NEWS
Searching for Sunken Treasure
Researchers at the University of Michigan and Thunder Bay National Marine Sanctuary are relying on underwater autonomous vehicles and machine learning to detect and map shipwrecks.
Written by Kayt Sukel
IN GUY RITCHIE’s 2025 ACTION-ADVENTURE FILM, Fountain of Youth, the lead characters, played by John Krasinski and Natalie Portman, use innovative engineering techniques to raise the RMS Lusitania from the depths of the sea. While the audience is treated to the sudden (and hair raising) lift of the ship, which had lain on the floor of the Celtic Sea since 1915, what’s not shown is the painstaking work it takes to detect and then map the remains of sunken vessels.
The Global Foundation for Ocean Exploration estimates there are more than 3 million shipwrecks across the globe and less than 1 percent have been explored. Today, government entities, commercial salvagers, preservation societies, archeologists, and, yes, the odd well-funded adventurer are working to find these wrecks. Some are doing so to recover important cargo, and others to understand vital parts of our history. But salvage operations are complex undertakings. Couple the depth and salinity of the waters with the power and pressure of currents—then add in the ravages of time—and there are various points where engineers must apply innovative techniques to safely identify and recover wrecks.
Katie Skinner, an assistant professor of robotics at the University of Michigan, focuses on the first parts of the salvage puzzle: detection and mapping.
An underwater drone surveys the massive propeller of a long-submerged shipwreck, revealing details of a vessel frozen in time beneath the ocean. Photo: University of Michigan
“My work focuses on marine robotics, specifically on sensing and perception,” she said. “We need to be able to interpret sensor data, collected from autonomous underwater vehicles (AUVs), and then use it for different downstream tasks like shipwreck detection and navigation mapping.”
Skinner has been working closely with the Thunder Bay National Marine Sanctuary, an organization responsible for protecting nearly 100 historic shipwrecks in Lake Huron. While this organization does not bring up the lost vessels, they do important surveying work to keep track of them over time. Historically, to both identify and map the wrecks, the Sanctuary’s teams required survey vessels using sonar and human divers—which was both cost and time intensive.
“They would do these large area searches using ships with multibeam sonar. They would then look at the data over days or even weeks to identify potential sites for a wreck. Then they must replan to go back out and investigate those sites using divers,” she said. “We are looking at ways to use robotics to improve that pipeline, making the process more efficient so we can scale it up more broadly while reducing costs.”
Skinner and her team have developed an underwater autonomous vehicle (UAV), equipped with side sonar that can “see” 100 meters on either side of the vehicle to do those surveys. This allows them to collect data in a more efficient manner. She and her colleagues are also developing AI algorithms to process that data to flag wreck sites.
“With AI, we can process all of the data right on board the ship and get some initial detections within minutes of getting the vehicle back on board,” she said. “This means we can make decisions while we are still in the field about what sites to return to.”
Developing those algorithms, however, is not without its challenges. As any SCUBA diver knows, underwater conditions can change on a dime, which can interfere with the sensing data collected. And, as machine learning systems rely on training data, Skinner has had to leverage simulation data to get her algorithms up to speed.
“One of the biggest challenges is that we don’t have access to large amounts of labeled data like we have on land. ”
—Katie Skinner, assistant professor of robotics at the University of Michigan
“One of the biggest challenges is that we don’t have access to large amounts of labeled data like we have on land,” she said. “We have a simulated environment, with a shipwreck, and we’re simulating side scan sonar in our data, which can help us train our models. But simulation doesn’t always look like the real data. There’s going to be a gap.”
Despite those obstacles, Skinner said she and her team are making progress. They are now focusing on how to do autonomous shipwreck mapping with UAVs, sending the robot to collect imagery within 2 meters of a wreck site and then plan a path to survey the extent of the shipwreck to collect data.
“To do that, the robot has to be quite advanced,” she said. “It has to be able to estimate its position as it’s traveling throughout the survey and then use the data it’s collecting to build a dense map of the shipwreck so that it can improve its navigation as it’s going.”
Skinner and her colleagues have gotten “pretty good” at underwater robotic surveying, but when it comes to shipwrecks and underwater environments, there are always new problems to solve. That’s part of the appeal, she said. And those working in ocean environments have to contend with additional design constraints from the corrosiveness of salt water to greater water depths. But she is excited about what the future holds.
“Marine robotics is such an exciting field because it allows to explore places that humans can’t go,” she said. “I hope that our work, and the other work happening in shipwrecks, inspires others to keep exploring and think about all the ways engineering and technology can help push boundaries in these different environments and different domains.”
Kayt Sukel is a technology writer and author in Kansas City.
© 2026 The American Society of Mechanical Engineers. All rights reserved.