University of Minnesota Robots Estimate Diver Breathing
University of Minnesota Twin Cities researchers announced on July 24 an AI vision system that estimates a scuba diver's respiration rate from exhaled bubbles captured by an underwater robot. The study compares four bubble-classification methods using synchronized underwater audio and video, then converts visual detections into breaths-per-minute estimates; the published evidence validates respiration estimation, not stress diagnosis or emergency prediction.
University of Minnesota Twin Cities researchers announced July 24 that they developed an AI-enabled underwater-robot system to estimate a scuba diver's respiration rate by analyzing exhaled bubbles in camera footage. The work, published in The International Journal of Robotics Research, addresses contactless physiological monitoring in underwater human-robot collaboration.
The study, "Robotic estimation of single scuba diver respiration rate for safety in underwater human-robot collaboration," estimates Human Respiration Rate, or HRR, from the visual presence and timing of bubbles generated during scuba exhalation. According to the journal abstract, the researchers collected an audio-visual dataset from several locations and water bodies, evaluated four methods for classifying bubbles in images, and used a respiration-state algorithm to calculate breaths per minute. They compared the resulting estimates with measurements from human analysts.
Training vision with synchronized audio
Underwater video presents difficult conditions for computer vision, including murky water, variable lighting and bubbles that change shape as they rise. The University of Minnesota says the team manually categorized thousands of images and used synchronized audio of regulator exhalations to support a "fuzzy labeling" process for inhalation and exhalation states.
The university reports that data collection spanned Minnesota lakes and the Caribbean Sea near Barbados. During field trials, a system called HREyes categorized breathing as below 14 breaths per minute, 14 to 20 breaths per minute, or above 20 breaths per minute, allowing the robot to communicate the observed band to a diver.
The journal article describes robotic visual HRR estimation in scuba-diving scenarios as previously unexplored. Its core contribution is a perception pipeline that turns visual evidence of exhalation into an estimate of a physiological rate.
Respiration is not a diagnosis
The available evidence supports automated respiration-rate estimation, but it does not establish a clinically validated stress diagnosis or emergency-warning system. The University of Minnesota describes changes in bubble frequency and volume as possible indicators of stress, hyperventilation or exhaustion, while the journal abstract reports comparison with human respiration-rate analysis rather than clinical outcomes.
That distinction matters for a safety-related system. Reliability would need to be established across water conditions, equipment configurations and diver populations before the estimates could support high-stakes intervention. The researchers say they next plan to combine breathing-rate data with analysis of diver movement and eventually assess more than one diver in the robot's field of view.
For robotics and ML practitioners, the research demonstrates a useful multimodal-supervision pattern: audio provides annotation support when the target visual event is ambiguous. Comparable field-robotics systems face the same challenge, where synchronized sensors can improve labels during training even when deployment relies primarily on one sensor modality.
Key Points
- 1The system estimates breaths per minute from exhaled-bubble video, extending underwater robots beyond navigation and inspection into human-state monitoring.
- 2Synchronized regulator audio supported fuzzy labels for ambiguous visual footage, demonstrating a practical multimodal-supervision method for underwater computer vision.
- 3The study compares respiration estimates with human analysis; stress diagnosis and emergency prediction require separate clinical and operational validation.
Scoring Rationale
This is a computer-vision and multimodal-labeling contribution for an operationally difficult robotics setting. Its immediate practitioner impact is specialized, but the contactless sensing approach and audio-assisted annotation workflow are relevant to field robotics researchers.
Sources
Primary source and supporting public references used for this report.
Practice with real Hotels & Lodging data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Hotels & Lodging problems
