Towards Physical Underwater Robotic Assistance for Scuba Diver Movement in Confined Spaces

arXiv:2610.01906 · cs.RO · Submitted 2026-10-01 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "Towards Physical Underwater Robotic Assistance for Scuba Diver Movement in Confined Spaces".

Dev: A novel wearable robotic system, RADMCS, is introduced to assist scuba divers in maintaining safe standoff distances from subsea structures in confined or hazardous underwater environments.

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: So, we're diving into the paper "Towards Physical Underwater Robotic Assistance for Scuba Diver Movement in Confined Spaces," and it sounds like they’ve put forward a wearable system called RADMCS that helps divers keep a safe distance from underwater structures. Rosa here, I'm curious if this kind of external assistance actually works well outside of a controlled lab setting, and for how long can we expect this to be reliable in real-world diving situations?

Dev: That’s a fair question, Rosa; as the controls engineer, my first thought is always about the loop rate and latency when we move from simulation to actual underwater application. If the feedback loop has too much lag or jitter, it could easily become dangerous for a diver in a real scenario.

Taro: From an autonomy research standpoint, I'm interested in what happens when things go wrong; if the environment misbehaves unexpectedly, does RADMCS have any built-in mechanisms to handle that uncertainty?

Rosa: Exactly, Taro; the core of this paper introduces RADMCS as a wearable robot that uses perception from monocular depth estimation and thruster feedback to give divers haptic guidance. The main goal is to assist divers in maintaining a fixed distance from objects like coral reefs or structures without needing complex exoskeletons.

Dev: I see how the system works conceptually, using thrust actuation to communicate directional information through physical sensation; but I need to know how stable that distance estimate is when the environment isn't perfectly clear or when currents get strong.

Taro: That stability is key, because if the perception fails in a complex cave system, we need something that can keep guiding the diver safely rather than just giving noisy signals.

Rosa: The paper details their control methodology, which uses a linear control algorithm with saturation to limit thrust outputs and employs an exponential moving average filter to smooth out distance estimates. They also set up a maximum likely distance heuristic, DMAX, and a deadzone called δdeadzone where the robot provides no feedback if the diver is already moving correctly.

Dev: The use of an EMA filter helps smooth things out, but I have to ask about the stability when that filter is smoothing out real disturbances from ocean currents; does that filtering introduce any undesirable phase lag in our control loop?

Taro: That’s a critical point for any system operating in fluid dynamics; we need to know if smoothing the input data compromises our ability to react quickly enough when an obstacle suddenly appears.

Rosa: The experimental results show that relatively low thrust values, approximately ten percent of maximum, were enough for the robot to guide a human’s movement through physical sensation across four participants and all three configurations in both open water and closed-water environments.

Dev: Ten percent of maximum thrust is quite low; that suggests a very efficient way to provide guidance without overwhelming the diver with unnecessary force, but I wonder if that low threshold is robust enough when comparing clear water to more murky conditions.

Taro: That robustness across different visual qualities is what we need to test rigorously; if it works well in clear water but gets confused by turbidity, then its practical use outside of ideal conditions is limited.

Rosa: They also tested the form, fit, and function in open water field experiments where one participant mentioned a device moved out of place during testing and caused a pinch point at the diver’s head due to forward thrust coupling.

Dev: That feedback about physical interference during testing highlights an integration challenge that we need to address if we're going to deploy this commercially; how much does that physical interaction impact the actual control performance?

Taro: It shows that even if the guidance mechanism is mathematically sound, the physical coupling between the robot and the diver’s equipment can introduce unexpected dynamic issues during movement.

Rosa: The paper concludes by summarizing these findings, suggesting RADMCS shows promise for foundational work in robotic-assisted navigation underwater and establishing a platform for physical human-robot interaction.

Dev: So, to wrap up on this paper, it seems the authors have shown a feasible way to use low-thrust haptic feedback for lateral control in confined spaces using current perception techniques. The implication is that we can move beyond just depth control into true movement assistance.

Taro: I think the real impact here is establishing a baseline for how physical perturbation can be used to guide human operators, which could open up avenues for more complex autonomy in these environments later on.

