How Does a Robot Sub Decide Which 16 Photos to Send Home?
A robot sub surveying the seafloor takes thousands of photos in a few hours. Out at sea with no ship nearby, its only link to shore is a satellite that carries less than a hundred bytes a second, so nobody sees those photos until the sub comes back.
Researchers at the University of Southampton, the National Oceanography Centre, Ocean Infinity and Voyis built a way around that: the sub sorts its own photos with a neural network and sends home a summary, a handful of photos that stand in for the rest plus a line of data for every photo it sorted.
Off Gran Canaria, an earlier version of the method summed up almost three hours of photos in 90 kilobytes, and the summary reached shore by satellite in about 34 minutes. In this independent explainer, I walk through how it works and what the paper shows.
In the video
- Why surveys make so many photos: a camera has to be within about 10 meters of the bottom, so each photo covers only a small patch
- The gap: the camera makes about 100,000 times more data than the satellite link can carry, and even hard JPEG compression leaves about 1,000 times too much
- Two modes: in summary mode the sub picks the photos to send by itself; in query mode someone on shore picks an example photo and asks, "have you seen anything like this?"
- How the sub sorts: a neural network turns each photo into a list of numbers, so look-alike photos get similar lists. The sub groups them, splits each group again, and picks the photo nearest the middle of each small group
- Packing it up: small messages that can arrive in any order, and a lost one only leaves a gap on the web page on shore
- Three field trials: Gran Canaria; Plymouth, where a cable and a mock repeater were laid on the seabed and the results went over Wi-Fi after an electrical fault kept the sub from using its satellite modem; and Shetland, 151 kilometers in three days, steered over the horizon by satellite
- The offline check on the Plymouth photos: asked for photos like the mock repeater, the general-purpose network (DINO) found almost all 12 in the set, and the site-trained one (SimCLR) found none. The paper doesn't say why
- What the paper doesn't show yet: each trial ran a different version or link, the accuracy numbers come from one site's photos, and it doesn't yet show anyone changing a mission because of a summary
- What a summary isn't: a replacement for the full photos, which still get processed once the sub is back
The paper
"Remote Awareness of Seafloor Images Collected by AUVs over Low-Bandwidth Communication Links"
Adrian Bodenmann, Cailei Liang, Miquel Massot-Campos, Samuel Simmons, Alexander B. Phillips, Alberto Consensi, Matthew Kingsland, Rashiid Sherif, Stan Brown, Adam Riese, Blair Thornton
University of Southampton, National Oceanography Centre, Ocean Infinity, Voyis Imaging and The University of Tokyo
IEEE Journal of Oceanic Engineering (early access, 2026), doi:10.1109/JOE.2026.3708604
Free version: arXiv:2607.18013, July 2026
Funded by the European Union's Horizon 2020 research and innovation program (grant 101000858, TechOceanS) and UK Research and Innovation (project 10110715, OASIS).
The paper goes much deeper than this video: how the photos are encoded, clustered and compressed into satellite messages, the vehicles and cameras on each trial, the run-by-run tables behind the summary results, and the query results for every example photo. If you run seafloor imaging surveys or work on AUV autonomy, it's worth reading in full.
This is an independent explainer. I'm not affiliated with or endorsed by the authors, the University of Southampton, the National Oceanography Centre, Ocean Infinity, Voyis Imaging or The University of Tokyo, and the research is all theirs. The animations are my own illustrations of what the paper describes, not footage or photos from the study: the photo tiles, maps and tracks are illustrative, and the two charts are redrawn from the paper's Figs. 7 and 8. Any mistakes in the explanation are mine.
I'm Bora Celik from Piccard. I explain new research in ocean science and robotics.