In the dark waters 600 meters below Monterey Bay, a torpedo-shaped robot called Mesobot drifted silently behind a pulsing siphonophore, its low-light cameras recording every contraction of the colony’s translucent body. No pilot guided the vehicle. An onboard algorithm, trained on hundreds of thousands of labeled deep-sea images, kept the animal centered in the frame while the robot matched its pace and heading. The encounter lasted hours, not the minutes a crewed submersible typically manages before its thrusters scatter fragile midwater life.
That scenario captures a shift now supported by multiple peer-reviewed studies and institutional programs: autonomous underwater vehicles equipped with artificial intelligence can detect, follow, and classify deep-sea animals with minimal human direction. For marine biologists who have spent decades relying on expensive, human-piloted submersibles for brief glimpses of the ocean’s least-explored zone, the technology opens a fundamentally different way to collect data.
Teaching robots to see in the deep
Several independent research efforts have converged on the same problem: giving underwater robots the ability to recognize marine animals in real time without requiring a person to label every frame of video.
A study led by researchers at the Woods Hole Oceanographic Institution and published in the International Journal of Computer Vision describes an evaluation framework for semi-supervised visual tracking of marine animals using an autonomous underwater vehicle. “The key challenge is reducing the annotation burden while maintaining reliable tracking in open water,” the paper explains, noting that semi-supervised methods, which blend a small set of human-labeled examples with large volumes of unlabeled footage, can sharply cut the volume of manual labeling required without sacrificing performance. The researchers demonstrated the approach running onboard an AUV during field trials.
Those tracking algorithms depend on high-quality training imagery. FathomNet, a large labeled underwater image database described in a Scientific Reports paper archived through CaltechAUTHORS, serves as a primary resource. Kakani Katija, a principal engineer at the Monterey Bay Aquarium Research Institute and a lead contributor to FathomNet, has described the database as an effort to “democratize access to deep-sea visual data” so that machine-learning researchers outside traditional oceanographic institutions can contribute to ocean science. The database contains expert-verified annotations spanning hundreds of deep-sea species, and its open data and code availability statements let other research groups inspect the dataset, reproduce results, and adapt models for different vehicles or habitats. Trained models built from FathomNet data have already been integrated with robotic vehicles to enable automated species detection during dives.
Hardware built to be invisible
Algorithms alone are not enough. Many midwater organisms, from fragile larvaceans to light-sensitive dragonfish, flee or freeze when hit by the bright lights and loud thrusters of a conventional remotely operated vehicle. That behavioral disruption has long been a blind spot in deep-sea biology: scientists could only study animals that tolerated being approached.
Mesobot, developed by the Monterey Bay Aquarium Research Institute, was engineered specifically to close that gap. The hybrid vehicle can operate on a tether or switch to fully autonomous mode. According to MBARI’s institutional reporting, it carries low-light 4K cameras and uses red-spectrum lighting and ultra-quiet propulsion designed to avoid startling the organisms it tracks. The engineering goal is to hover in the midwater column and follow individual animals for up to 24 hours, long enough to observe feeding, migration, and predator-prey interactions that have never been recorded continuously.
A separate peer-reviewed study in Frontiers in Marine Science details a complementary approach called tracking-by-detection. Researchers used a stereo camera system and trained an object detector across multiple taxonomic groups, then ran field trials in Monterey Bay National Marine Sanctuary using a proxy ROV. The experiments showed that a detection model trained on curated imagery could guide a vehicle to keep animals centered in the camera’s field of view during real dives, demonstrating that computer vision can close the loop between perception and vehicle control, not just annotate footage after the fact.
The data pipeline connecting it all
Behind the vehicles and algorithms sits shared infrastructure that makes the work reproducible. The Woods Hole Oceanographic Institution’s Autonomous Robotics and Perception Laboratory, known as WARP, hosts the VMAT dataset referenced by the International Journal of Computer Vision paper, providing a canonical source of curated video sequences for evaluating marine-animal tracking algorithms. MBARI’s Video Annotation and Reference System (VARS) catalogs decades of deep-sea video with standardized descriptive and quantitative annotations, supplying the labeled training images that feed into AI models.
Together, these platforms create a pipeline from raw video to standardized annotations to machine-learning-ready datasets that can be shared across institutions. That openness matters: it means no single lab controls the benchmarks, and outside researchers can stress-test published claims with their own data.
Where the technology still faces hard limits
For all the progress, significant gaps remain between controlled demonstrations and routine operational use.
According to NOAA Ocean Exploration, a deployable “detector/supervisor/agent” AI stack has been integrated with underwater vehicles to detect and autonomously follow animals with limited human oversight. The program page references MiniROV tests in Monterey Bay in October 2024 and a subsequent January window, but no primary performance metrics from those tests have appeared in a peer-reviewed venue as of June 2026. The page functions as a program preview, not a results report.
Mesobot’s 24-hour tracking ambition, while described in MBARI’s institutional materials, lacks independent peer-reviewed evaluation across varied deep-sea conditions. The most detailed public account of the vehicle’s design dates to MBARI’s 2019 annual report. Whether Mesobot has achieved sustained autonomous tracking at that duration in practice, and how consistently it performs across different oceanographic settings, is not confirmed by available published sources.
The Frontiers in Marine Science tracking-by-detection study does not specify the taxonomic limits of its object detector. The model was trained across multiple groups, but the exact number of species or genera it can reliably distinguish, and at what accuracy thresholds, is not detailed. Without those figures, it is difficult to judge whether the system is best suited for broad functional categories (gelatinous versus crustacean, for example) or for the finer-grained species-level identification that many ecologists need.
A broader open question is generalization. Most demonstrations have taken place in Monterey Bay, where institutions hold decades of video archives and well-characterized species assemblages. How these AI systems perform in unfamiliar regions, with species absent from training sets, or under variable lighting and current conditions has not been quantified in published benchmarks. The gap between a successful regional trial and reliable global deployment is real, even as the individual technical building blocks are now proven.
Why patience and stealth may reshape deep-sea biology
The core components of AI-enabled animal tracking in the deep ocean are technically demonstrated. Peer-reviewed studies show that underwater robots can detect and follow marine life without continuous human control, and open datasets like FathomNet and VMAT give the broader research community tools to build on that work.
Still, the distance between a promising Monterey Bay trial and a system that operates reliably across the world’s ocean basins is considerable. Forthcoming studies reporting long-duration deployments, cross-habitat benchmarks, and transparent error analyses will be the real test. Until those results arrive, the technology is best understood as a powerful new capability for targeted research, not yet a turnkey solution for monitoring the deep sea at scale.
For the animals themselves, the stakes are straightforward. The midwater zone, Earth’s largest living space by volume, remains one of its least observed. Robots that can watch without disturbing, for hours instead of minutes, stand to reveal behaviors and ecological relationships that human-piloted vehicles have never had the patience or stealth to capture.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