Rosa: Indeed, and while the paper is solid in demonstrating the concept across different settings, future work will focus on things like deep learning-based depth estimation and exploring control strategies for more complex maneuvers in simulated confined spaces.

Dev: I'm looking forward to seeing how they tackle those higher-level behaviors; right now, I’m focused on making sure the latency of that low-thrust command is as tight as possible.

Taro: I hope those future explorations lead to a system that can truly handle unpredictable environmental misbehaviors without requiring constant manual intervention from the diver.

Rosa: That’s what we want to see, and it sounds like RADMCS provides a really tangible starting point for physical interaction research underwater. We'll keep an eye on their next steps as they push this platform forward.

The paper's summary: Rosa: So, to recap, the RADMCS system is a wearable robot designed to help scuba divers maintain safe distances from underwater structures by giving them physical feedback when they get too close to walls or objects, using depth estimation and thruster control.

Dev: That's right; basically, it uses monocular depth sensing combined with force-feedback from submersible thrusters to guide the diver through physical sensations that tell them which way to go.

Taro: I’m really interested in the practical implications of this level of assistance; if a diver is navigating a tight cave system, having that haptic cue might make a huge difference in avoiding an accident.

Rosa: Exactly, Taro; imagine exploring a coral reef or navigating a dark cave system without bumping into something dangerous because you get that subtle physical nudge telling you to back off.

Dev: From my side of things, the methodology shows they use linear control with saturation and an EMA filter to smooth out those distance estimates before translating them into PWM signals for the thrusters.

Taro: That smoothing is interesting; I wonder if that filtering compromises the system's ability to react instantly when a sudden current pushes a diver off course; does it introduce any dangerous lag?

Dev: That’s a valid concern, Taro; we have to look closely at how that EMA filter affects the loop rate and latency, because in real-time control, even small delays can cause instability.

Rosa: The experimental results show that even with those smoothing techniques, they found that relatively low thrust values—about ten percent of maximum—were enough for the robot to provide perceptible guidance across all three test configurations.

Taro: Ten percent of maximum thrust is quite low; I’m curious if that level of intervention is sufficient when the visual data quality starts getting really bad, like in murky water where depth estimation gets fuzzy.

Dev: That’s exactly where I see the weakness; if the perception input degrades significantly, those low-thrust commands might become less reliable for a diver who needs precise guidance.

Rosa: The researchers did conduct tests in both clear open water and more challenging ocean environments, and they found that when things get noisy or distorted visually, the robot still tightly couples to the diver’s physical sensing capabilities.

Taro: That suggests that even if the visual input isn't perfect, the system can adapt by relying more on direct physical feedback from the diver’s movement itself.

Dev: So it seems like RADMCS is built to be somewhat robust against minor visual noise, but we still need to figure out how to make that transition seamless when things get really unpredictable.

Rosa: Right, and the paper concludes by saying this work lays a foundation for physical human-robot interaction underwater navigation.

Taro: That foundational aspect is what excites me; establishing a platform for physical guidance could open up so much more complex autonomous behaviors in these environments down the line.

Dev: I agree; having that physical interface makes the control problem inherently more tangible, which is useful for testing how we model human-robot dynamics.

Rosa: And looking ahead, the authors clearly point toward using deep learning for depth estimation and developing better models for complex behaviors in confined spaces as their next big steps.

Taro: That’s where things get really interesting; if they can integrate a system like that with learned policies, we could see assistance that's proactive rather than just reactive to distance errors.

The paper's improvements: Rosa: So, to summarize, the paper doesn't just stop at describing what they built; they lay out several ways to make RADMCS even better for those challenging underwater conditions we talked about earlier.

Dev: Exactly; they suggest moving beyond their current linear control algorithm and swapping it out for a Reinforcement Learning based controller, specifically Proximal Policy Optimization, to handle the non-linear dynamics of human swimming.

Taro: That makes sense; if we can learn an optimal policy that accounts for those unpredictable disturbances in currents, it could really help with proactive guidance instead of just reacting to errors.

