Most discussion of autonomous collision avoidance is really discussion of path planning around obstacles. That is a solved-enough problem. The unsolved problem is that the International Regulations for Preventing Collisions at Sea are not written about obstacles. They are written about vessels, their types, their obligations to one another, and the judgement of a person keeping a lookout.
Rule 5 is a sensing requirement disguised as a procedural one
Rule 5 requires every vessel to maintain a proper lookout by sight and hearing as well as by all available means appropriate in the prevailing circumstances, so as to make a full appraisal of the situation and of the risk of collision. Read carefully, that is a specification.
It requires sight — not merely detection. It requires a full appraisal of the situation, which means determining not just that something is there but what it is and what it is doing. And it requires all available means, which is why a system that relies on radar alone and ignores its cameras is arguably non-compliant in exactly the conditions where cameras contribute most.
Rule 7 raises the bar again
Rule 7 requires the use of all available means appropriate to determine if risk of collision exists, and states that if there is any doubt, such risk shall be deemed to exist. That last clause is doing enormous work for an autonomous system designer.
A machine's natural failure mode under uncertainty is to act on its best guess. Rule 7 says the opposite: under doubt, assume risk. Encoding that faithfully means the perception layer must produce calibrated uncertainty rather than a confident label, and the decision layer must treat low confidence as a positive indication of risk rather than as an absence of evidence. Systems that quietly threshold away low-confidence detections have inverted the rule.
Vessel type changes the obligation, and only vision establishes it
This is the point where radar-only autonomy runs out of road. Consider two contacts, identical on radar: the same size, the same range, the same course and speed.
- One is a power-driven vessel underway. In a crossing situation, ordinary give-way and stand-on obligations apply under Rule 15.
- The other is a trawler with her gear out — a vessel engaged in fishing under Rule 18, which a power-driven vessel must keep out of the way of.
The obligations are different, and in some geometries they are opposite. Nothing in the radar return distinguishes them. What distinguishes them is the day shape she is carrying, or at night the light pattern she is showing. That is a vision problem, and there is no way around it.
The same holds for a sailing vessel under Rule 12, a vessel restricted in her ability to manoeuvre, a vessel not under command, a pilot vessel, and a dredger with an obstructed side. Each is identified by lights and shapes. Each changes what you are legally required to do.
An autonomous vessel that cannot read navigation lights is not keeping a lookout in any sense the regulations recognise. It is avoiding radar returns and hoping the categories line up.
Rule 8 wants action a human would notice
Rule 8 requires that action to avoid collision be positive, made in ample time, and — where circumstances allow — large enough to be readily apparent to another vessel observing visually or by radar. It explicitly cautions against a succession of small alterations.
This is directly at odds with how an unconstrained optimiser behaves. Minimising deviation from track subject to a clearance constraint produces exactly the succession of small alterations Rule 8 warns against. Compliance requires the planner to be penalised for small manoeuvres, not rewarded for them — and to commit early enough that the other vessel's bridge team can read the intention.
Rule 19 changes the whole framework in restricted visibility
In restricted visibility the stand-on concept largely disappears; Rule 19 imposes different obligations on a vessel that detects another by radar alone. An autonomous system must therefore know which visibility regime it is in and switch rule sets accordingly — and its own perception degradation in fog is itself evidence about the regime. There is a neat recursion here that is easy to get wrong: the sensor that is failing is also the sensor telling you the rules have changed.
Why the log matters more than the manoeuvre
When an autonomous system is put in front of a trials authority, a flag administration or an accident investigator, the question is never simply whether it avoided the collision. It is whether the action it took was the action the rules required, and how that can be demonstrated.
That is an argument for explicit rule encoding over learned end-to-end policy, at least for the decision layer. A learned policy that empirically avoids collisions cannot tell you which rule it was applying, and cannot be audited against a regulation written in natural language for human judgement.
Our own position is a split one: learned perception, explicit reasoning. The models do what models are good at — detecting, classifying, reading lights and shapes, tracking through occlusion. The rule engine does what symbolic systems are good at — applying a written body of rules consistently and citing which one it applied.
Every manoeuvre VARUNA-AV proposes is logged with its governing rule, the give-way determination and the closest point of approach it is protecting. This is less elegant than a single end-to-end network. It is considerably easier to defend at a trials board, which is the only elegance that matters here.
The honest caveat
None of this makes an autonomous lookout equivalent to a competent officer of the watch. COLREGs are built around seamanship and the ordinary practice of seamen, a standard that explicitly contemplates departing from the rules when necessary to avoid immediate danger under Rule 2. That judgement is not something we can claim to have encoded, and we would be sceptical of anyone who did.
What a machine offers is consistency, endurance and an audit trail. What it does not offer is seamanship. Systems built on the assumption that those are the same thing will be found out, and the finding out will happen at sea.