MIT Develops Human-Robot Teams for Underwater Cable Inspection

Researchers at MIT Lincoln Laboratory are advancing the integration of human divers and autonomous underwater vehicles (AUVs) to address complex tasks such as the inspection and repair of underwater power cables. The initiative, led by principal investigator Madeline Miller within the Advanced Undersea Systems and Technology Group, aims to leverage respective human and machine strengths to aid maritime missions relevant to the U.S. military, including infrastructure maintenance, search and rescue, harbor entry, and countermine activities.

While divers possess unmatched dexterity and object recognition abilities underwater, they are limited in terms of computational capacity and speed, especially in challenging ocean conditions. In contrast, AUVs offer high processing power, rapid movement, and endurance but struggle with tasks involving nuanced manipulation or object identification in murky waters. To bridge these gaps, the MIT team is developing new hardware and AI-driven algorithms focused on navigation and perception.

A significant challenge relates to ensuring precise navigation for both humans and machines. Traditional diver navigation tools such as compasses and fin-kick counts are imprecise, particularly where visibility is low and landmarks are absent. Robots, for effective guidance, need the ability to perceive and interpret their dynamic, often opaque, underwater environment. The limitations of both optical and sonar sensors—ranging from poor image quality to a lack of comprehensive labeled datasets for training AI models—have complicated this task. Underwater, sonar images show only outlines and shapes, making AI classification difficult, particularly when objects are degraded or fragmented.

Building on earlier efforts by the MIT Marine Robotics Group, Miller's team integrated advanced navigation algorithms into AUVs for field testing, initially using boats to simulate diver movement before transitioning to tests with human divers. Their research revealed that ocean currents and real-world conditions demand sophisticated sensing capabilities on both robots and humans. Calculations that worked in controlled environments became ineffective in turbulent waters, underscoring the need for robust data fusion and adaptive algorithms.

On the perception front, the team developed an AI classifier capable of fusing sonar and optical data and designed to request feedback from human divers when uncertain. This interactive approach allows divers to validate or correct AI-driven object classifications, a process that relies on underwater acoustic modems to enable communication. However, the severely limited bandwidth of such modems means information must be compressed—transferring even a single image would otherwise take minutes.

The sensor payload created for these tests combined commercial off-the-shelf sensors and computing hardware, including sonar, optical sensors, and acoustic modems for diver tracking and data transfer. Field deployments included collaborative testing with the University of New Hampshire as well as deliberate, diver-focused trials in both the open ocean and on rivers. Recent experiments at Michigan Technological University’s Great Lakes Research Center involved human divers using a prototype ‘tube-let’ device equipped with multiple sensors to gather navigation data for the AUVs’ algorithms.

Another line of research, led by PhD student Caroline Keenan, explored whether optical recognition systems—capable of identifying objects in clear water—could be used to train AI models for sonar data analysis. This knowledge transfer seeks to minimize the labor involved in labeling and training AI for underwater perception.

As the internally funded phase of the project concludes, the MIT team is now seeking external partners to further refine the technology and transition it for military or commercial use. According to Miller, effective human-machine collaboration in the undersea domain is increasingly critical as global reliance on subsea infrastructure grows and as geopolitical competition intensifies. Advances in AI-powered human-robot teaming offer a promising path to enhanced maritime security and infrastructure resilience.

mit.edu