Rosa: Plus, they propose using a Deep Neural Network for depth estimation instead of relying only on classical PnP methods, which should make the system much more robust against underwater distortion and poor lighting.

Dev: I think integrating that DNN would significantly improve the stability of the distance estimate, even when visual features are sparse or moving quickly in turbulent water; it should help mitigate those unphysical jumps we were worried about.

Taro: If you can build a better depth estimator, then we could potentially move toward a system that learns to anticipate needs and issues corrective thrust commands long before the diver gets into danger.

Rosa: They also suggest developing a more sophisticated state-space model within the RL agent to help it predict how the diver will move next, which would be a big step in understanding human dynamics.

Dev: A predictive model sounds promising for reducing latency; if the AI can anticipate the need for thrust adjustments based on predicted movement, we could optimize that control loop rate dramatically.

Taro: That ties back to my earlier point about anticipating needs; it moves the system from being a reactive tool to something that can be more intelligently pre-emptive in complex scenarios.

Rosa: They also talk about creating a generative model, like a GAN, to reconstruct high-fidelity three dee maps from sparse camera inputs, which should improve how accurate the long-term distance estimation becomes.

Dev: A persistent map reconstruction would certainly give the system better context over time, reducing the reliance on immediate visual cues and making those low-thrust haptic cues more meaningful.

Taro: If we combine that with a learned policy for guidance, imagine a scenario where the robot can use that three dee map to navigate around submerged obstacles autonomously.

Rosa: That’s exactly what I mean; it transitions RADMCS from being just a distance maintainer to an intelligent navigational co-pilot in those complex cave environments we discussed.

Dev: The implication is that the next generation of this system could move toward truly autonomous navigation assistance, not just simple haptic guidance.

Conclusion: Rosa: So, we've covered quite a bit on the paper "Towards Physical Underwater Robotic Assistance for Scuba Diver Movement in Confined Spaces," and to recap, they’ve shown how a wearable system can use depth perception and thruster feedback to provide low-thrust haptic guidance to divers.

Dev: That’s right; the core idea is using physical movement cues delivered through thrust actuation to help divers maintain safe standoff distances from underwater structures.

Taro: The implications for autonomous navigation in confined spaces are quite significant; if this level of assistance becomes a standard tool, it could fundamentally alter how we approach remote exploration and research underwater.

Rosa: I agree; it really demonstrates the potential for physical human-robot interaction in a way that feels intuitive, moving beyond just screen-based telemetry.

Dev: From my end, the control aspect is what makes this interesting; getting that low-level control loop tight enough to handle those thruster PWM signals reliably under dynamic conditions is a real engineering challenge they tackled.

Taro: It shows that even with relatively simple physical feedback mechanisms, we can achieve meaningful assistance when you focus on the right perception inputs.

Rosa: And looking at the future work they mentioned, focusing on deep learning for depth estimation and more sophisticated control policies really points toward making this a much smarter system down the line.

Dev: I'm keen to see how they refine those control strategies; moving towards adaptive deadzones and better current prediction will be crucial for real-world deployment.

Taro: I hope we see that shift toward proactive guidance; that’s where the real autonomy is, not just reactive distance correction.

Rosa: Overall, "Towards Physical Underwater Robotic Assistance for Scuba Diver Movement in Confined Spaces" provides a really solid foundation for exploring how physical interaction can be used to guide human operators in hazardous environments.

Dev: It’s a great piece of work that bridges the gap between perception and practical control implementation in an underwater context.

Taro: I think it opens up avenues for much more advanced, adaptive autonomous systems in challenging physical spaces.

Rosa: And that wraps up our discussion on RADMCS; I think this paper shows we’re getting closer to a wearable system that can truly assist divers in complex underwater scenarios without needing bulky gear.

Dev: Indeed, it’s an interesting blend of perception and actuation, and I’m curious to see how their future work addresses those latency issues as they move toward more complex behaviors.

Taro: I just think the potential for physical guidance is huge; we should be watching this space for systems that can use these haptic cues to guide human operators in even more extreme environments.

