Underwater robots may one day help keep scuba divers safe simply by watching the bubbles escaping from their breathing equipment. Researchers at the University of Minnesota have demonstrated a vision system that estimates how often a diver breathes without requiring a new sensor attached to the person's body. The project turns an ordinary visual clue, a stream of bubbles, into information a robotic companion can use. It is an inventive piece of research with obvious appeal, but it is not a commercially available underwater rescue service or a replacement for established diving precautions.
The university's report on the underwater robot describes tests involving water in Minnesota and the Caribbean, with a system designed to identify exhalations from camera footage. The team also used sound recordings to help prepare its training material. Unlike a fictional robot that can diagnose every problem from a glance, this system focuses on one measurable sign: breathing rate. That narrow objective is part of what makes the experiment credible, and it also defines its limitations.
Underwater Robot Breathing Works Through Bubbles
A diver breathes through a regulator and exhales air into the water. From a distance, the resulting bubbles can be visible even when the equipment and clothing make it difficult to attach ordinary sensors. The robot's camera looks for visual patterns associated with those exhalations and converts the observations into an estimate of breaths per minute. To a human observer, the idea is intuitive. The engineering difficulty lies in making that estimate reliably while both the diver and the camera are moving.
Underwater footage can be murky. Light changes with depth, the water contains moving particles and bubbles do not always rise neatly through an empty background. The researchers therefore worked with recorded examples collected across different environments. Their training approach used audio cues from breathing regulators to help identify relevant moments in video. The goal was to teach the vision system which changing patches of an image actually represent a breath rather than a wave of sediment or unrelated movement.
The work is described in The International Journal of Robotics Research, where the study was first published online in January 2026. The university reported on its development in July. Those dates matter because this is an evergreen account of a research advance, not a supposed announcement made today. The paper concerns estimating respiration rate during underwater human-robot collaboration. Its publication does not establish that the system is approved as medical equipment or ready for everyday dive centres.
Why Divers Cannot Always Wear Ordinary Sensors
Monitoring breathing is familiar on land. In a hospital or sports setting, a person may be connected to equipment or wear a device that measures relevant signals. Underwater conditions create different obstacles. Diving suits, equipment placement and the physical properties of water complicate the use of sensors that depend on direct contact or ordinary wireless communication. A robot that can estimate breathing without touching its human companion offers a potentially useful additional source of information.
That does not mean visual information alone reveals why a diver is breathing quickly. Excitement, exercise, cold, anxiety and other factors may influence a person's rate. A number on a screen has to be interpreted in context. The researchers describe the system as a step towards monitoring wellbeing, not as a machine that can instantly distinguish every possible medical emergency. The important potential is earlier awareness of a meaningful change, followed by appropriate human judgement and established safety responses.
A companion robot might be able to maintain an observation position while a diver concentrates on a task. That could prove interesting for scientific work, inspection or training exercises, where human attention is already divided. Yet the robot must itself be reliable and must not distract the diver or introduce new hazards. Any real deployment would need careful evaluation of its behaviour around people, lines, visibility constraints and the consequences of false alarms. Watching bubbles is only part of an operational safety system.
Reading a Moving Scene Is Harder Than Watching a Person
Imagine trying to count breaths through a window while the person is swimming, the camera is moving and the view fills with drifting reflections. A human can often use context to guess what is happening. A computer vision system needs methods for deciding which visual changes correspond to the event it was trained to recognise. It must also decide when the evidence is too weak to make a useful estimate. A model that always supplies an answer, even with a poor view, could be less helpful than one that admits uncertainty.
The Minnesota team assembled examples from more than one body of water, including sites with different clarity and conditions. This is a sensible research choice because a model trained on perfect swimming-pool footage might fail on an actual expedition. Still, testing several environments is not the same as proving performance across all seas, depths, lighting conditions and equipment. The study should be understood as an encouraging demonstration whose limits have to be investigated further.
There is also the practical problem of following the right person. The published work focuses on a single scuba diver, according to the paper title and university description. Watching several people at once would require identifying whose bubbles belong to whom and maintaining that distinction as the group moves. A robot may need to handle overlap, hidden faces and streams of air crossing in the same frame. Expanding to multiple divers is an ambition discussed by the researchers, rather than an ability that can be assumed from the initial results.
The team described a communication component that could report breathing categories to the diver or supporting system. The categories are a way to convert an estimate into something potentially useful. They should not be mistaken for clinical diagnoses or universal rules that apply identically to every individual and diving situation. A practical safety product would need evidence that its alerts lead to better decisions without creating unnecessary distraction. That would require testing beyond demonstrating that bubbles can be counted.
A New Sort of Robot Partner
Robots in water are often imagined as tools for surveying the seabed, inspecting structures or carrying cameras where people cannot easily go. A machine that watches a diver's breathing suggests a different relationship. It becomes a partner observing the human's condition while they carry out another task. This is distinct from experimental robot fish designed to study movement, which use aquatic machines to investigate how bodies travel through water. Both fields show how researchers are using physical robots to ask questions that cannot be answered by a software simulation alone.
The observation could potentially be combined with a diver's movements. Someone swimming unusually slowly while breathing rapidly may require a different interpretation from a person moving vigorously at a normal rate for that activity. The Minnesota researchers have identified movement analysis as a direction for subsequent work. It would be premature to describe a finished combined system, but the idea shows why breathing rate is a useful starting point rather than a complete portrait of safety.
The approach also raises design choices about how a robot communicates. Underwater sound and light signals need to be clear without alarming people unnecessarily. A useful warning should help a diver make a decision rather than interrupt them with ambiguous data. Operators might also want to know whether a reading was continuous or interrupted by poor visibility. In high-consequence settings, how a system expresses uncertainty can be as important as its average performance in a controlled trial.
What the Technology Could Change
If future tests show robust results, a companion robot could provide an extra observation channel during some dives. It might help a team notice an unexpected breathing pattern while working on an inspection or research task. A remote supervisor could also benefit from information that is hard to gather by ordinary visual monitoring. These are potential applications, not claims of proven injury prevention. No research team can infer fewer accidents merely because a model recognises bubbles in a demonstration.
Elsewhere, researchers are exploring ways for AI to observe marine life without relying entirely on a person watching continuously. LiveAIWire's report on AI-assisted orca detection is one example of a different monitoring problem in the same challenging environment. Satellite analysis of fishing vessels offers another. The shared attraction is turning large volumes of observations into something useful, but the data and consequences vary greatly. A wildlife sighting can be checked later; a warning involving a diver demands much more immediate reliability.
A sensible diving operation would continue to rely on training, buddy systems, equipment checks and accepted safety procedures. An AI observation tool should supplement those protections rather than encourage riskier behaviour. Some of the most important questions are mundane: how long does the robot operate, how often does it lose sight of the diver, and what happens when its battery fails? Until those are answered in real deployments, the research remains a technical contribution rather than a promise of complete protection.
The memorable image is a robot watching a stream of bubbles and translating it into a breathing rate. It makes sophisticated machine vision feel concrete because the cue is something anyone who has seen scuba diving can understand. The scientific achievement is not that a machine can magically read a person's mind underwater. It is that researchers have found a practical way to estimate one useful bodily signal in a setting where ordinary sensors are difficult to use. That is promising precisely because its purpose is limited and testable.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.