Norway's salmon farms fight sea lice with submerged robots that watch fish continuously, recognise a louse on a moving salmon using machine vision, and respond in the water without handling the fish. The best-known approach is the laser node: a sealed unit suspended inside a sea pen that identifies a parasite on a passing salmon and fires a short optical pulse at it, while other submersible units focus on cleaning nets, counting lice, or keeping fish below the lice-rich surface layer.
Sea lice (*Lepeophtheirus salmonis*) are small marine copepods that attach to farmed and wild salmon and feed on skin and mucus. They are the single most expensive biological problem in Norwegian salmon farming, and the traditional treatments — chemical baths, warm-water flushes, mechanical brushing — all require crowding and pumping the fish, which stresses and sometimes kills them. In-pen robotics exists to replace the treatment event with continuous, low-contact management.
Table of Contents
- What the submerged units actually do
- Why vision is the hard part
- The other robots in the pen
- What the evidence supports, and where the limits are
- How to evaluate a system if you are buying or specifying one
- Who this affects beyond the farms
- Frequently Asked Questions
What the submerged units actually do
A laser delousing node is a pressure-housed assembly of cameras, lights, image-processing hardware and a low-power laser, hung on a line inside a net pen and powered from the feed barge or a surface float. Cameras run constantly. When a salmon swims through the field of view, the vision model has to do three things in a fraction of a second: separate fish from background, locate and classify a louse against the fish's own skin pattern, and confirm the target is a parasite rather than a wound, a scale, or a shadow. If the system is confident, it aims and fires.
The pulse is short and the energy is tuned so that the louse's pigmented body absorbs it while the salmon's reflective, silvered skin does not. The salmon is unrestrained and typically shows no reaction; the louse is killed or detached and falls away. The engineering difficulty is not the laser. It is doing reliable target recognition in moving, turbid, poorly-lit water, on a fish that is itself moving, with a shot window measured in milliseconds and a false-positive cost measured in fish welfare.
Why vision is the hard part
Lice are a few millimetres across and sit mostly on the head, dorsal area and behind the fins. A model trained on clear-water imagery degrades quickly in a Norwegian winter pen, where algae, plankton blooms, suspended feed particles and low ambient light reduce contrast. Biofouling on the housing's optical window adds another layer of degradation that worsens over weeks. Systems address this with their own lighting, with image enhancement before classification, and with a confidence threshold that makes the unit withhold a shot rather than take an uncertain one.
A missed louse costs nothing. A misclassified one aimed at an eye or a gill does. This is also why vendors distinguish between adult female lice, which carry visible egg strings, and smaller mobile or sessile stages. Adult females are the regulatory and biological priority because they reproduce, and they are the easiest stage to recognise. A system that reliably removes gravid females suppresses the pen's reproductive output even if it never sees the smallest stages.
The other robots in the pen
Laser nodes are the headline technology, but they share the water with several other machines that attack the same problem differently: That last category matters for anyone evaluating the field. Prevention by depth control and physical barriers often does more for lice numbers than removal does, and the two are complementary rather than competing.
- **Net-cleaning ROVs** — remotely operated or semi-autonomous crawlers and thrusters that remove biofouling from the netting. Cleaner nets improve water exchange and reduce the surfaces where lice larvae and organic matter accumulate.
- **Inspection and mortality-collection units** — submersibles that patrol the pen, check the net for holes, and retrieve dead fish from the bottom cone, reducing the need for divers.
- **Lice-counting cameras** — fixed or winched camera rigs that estimate lice load automatically, replacing or supplementing the manual counts that require netting and anaesthetising sample fish.
- **Submerged feeding and lighting systems** — these do not touch lice at all. They exploit the fact that infective lice larvae concentrate in the top few metres, and use feed depth, lights and "snorkel" barriers to keep salmon below that layer.
What the evidence supports, and where the limits are
In-pen robotics is best understood as pressure management, not eradication. A laser node treats the fish that happen to swim past it, so coverage depends on how many units a pen has, how fish distribute themselves in the water column, and how much of the time the system is operating with clear optics. Pens are large, stocked densely, and fish move in shoals — so the realistic claim is a sustained reduction in lice load that delays or reduces the need for a full mechanical or thermal treatment, rather than a pen that reaches zero. The welfare argument is the strongest one.
Crowding, pumping and thermal treatment carry well-documented mortality and injury costs in Atlantic salmon. Anything that reduces the number of handling events in a production cycle has a measurable welfare benefit independent of how many lice it removes. The limits worth stating plainly: capital cost per pen is high; the units need power, servicing and optical cleaning; performance in poor visibility is worse than in demonstration footage; and independent, peer-reviewed field comparisons remain thinner than vendor material. Norway's lice regime is also regulatory, not just technical — farms are held to counting and reporting obligations and to lice limits that drive treatment decisions, so a technology's value is partly a question of whether it keeps a site under its threshold.
How to evaluate a system if you are buying or specifying one
Treat vendor detection accuracy figures the way you would treat any computer-vision benchmark: ask what the test distribution was. A classifier validated on clear-water tank imagery tells you little about a December pen in a fjord.
- Ask for detection and shot-accuracy data broken down by water turbidity and ambient light, not a single headline number.
- Ask for the false-positive rate and what the system does when confidence is low — withholding a shot should be the default.
- Establish the servicing interval for optical windows, and who performs it.
- Ask for lice-count trends across a full production cycle at a comparable site, including winter, not a four-week trial.
- Confirm how the unit's counts reconcile with the manual counts your reporting obligations are based on.
Who this affects beyond the farms
Sea lice from farms are a conservation issue as well as a production one, because larvae drift from pens into the migration routes of wild salmon and sea trout smolts. Any technology that lowers the reproductive output of lice inside pens reduces that external pressure, which is why lice policy in Norway is tied to the health of wild stocks in the surrounding production area and not only to the farmer's own fish.
For robotics and automation practitioners, the sector is a useful case study in a specific class of problem: real-time perception in a degraded optical medium, with an actuator whose misfire has a biological cost, deployed unattended for months. The constraints that shape these systems — fouling, power budget, servicing access, and a conservative confidence threshold — recur in most submerged autonomy work, and salmon farming is one of the few places funding them at commercial scale.
Frequently Asked Questions
Does the laser hurt the salmon?
The pulse is tuned so the louse's pigmented body absorbs the energy while the salmon's reflective skin reflects it. Systems also withhold a shot when target confidence is low, since a misaimed pulse near an eye or gill is the main welfare risk.
Can robots replace chemical and thermal delousing entirely?
Not currently. In-pen robotics suppresses lice load and reduces how often a site needs a full treatment, but farms generally retain conventional methods as a fallback when numbers rise.
Why do submerged feeding and lights help if they do not touch the lice?
Infective lice larvae concentrate in the top few metres of the water column. Keeping salmon feeding and swimming below that layer reduces how many larvae ever reach them.
How is lice counting automated?
Camera rigs in or above the pen estimate lice load from imagery of passing fish, reducing the manual sampling that requires netting and anaesthetising individual salmon. Counts from these systems still have to reconcile with the manual counts that reporting obligations are based on.



