must know it. Established
- <strong>Load and sea-state dependence.</strong> The same rudder order produces different turns at different drafts and sea states. Established
The transfer problem
The sim-to-real gap is the name for the failure: a policy trained on simple dynamics deploys on a real ship whose response it never learned. The mitigations are the twin-literature's answers — train on higher-fidelity models (3-DOF, 6-DOF manoeuvring models), randomize the dynamics across training so the policy is robust to its uncertainty, and validate on progressively more realistic twins before the sea trial. The policy that survives the high-fidelity twin is the one that earns the water. Research demonstration
The safety-margin implication
Real dynamics don't just affect performance — they define the safety envelope. A collision-avoidance decision must account for the ship's response time: 'if I order the turn now, where will the ship be in 30 seconds' is a dynamics question, not a geometry one. The policy that treats the ship as a point produces margin calculations that are wrong by the response lag — which in a close-quarters encounter is the difference between passing clear and not. The honest system carries the ship's dynamics inside its margin. Maritime-GPT analysis
What we know
- The kinematic point model omits the dynamics that matter in close quarters.
- Real ship response is lagged, load-dependent, and sea-state-dependent.
- Higher-fidelity training and dynamics randomization close the transfer gap.
What we don't
- How much fidelity is 'enough' for a policy to transfer safely.
- Whether 3-DOF trained policies hold on real-ship trials.
- The certification standard for dynamics fidelity in the safety case.
Being tested
- High-fidelity dynamics in RL training environments.
- Dynamics-randomized policies for transfer robustness.
- Real-ship validation of simulation-trained collision avoidance.
Maritime-GPT analysis
The dynamics gap is collision-avoidance research's least-glamorous, most-important problem — the difference between a paper's collision rate and a ship's safety record is often just the fidelity of the model it trained on. A point-mass policy is a good first draft; a dynamics-aware one is the deployable version. The research that's honest about the gap is the research worth building on; the kind that reports sim numbers on point masses is marketing wearing a DOI. 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
Recent work moving toward realistic vessel dynamics.
The manoeuvring-model fidelity that dynamics-aware training needs.
The training this dynamics discussion constrains.