AI can forecast a berth conflict, rank a machinery alert or propose a fuel-efficient route. It cannot make a narrow channel wider, turn uncertain sensor data into a safe lookout or transfer the master’s and operator’s responsibilities to a software supplier.
The practical pattern is observe → predict → constrain → authorise → act → verify. The model should have a declared operating domain, a defined human or engineered authority above it and a safe result when evidence is missing.
The sources and regulatory position below were checked on 31 July 2026. Maritime duties depend on vessel type and size, flag, operating area, cargo, activity, port and the coastal state. UK rules may apply to a UK ship or an operation in UK waters, while international conventions and local harbour directions may apply simultaneously. Engage the Maritime and Coastguard Agency, classification society, harbour authority and other competent bodies for the actual concept.
Separate Five Different Maritime Decisions
“Autonomous shipping” obscures systems with very different consequences. Write a decision record before choosing a model.
| Use case | AI may support | Authority that remains outside the model | Primary evidence |
|---|---|---|---|
| Port planning | Arrival, berth, yard or equipment forecast | Harbour limits, safe capacity, pilotage, traffic and emergency controls | Port call, tide, weather, berth, labour and equipment records |
| Voyage advice | Route, speed or arrival-window options | Master’s safe-navigation decision and current passage plan | Approved charts, notices, weather, traffic, draught and vessel limits |
| Machinery maintenance | Fault detection and work prioritisation | Competent inspection, isolation, repair and return to service | Calibrated condition data, maintenance and failure history |
| Remote or autonomous operation | Perception, planning or bounded control | Certified safety architecture, remote operator and emergency authority | Sensor coverage, operating domain, trials, hazards and assurance case |
| Emissions support | Fuel, speed and voyage comparison | Applicable monitoring plan, verified report and compliance decision | Calibrated fuel/activity data and regulated calculation method |
Record purpose, vessel or port boundary, forecast horizon, data issue time, model version, permitted action, uncertainty, approver and rollback route. A berth optimiser must not silently become a vessel-traffic instruction; an anomaly score must not close a maintenance defect.
Put Port Optimisation Inside the Marine Safety System
A port model may estimate arrival times, allocate cranes or suggest yard moves. Its objective cannot be “maximum throughput” without safety, environmental and labour constraints.
The April 2025 Ports and Marine Facilities Safety Code sets a national standard for UK ports, harbours and marine facilities. It expects proportionate risk assessment, defined dutyholders and a marine safety management system. The MCA’s January 2026 compliance exercise confirms the code’s current operational importance while noting that it is not itself a statutory requirement. Applicable legislation and local powers still need separate mapping.
Add the AI change to the port’s formal hazard process. Consider:
- conflicting vessel movements and wake effects;
- under-keel clearance, berth geometry, tide and weather;
- towage, pilotage and mooring resources;
- cranes, vehicles and people sharing work areas;
- hazardous cargo and emergency access;
- degraded communications or position feeds;
- congestion moved from berth to gate, road or anchorage; and
- incentives that pressure staff to accept an unsafe sequence.
Safe capacity is an engineering and operational determination, not yesterday’s maximum count. Preserve harbourmaster and vessel-traffic-service authority, and show operators why a recommendation changed. If a feed is stale or a constraint is unresolved, return “no plan” rather than fill the slot.
Measure berth-plan stability, estimated-time-of-arrival error, conflict rejections, unplanned rehandles, operator overrides and safety events. Compare the complete port call, not only crane moves per hour. Our warehouse [automation guide](/blog/logistics-ai-warehouse-automation-supply-chain-uk) covers the landside handoff in more detail.
Treat Route Optimisation as Advice With Uncertainty
Fuel-minimising routes can conflict with navigational safety, schedule, weather avoidance, cargo limits, emissions rules or commercial instructions. Optimise only after hard constraints are applied.
The voyage evidence pack should include:
- approved and corrected navigation data;
- vessel dimensions, draught, stability and manoeuvring limits;
- machinery availability and fuel characteristics;
- weather and ocean forecast issue times and uncertainty;
- traffic, routeing, restricted and environmentally sensitive areas;
- port arrival window and contingency;
- the current passage plan; and
- every recommendation accepted, changed or rejected.
Back-test by voyage and season so positions from the same passage do not leak between training and evaluation. Replay heavy weather, forecast shifts, traffic density, sensor disagreement and missed port windows. Report fuel and time against a credible baseline, plus route deviation, unsafe-suggestion rejection and forecast error.
Do not advertise a percentage fuel saving from simulation alone. A fair counterfactual uses the same vessel, cargo, weather, port window and operational restrictions. Report extra distance, schedule effects and any emissions displaced to tugs, waiting or terminal activity.
The Environment Agency’s June 2026 UK ETS maritime compliance guidance defines regulated scope, monitoring plans, reporting, verification and surrender duties for operators it regulates. Scope includes thresholds and exclusions, and other UK regulators cover their jurisdictions. An AI estimate can support data checks; it does not replace the approved monitoring method or independent verification.
For the road and fleet side of route decisions, see our UK transport optimisation guide.
