AI underwater robots can now track diver stress via exhaled bubbles

University of Minnesota Twin Cities researchers developed an AI-based system that enables camera-equipped underwater robots to estimate a scuba diver’s respiration rate by observing bubbles exhaled through the regulator. The study, published in The International Journal of Robotics Research, is described as the first use of robotic vision to estimate a diver’s human respiration rate. The approach is intended to give divers a robotic safety partner that can monitor physiological stress without physical contact. Diving can involve exhaustion and potentially dangerous respiratory distress, while conventional medical sensors may not work reliably underwater because wetsuits and drysuits interfere with body contact. Wireless transmission through water is also limited. By measuring the frequency and volume of exhaled bubbles, the robot can identify indications of stress, hyperventilation, or exhaustion in real time. To train the model, the team used a fuzzy-labeling method for underwater images. Researchers manually classified thousands of images and synchronized them with the recognizable audio signals from regulator exhalations, helping the system learn how breaths appear in visually challenging water conditions. The dataset included audio and visual recordings gathered in varied environments, temperatures, and water clarity levels, including Lake Superior in Duluth, Square Lake in Stillwater, and the Caribbean Sea near Barbados. In field trials, autonomous underwater vehicles used a communication system called HREyes to report breathing status. It classified rates below 14 breaths per minute as below-normal, 14 to 20 as normal, and above 20 as above-normal. The researchers plan to combine respiration estimates with analysis of diver movement to create a broader wellness profile for deep-sea exploration safety. The team included Junaed Sattar, lead author Demetrious Kutzke, and undergraduate researcher Vennela Dupati.