MASS, the COLREGS, and Automation Transparency

emerging Updated: 2026-06-07
mass autonomous-systems colregs collision-avoidance operational-design-domain degrees-of-autonomy automation-transparency anthropomorphism machine-learning governance maritime

MASS, the COLREGS, and Automation Transparency

Status: emerging
Last updated: 2026-06-07
Sources: Porathe2019_Mass_And_The_Colregs.Pdf
Tags: [mass, autonomous-systems, colregs, collision-avoidance, operational-design-domain, degrees-of-autonomy, automation-transparency, anthropomorphism, machine-learning, governance, maritime]

Summary

Porathe (2019) asks whether the international collision regulations (COLREGS) must be rewritten in quantified terms before autonomous ships can navigate by them. His answer is no: like other legal text, the rules should stay general and open to interpretation, while it is the collision-avoidance algorithms — not the law — that must be made precise and quantitative, for example by learning the local meaning of "early and substantial action" from AIS traffic data within a defined Operational Design Domain. Because autonomous ships will share the sea with human-crewed ships for a long time, the paper's central design principle is automation transparency: a Maritime Autonomous Surface Ship (MASS) must behave predictably and make its mode, situation awareness, and intentions understandable to nearby humans. This is a governance-and-design argument about deploying AI safely in a regulated, safety-critical domain.

Body

Context

Porathe (2019) is a discussion paper occasioned by Norway's YARA Birkeland, an unmanned autonomous container feeder then scheduled to sail autonomously by 2022 through complex inshore waters. The lens is the gap between machine decision-making and a body of qualitative law: the author works clause by clause through the COLREGS to show where a collision-avoidance programmer meets undefined terms (PDF pp. 2–3, orig. pp. 512–513). Within this knowledge base the article seeds the autonomous-systems and governance strands — it is the first MASS source here — and pairs naturally with Ironies Of Automation: where Bainbridge warns that automation undermines the human left to supervise it, Porathe addresses the inverse coordination problem of autonomous agents that must remain legible to the humans around them. It borrows the self-driving-car concepts of Operational Design Domain and graded degrees of autonomy, and connects to the maritime collision-avoidance and remote-operations work held in sibling knowledge bases.

Key Points

Autonomy is a continuum, anchored to context. Porathe distinguishes unmanned, automatic, and autonomous operation and notes that today's "manual" ships already run at high automation (track-following autopilots). A ship may be periodically unattended, remotely monitored, or remotely controlled from a Shore Control Centre, evolving gradually toward higher autonomy (PDF pp. 1–2, orig. pp. 511–512). The Operational Design Domain (ODD), borrowed from self-driving cars, defines the prepared lanes and conditions within which a ship may run autonomously, navigating manually or under remote control elsewhere (PDF p. 2, orig. p. 512).

The COLREGS are deliberately qualitative, and that is the programmer's problem. Like legal text generally, the rules are written generally so as to apply across situations, with precise interpretation left to "the ordinary practice of seamen" (Rule 2). Rule 2 simultaneously requires obeying the rules and departing from them when necessary to avoid collision, without specifying when (PDF pp. 2–3, orig. pp. 512–513). Rules 15–17 (crossing situations) hinge on "early and substantial action" — undefined in miles, degrees, or minutes — and Rule 19 (restricted visibility) adds soft terms such as "safe speed" and "ample time," plus the problem that "restricted visibility" maps onto sensor visibility (cameras, radar, LIDAR) rather than the human eye (PDF pp. 3–5, orig. pp. 513–515). Local custom complicates matters further: in some areas fast ferries keep clear of everything, contrary to a literal reading of the rules (PDF p. 4, orig. p. 514).