Define the Autonomous Vessel’s Operating Domain
“Uncrewed”, “remotely operated” and “autonomous” are not interchangeable. Define:
- vessel, length, construction and certification basis;
- waters, charted route, distance from refuge and traffic density;
- wind, sea state, visibility, current and daylight limits;
- communications coverage and maximum acceptable latency;
- sensor set, range, blind sectors, environmental limits and redundancy;
- remote-control centre, watch arrangement and operator workload;
- functions performed onboard, ashore and manually;
- minimum-risk condition and recovery; and
- search-and-rescue, pollution and casualty response.
The MCA’s June 2026 Maritime Autonomous Surface Ships Innovation Hub guidance explains that the applicable route depends on vessel length, operating area and use. It directs developers to the relevant codes and evidence for alternative design, watchkeeping equivalence, remote operations and whole-ship integration. Its Plymouth prototype areas are controlled trial arrangements, not permission for unrestricted deployment.
The IMO’s non-mandatory MASS Code was announced in the UK on 1 July 2026. It covers matters including certification, software, security, manning, training, watchkeeping, navigation and remote operations. At this guide’s cutoff it is an international non-mandatory framework entering an experience-building phase; the announced mandatory version is scheduled for 2032. Do not present compliance with a pilot or voluntary code as a current licence to trade autonomously.
Build an assurance case from hazards to controls and test evidence. Simulation is useful for breadth, but trials must show real sensor effects, vessel dynamics, communication loss and human response. Every scenario needs expected behaviour and a pass criterion.
Make Perception Fail Clearly
Marine perception must distinguish a vessel, buoy, person, land, weather return or sensor artefact soon enough to support the next safe action. Average object-detection accuracy is inadequate.
Test by range, relative bearing, closing speed, target size, sea state, glare, darkness, rain, fog, clutter and sensor availability. Include fishing gear, small craft, non-cooperative objects and partially occluded targets where relevant to the domain.
Report:
- probability of detection by target class and range;
- false and late alert rates per operating hour;
- track continuity and position error;
- time-to-action margin;
- sensor disagreement and degradation rate;
- minimum-risk-state success; and
- remote-operator detection, acknowledgement and intervention time.
Keep collision regulations and the approved navigation function above a learned predictor. A model classification is one input to situational awareness; it is not a declaration of right of way. Uncertainty must drive slower speed, wider margins, operator escalation or domain exit—not confident extrapolation.
Design the Remote Operator as Part of the Vessel
Moving the bridge ashore changes rather than removes the human problem. Test the complete team, interfaces, workload, communications and handover.
Show the operator source timestamps, sensor health, vessel mode, control authority, limits and the reason for an alert. Avoid alarm floods that make the remote watch nominal. Specify the maximum number and mix of vessels an operator may supervise, then validate it under simultaneous faults rather than normal cruising.
Use drills for lost communications, degraded positioning, conflicting commands, propulsion fault, fire, flooding and recovery of control. Log every authority transfer. The minimum-risk condition must work without an optimistic cloud dependency.
The MCA’s Code of Safe Working Practices for Merchant Seafarers remains relevant to UK-flagged merchant-ship work and onboard safety. Removing crew from one task does not remove duties to people who maintain, board, recover or support the vessel.
Secure Navigation and Port Systems
Route, engine, crane and access systems are cyber-physical. Separate safety-critical networks from business and guest access; control vendor connections; use strong identities and least privilege; sign and validate software and model updates; monitor unusual commands; and keep tested offline recovery.
The Department for Transport’s cyber security code of practice for ships, updated in 2023, covers assessment, resilience, incident handling and recovery in alignment with IMO requirements. Apply it to the whole autonomous chain: sensors, vessel networks, satellite links, remote centre, cloud services and suppliers.
Test spoofed position, delayed weather, corrupted model files, replayed commands, credential loss and unavailable vendor support. The response must preserve safe navigation and evidence, not merely restore the dashboard. Our AI cybersecurity guide provides a broader detection-and-response workflow.
Set Measurable Release Gates
Release one vessel, port function and operating mode at a time:
| Gate | Minimum release evidence |
|---|---|
| Regulatory route | Flag, waters, vessel code, port directions, class and approval owners mapped; required engagement complete |
| Operating domain | Geographic, weather, traffic, sensor, communications and staffing limits machine-checkable and documented |
| Port safety | Model change assessed in the marine safety system; zero unresolved movement conflicts in scenario testing |
| Navigation | All proposed routes pass independent chart, vessel, weather, traffic and passage-plan constraints |
| Perception | Pre-agreed detection, false-alert, tracking and time-to-action thresholds pass every critical domain slice |
| Human control | Remote workload, intervention, handover and emergency drills pass under simultaneous credible failures |
| Minimum-risk state | Communications, positioning, sensor and model failures reach the defined safe condition in every test |
| Cyber resilience | Network separation, privileged access, update integrity, logging, response and offline recovery demonstrated |
| Emissions | Claimed benefit reproduced against a matched voyage baseline; regulated data reconciles to the monitoring plan |
| Operations | Named watch, maintenance, incident, rollback and regulator-notification processes active before live influence |
Monitor domain exits, route rejections, perception misses, late alerts, loss-of-link events, remote interventions, operator workload, port-plan overrides, fuel and schedule outcomes, security events and near misses. Pause the affected function when a critical threshold is breached or a model, sensor, vessel or rule changes outside its validated configuration.
Maritime AI earns trust by making uncertainty and authority visible. The goal is not a ship that appears independent or a port that always looks full. It is an operation that stays inside its evidence, protects the watch and reaches a known safe state when automation cannot continue.
