How robots and divers work together to fix underwater power cables

The recent blackout on a remote island highlighted a critical challenge: locating a break in an underwater power cable. Conventional solutions involve either retrieving the entire cable or deploying remotely operated vehicles (ROVs) to inspect the line. But what if an autonomous underwater vehicle (AUV) could map the cable and identify the fault location, guiding a diver directly to the problem?

This is the focus of a groundbreaking project at MIT Lincoln Laboratory, funded through an internal research initiative on autonomous systems. The initiative is led by the Advanced Undersea Systems and Technology Group, which is exploring ways to combine human expertise with robotic precision to enhance maritime operations for the U.S. military. These operations include inspecting and repairing critical infrastructure, conducting search and rescue missions, securing harbor entries, and neutralizing underwater threats like mines.

The strengths and limitations of humans and robots underwater

Underwater missions often require human divers because robots lack the dexterity needed for tasks like repairing infrastructure or disarming explosive devices. As principal investigator Madeline Miller notes, "Divers and AUVs generally don’t team at all underwater." Even remotely operated vehicles (ROVs) struggle with complex manipulation tasks due to the limitations of their manipulators in underwater environments.

While humans excel in object recognition and hands-on repair, they face significant challenges underwater. Heavy equipment slows movement, and complex calculations are difficult to perform in low-visibility conditions. Robots, on the other hand, offer advantages in processing power, speed, and endurance—capabilities that complement human skills. The goal of Miller’s team is to bridge these strengths by developing advanced navigation and perception systems for underwater human-robot collaboration.

Challenges in underwater navigation and perception

Navigating underwater is inherently difficult. Divers often rely on a compass and estimates of fin kicks to track their position, but landmarks are scarce, and visibility can be poor due to darkness or suspended biological matter. Without reliable reference points, divers risk becoming disoriented or lost. Robots can assist by perceiving their environment, but traditional optical sensors like cameras fail in dark or murky conditions. Acoustic sensors such as sonar provide some visibility but produce images that lack color and detail, showing only shapes and shadows.

A major hurdle is the lack of large, labeled datasets for sonar imagery, which are essential for training perception algorithms. Even with sufficient data, the dynamic ocean environment complicates object recognition. For example, a downed aircraft broken into fragments or a tire covered in mussels may no longer resemble their original forms, making it difficult for AI to classify them accurately.

Miller emphasizes the need for solutions that work in expeditionary environments—areas with little or no prior mapping. "For missions like harbor entry, we might only have a satellite map, with no underwater data," she explains. "In such cases, we need systems that can adapt in real time."

Developing navigation algorithms for human-robot teaming

Miller’s team built upon earlier work by the MIT Marine Robotics Group, led by John Leonard, which developed diver-AUV teaming algorithms. These algorithms were initially tested in simulations and calm-water field trials using human-paddled kayaks as stand-ins for both divers and AUVs.

The next phase involved integrating these algorithms into a mission-capable AUV and testing them in realistic ocean conditions. Early tests used a support boat as a diver surrogate before progressing to actual divers. The team quickly discovered that ocean currents introduced complexities that required additional sensing on the diver’s part.

"The original algorithms were designed for ideal conditions," Miller says. "They only needed to calculate the distance to the diver at intervals to estimate positions over time. But real ocean conditions—with currents, waves, and unpredictability—make this optimization problem far more complex."

Enhancing perception with AI and real-time feedback

To improve underwater object recognition, Miller’s team is developing an AI classifier capable of processing both optical and sonar data during missions. When the classifier encounters uncertainty—such as an object it cannot confidently identify—it sends a bounding box and a tentative label to the diver for verification. The diver can then respond with corrections or additional details, creating a feedback loop that improves the AI’s accuracy over time.

This system relies on underwater acoustic modems for communication between the AUV and diver. However, underwater acoustic communication is notoriously slow. Transmitting an uncompressed image could take tens of minutes, far too long for real-time collaboration. To overcome this, the team is exploring ways to compress essential information into minimal, actionable data packets that fit within the constraints of low bandwidth, high latency, and limited hardware capabilities.

Prototyping and future testing

The team has conducted field tests along the coastal waters of New England, including open-ocean trials near Portsmouth, New Hampshire. These tests involved the University of New Hampshire’s research vessels, Gulf Surveyor and Gulf Challenger, acting as diver surrogates. The sensor-equipped AUV, designed to integrate seamlessly with U.S. Navy AUVs, carries a payload of commercial off-the-shelf (COTS) sensors and processing boards, including sonar, optical cameras, an acoustic modem, and computing modules.

Looking ahead, the team aims to refine their algorithms and hardware, focusing on improving real-time adaptability in dynamic underwater environments. The ultimate goal is to transition this technology into operational use, enabling faster, safer, and more efficient underwater missions for both military and civilian applications.

As Miller puts it, "The future of underwater missions lies in the seamless collaboration between humans and robots—combining the best of both worlds to tackle challenges that neither could solve alone."