Porathe's resolution separates the law from the algorithm. Quantifying the COLREGS once and for all would produce a regulatory text that was both very long and still incomplete — the "unknown unknowns" would keep appearing — so the rules should stay general while the collision-avoidance application is made precise and quantitative (PDF p. 5, orig. p. 515). A programmer could deduce local thresholds for "early" and "substantial" by mining AIS data for an ODD, using the idea of a ship "safety zone"; the Nautical Institute's guideline of a 2-mile CPA in open sea and 1 mile in restricted waters is one such concrete value (PDF p. 4, orig. p. 514). He suggests "lifelong" machine learning would let the onboard AI grow more experienced, while noting the IMO is unlikely to certify an AI that does not behave in a predetermined, predictable way (PDF p. 5, orig. p. 515).

Because manned and autonomous ships will coexist, transparency is the safety mechanism. The AI may extrapolate far further ahead than a human and so make manoeuvres that surprise a nearby officer ("automation surprise"); to avoid this, MASS should behave in a humanlike, predictable manner and signal its state (PDF p. 5, orig. p. 515). Porathe proposes three transparency measures: countering misplaced anthropomorphism by guaranteeing the ship always follows COLREGS so its behaviour can be learned; an identification signal that the ship is in autonomous mode (an "A" on the AIS icon, a purple masthead light at night — the same colour debated in the autonomous-car industry); and sharing intentions through route exchange, so other ships, VTS, and coastguards can see what the MASS has observed and where it intends to go (PDF pp. 5–7, orig. pp. 515–517).

Conclusion

Porathe (2019) concludes that the challenge of autonomous collision avoidance lies in the qualitative nature of the COLREGS set against the quantitative needs of real situations, and that the answer is not to quantify the law but to make the algorithms precise and the autonomous ship transparent. The onboard AI's potential to be "smarter" than humans — extrapolating further, keeping more options open — is itself a hazard if it produces incomprehensible manoeuvres; the software should therefore aim to behave in a humanlike way and to advertise its mode, situation awareness, and intentions. Only if other mariners can understand how a MASS works, Porathe argues, is peaceful coexistence at sea possible. The position is a clear governance stance: certify predictable behaviour, design for legibility, and keep autonomy bounded to a defined operational domain.

  • Ironies Of Automation — Bainbridge's complementary critique of automation; both treat the human–automation boundary as the central safety problem

References

International Maritime Organization (1972) Convention on the International Regulations for Preventing Collisions at Sea, 1972 (COLREGS). London: IMO. To be validated.

Lee, G.W.U. and Parker, J. (2007) Managing Collision Avoidance at Sea. London: Nautical Institute. To be validated.

Porathe, T. (2019) 'Maritime Autonomous Surface Ships (MASS) and the COLREGS: Do We Need Quantified Rules Or Is "the Ordinary Practice of Seamen" Specific Enough?', TransNav, the International Journal on Marine Navigation and Safety of Sea Transportation, 13(3), pp. 511–517. doi: 10.12716/1001.13.03.04. porathe2019mass

Rødseth, Ø.J. and Nordahl, H. (2017) Definitions for Autonomous Merchant Ships. Version 1.0. Trondheim: Norwegian Forum for Autonomous Ships. To be validated.

Open Questions

  • Porathe favours learning "early and substantial" thresholds from AIS data per ODD, yet also notes the IMO is unlikely to certify an AI that is not predetermined and predictable. How a lifelong-learning collision-avoidance system could be certified is unresolved and central to deploying autonomous navigation.
  • The transparency proposals (identification light, route exchange) presuppose that human operators will correctly read and trust autonomous-mode signals. Whether such signalling actually improves coordination, or merely shifts the interpretation burden, is untested here.
  • The argument that algorithms should be quantitative while the law stays qualitative leaves open who is liable when a certified algorithm's interpretation of a soft rule causes a collision — a governance gap the paper raises (negligence / "good seamanship") but does not resolve.
  • Cross-KB: the collision-avoidance and shore-control-centre machinery this paper assumes is treated in detail in the remote-operations corpus (shared-control collision avoidance, ROC concepts, automation transparency). Linking those would connect the AI/governance view here to the human-factors-of-supervision view there.