What underwater robots can do after a flood like Nepal’s, and what they cannot

On the morning of 26 August 2026, part of a glacier came off Langtang Lirung. The collapse put enough rock and ice into the valley to register on seismometers around the world, and the resulting debris flow travelled close to 100 km down the Lende Khola and the Trishuli, destroying the Gyirong border crossing and striking settlements along roughly 72 km of river. Nepali authorities have reported over 1,300 dead and several thousand still missing. Some bodies were carried 240 km into India.

We build simulation software for marine robots. It would be easy, and wrong, to write a post arguing that our field would have changed that morning. It would not have. What follows is an attempt to be precise about where underwater robotics is irrelevant to an event like this, and where it is genuinely useful — because the second category is larger than most people outside the field assume, and it is almost entirely in the weeks afterwards rather than the hours.

The first hours: no

A glacial debris flow is not water. It is a moving slurry of rock, ice, sediment and whatever it has already destroyed, travelling fast enough to remove bridges and roadbed. Nothing remotely resembling an underwater vehicle survives contact with it, and nothing useful could be observed from inside it if it did.

The warning problem in the first hours is also not a robotics problem. It is a sensing and communications problem: gauges upstream, seismic monitoring, and a way to get a message to a village before the water does. The water-level station at Syabrubesi recorded the Trishuli well below its official warning level shortly before the surge arrived. That gap — between the instrument reading normal and the valley being gone — is the thing worth engineering against, and it is answered with monitoring networks, not submersibles.

So: on 26 August, no. It is worth saying plainly, because a great deal of disaster-adjacent technology marketing does not.

The weeks after: a genuine role

Once the flow stops, the valley is full of water that has to be worked in and cannot be seen into. This is where small remotely operated vehicles have a real and unglamorous job, and where the alternative is putting a diver into it.

Flooded powerhouse tunnels

The floods damaged a large number of hydropower facilities — reporting has put the figure at 14 hydropower and solar projects totalling roughly 748 MW, with other counts of damaged facilities considerably higher. At Trishuli 3A and neighbouring plants, rescue teams spent days working into tunnels, pumping water out, clearing silt, and advancing by rafts and boats through partially flooded headrace and access tunnels. Two men were pulled alive from a Trishuli 3A tunnel nine days after the flood.

A flooded tunnel is close to the ideal case for an observation-class ROV: confined, no visibility, unknown obstruction, and lethal for a diver. A vehicle on a tether with a small imaging sonar and lights can answer questions that otherwise require dewatering the whole structure first — how far the silt extends, whether the passage is open, whether there is an air pocket, whether anything is in there. It does not rescue anyone. It tells the people who will do the rescuing what is in front of them.

Impounded water and new lakes

Debris dams create lakes that were not there a week ago, and those lakes sit above people. In this event a landslide-formed barrier lake in China hampered rescue work, and a river in Nepal swelled as debris choked it. Understanding one of these requires knowing the shape of the impoundment and the condition of the plug holding it, both of which are underwater.

Multibeam or side-scan survey from a small boat gives bathymetry of a new lake in hours. That is the input to whether the barrier is stable, how much volume is behind it, and what happens downstream if it goes. It is survey work, not rescue work, and it is the kind of thing that decides whether a valley is evacuated a second time.

Bridges, piers and the road back

Officials confirmed 32 bridges and close to 25 miles of road swept away. Every surviving bridge in the affected reach is now a question mark: scour around the piers is invisible from the deck, and a pier that looks intact can be sitting on nothing. Bridge scour inspection is an established ROV and sonar task in ordinary conditions, and it is on the critical path to reopening a road.

Search in standing water

This is the application people think of first, and it deserves the most caution. Nepal has done it before: after the 2024 Simaltal landslide swept two buses into the Trishuli, the search effort included divers, water drones and magnets, and recovered a fraction of those lost over many days. Rivers are the hardest search environment there is — current, zero visibility, snagging debris, and a search area that grows every hour the current runs.

Sonar-equipped submersibles do contribute to this work, and families are owed the effort. But an honest account has to say that recovery rates in moving water are poor regardless of equipment, and that the more than 1,300 bodies recovered by early September in this event were overwhelmingly found by people on the ground, not by machines.

What actually limits this

The barrier is rarely the vehicle. Observation-class ROVs are affordable enough that a regional disaster authority can own several. The barriers are the ones nobody photographs:

• Nobody has flown one in that water. A pilot who has only worked clear reservoirs is not ready for a tunnel full of suspended silt with a live current and a tether that will snag.
• Turbidity defeats cameras completely. In flood water the optical payload is decoration. Everything useful comes from acoustics, and reading a sonar image is a skill that takes months.
• Access outruns capability. If the road is gone, the vehicle arrives on somebody’s back or not at all.
• Sovereignty and coordination. Nepal and China have been selective about outside assistance in this response. Capability that cannot be invited in is not capability.

The honest summary

Marine robotics would not have prevented what happened in Rasuwa and Nuwakot on 26 August. Claiming otherwise would be indecent. What it does is shorten the period afterwards in which nobody knows what is under the water: which tunnel is passable, which barrier is holding, which pier is standing, where the water goes next. In a response measured in weeks and an economic loss Nepali authorities have put between four and seven billion dollars, that is not a small contribution. It is just a different one from the one people imagine.

The figures in this piece are provisional and were current as of mid-September 2026. Casualty numbers in particular have moved substantially week to week and will keep moving.