<- Back to news

RADMCS steers scuba divers with low-thrust haptic cues

A tank-mounted wearable robot uses two underwater thrusters as directional haptic cues, but its wall-following behavior depends on a distance estimate that failed in one documented run.

Preprint · arXiv v1 · IRB in-water study with eight participants across different subsets · CC BY 4.0 manuscript · no code or data release verifiedSource date: Read the primary source ↗
underwater hapticswearable roboticsforce feedbackhuman-robot interaction
Diagram showing two tank-mounted underwater thrusters turning a diver toward or away from a wall according to camera distance error.
Original RoboSkin.ai schematic of the RADMCS distance-to-thrust feedback loop. It is explanatory artwork, not an experiment photograph.

In an October 1, 2026 preprint, University of Minnesota researchers Demetrious T. Kutzke and Junaed Sattar describe RADMCS: a wearable underwater robot that straps to a scuba tank and uses two thrusters to cue a diver to turn toward or away from a surface. An IRB-approved study involved eight participants across several pool and ocean tests. The most stable quantitative finding is perceptual: approximately 10% of maximum thrust was enough to produce detectable motion in the tested configurations. Paper and version record.

Explore tactile datasets for the broader comparison.

Key takeaways

  • RADMCS is a haptic navigation aid, not a propulsive exoskeleton: its thrusters perturb the diver so the diver responds with their own motion.
  • Pool threshold tests use four participants split across three thruster configurations; ocean threshold tests use two participants in two configurations.
  • A seven-participant pool wall-following study exposes a critical dependency: when visual distance estimation is lost, the last thrust command can persist.

From camera distance to physical cue

RADMCS carries two Blue Robotics T200 thrusters in either longitudinal or transverse arrangements with 225 mm or 325 mm spacing. A camera estimates distance to a target surface. After exponential filtering and a dead zone, the controller maps distance error to differential thruster commands. The tested setup uses a practical sensing range of 1.5 m and clips larger errors.

The mechanism deliberately closes the loop through a person. A positive or negative thrust moment asks the diver to yaw; it does not autonomously translate the diver to a coordinate. That distinction matters for haptic interfaces: the relevant output is a perceivable, directional body cue, not robot trajectory accuracy.

What the participant study shows

Eight divers participated overall, but not every person completed every test. In the pool, two participants tested the 225 mm longitudinal and transverse layouts, while two different participants tested the 325 mm transverse layout. Facility time prevented collection for the 325 mm longitudinal configuration. Participants were neutrally buoyant at 1.8 m depth and reported when they first perceived the force during ascending and descending PWM sweeps.

Open-water threshold testing used two participants, 5 m from shore at 1.5 m depth, with both 325 mm layouts. Despite 0.25–0.5 m wave-height variation, the reported perceptible level again clustered around 10% of maximum thrust. The authors explicitly leave two explanations unresolved: divers may have confused environmental motion with robot feedback, or the device may couple strongly enough that directional cues remain distinct.

The closed-water distance-maintaining test used seven participants wearing blackout masks beside a 7 m tarp carrying AprilGrid targets at 1 m intervals. In one illustrated run, participant P2 completed the course in 89.89 s while responding to left and right cues. In the P8 example, distance estimates disappeared at about 12 s and never recovered; the controller retained a high left-turn command until the 43.28 s run ended. The paper therefore supports guided movement when the estimate is stable, not robust autonomous wall following under arbitrary underwater vision.

RoboSkin analysis

RADMCS is a useful counterexample to the idea that more force always makes a haptic cue better. A low, detectable threshold can conserve power and reduce intrusive motion, but perception reliability becomes the dominant safety constraint. The controller needs an explicit stale-signal policy before the device can be treated as a dependable wearable robotics platform.

The sample structure also matters. “Eight participants” is the union of the study, not the denominator for every result. Threshold findings, wall-following examples and a form-fit test come from different subsets. Reporting those boundaries keeps a promising prototype from sounding like a validated navigation product.

Limitations and availability

This is an arXiv v1 preprint, not an independently validated dive system, and RoboSkin.ai did not test it. The paper reports a qualitative ocean form-fit trial with one participant at up to 80% thrust. The apparatus shifted, moved regulator hoses forward and created a painful pinch point near the participant's head. Those observations are design constraints, not cosmetic issues.

Distance sensing used structured AprilGrid targets rather than natural reef geometry for the pool control study. The tarp warped in circulation currents, autofocus affected measurements, and one documented run preserved an unsafe stale command after tracking loss. The manuscript is available under CC BY 4.0. The November 4 paper review did not verify an official code repository, CAD package, dataset or software license through the reviewed sources.

Separate cue perception from supervised navigation

The approximate 10% threshold uses the manufacturer's T200 thrust/PWM relationship at the stated supply voltage. It is not a measurement of force transmitted to the diver's skin, nor an accuracy score for identifying cue direction. Mounting, buoyancy and environmental motion affect the body response. The full protocol also states that a safety diver accompanied the blackout-mask pool runs and intervened to prevent lateral deviation beyond one meter. A successful illustrated course therefore includes supervision; it is not an unattended navigation result.

Evidence layerReceiving-team review question
Distance observationWas the visual estimate valid and current when each command was issued?
Physical cueWas the quantity a commanded thrust, measured body load or participant-reported sensation?
Human responseDid the participant identify the intended direction, and how was that response recorded?
Navigation outcomeWhich runs needed intervention, and were tracking loss and incomplete runs counted?
Wearable fitDid mounting movement, hose interference or discomfort change the interpretation?

The one-person ocean fit exercise used forward thrust for swimming assistance; it was not the seven-person pool wall-following protocol. Its much higher thrust setting cannot be combined with the perception threshold into one operating recommendation.

For a future evaluation, record observation age, command expiry, intervention and task outcome separately. A candidate stale-signal policy would need its own supervised validation; this paper does not establish a safe timeout or recovery rule. These are research-review questions, not instructions for operating a diving device. The tactile AI evidence chain distinguishes haptic output to a person from learned tactile perception and robot feedback.

*Source review updated November 4, 2026. Original publication dates are retained. Research results are author-reported; RoboSkin has not run these systems or reproduced their experiments.*