Rosa: Absolutely, and I'm really excited about seeing how they evolve this platform into something truly capable of navigating those intricate cave systems we talked about earlier.

Demetrious T. Kutzke, Junaed Sattar

cs.RO

Submitted: 2026-10-01

Updated: 2026-10-01

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 69/100

The gist: A novel wearable robotic system, RADMCS, is introduced to assist scuba divers in maintaining safe standoff distances from subsea structures in confined or hazardous underwater environments.

Key concepts

RADMCS: A wearable robotic system designed for scuba divers that mounts onto their tank. It uses two thrusters to provide directional guidance and force feedback, helping the diver stay at a preset distance from objects.
Haptic Feedback Mechanism: The robot communicates movement instructions by controlling the thrust of its thrusters using pulse-width modulation (PWM). This physical sensation allows the diver to feel the robot guiding them in a specific direction, such as turning left or right.
Depth Estimation and Control: The system uses monocular depth estimation to estimate distances. A linear control algorithm adjusts thruster outputs based on the error between the desired distance and the estimated distance, smoothing these estimates with an exponential moving average filter.

Terminology

Institutional Review Board — 3× in this paper
Calibration — 2× in this paper · explained in 7 episodes
Core — 2× in this paper · explained in 2 episodes
Exponential Moving Average (EMA) — 2× in this paper · explained in 2 episodes
Robotic Assisted Diver Movement in Confined Spaces — 2× in this paper
3D Gaussian Splatting — 1× in this paper · explained in 2 episodes
ACM Transactions on Applied Perception — 1× in this paper
Entanglement — 1× in this paper · explained in 3 episodes
Experimental Results and Threshold Sensitivity — 1× in this paper
Fidelity — 1× in this paper · explained in 2 episodes
Gaussian Splatting — 1× in this paper · explained in 4 episodes
Generative Model — 1× in this paper · explained in 2 episodes
Reinforcement Learning — 1× in this paper · explained in 4 episodes
Robustness — 1× in this paper · explained in 5 episodes

Summary

A novel wearable robotic system, RADMCS, is introduced to assist scuba divers in maintaining safe standoff distances from subsea structures in confined or hazardous underwater environments. This work addresses the need for lateral control during tasks like coral reef exploration or navigating cave systems by providing thruster-actuated directional guidance to the diver through physical haptic feedback.

The gist

RADMCS is a wearable robot that assists divers in maintaining a preset standoff distance from encompassing walls or objects by providing force-feedback to the diver when they get too close to such objects, leveraging perception techniques in monocular depth estimation and force-feedback from submersible thrusters to provide haptic feedback.

System Design and Integration

The RADMCS robot is designed for less invasive integration with existing scuba diving technology. It mounts onto the diver’s scuba tank using strapping mechanisms already employed by open-water scuba divers, eliminating the need for complex exoskeleton style suits or augmented swimming solutions that require additional training or instrumentation. The external assembly consists of a single cylindrical acrylic shell, depth rated to 100 m, and is affixed to the diver’s scuba tank using 3D printed parts that match the curvature of a common 'aluminum 80' scuba diving cylinder. Actuation is achieved using two thrusters (Blue Robotics T200 [27]), which can be configured longitudinally or transversely.

Haptic Feedback Mechanism

Robotic feedback is provided by controlling thrust actuation. Each thruster is controlled by a pulse-width modulation (PWM) signal, with specific ranges for full forward thrust and full reverse thrust. The robot indicates the required movement via specific PWM values: Turn left (yaw CCW about the ˆz-direction) T1 reverse and T2 forward, or Turn right (yaw CW about the ˆz-direction) T1 forward and T2 reverse. This mechanism allows the robot to communicate navigational information through the physical sensation of the robot moving the diver in the intended direction.

Visual Control Methodology

The control algorithm employs a linear control algorithm with saturation to limit thruster outputs after a certain distance. The goal is to produce thruster PWM values based on the calculated distance error, where δ = d′ − d (the difference between setpoint distance and current filtered distance estimate). To handle instabilities in depth estimation, an exponential moving average (EMA) filter is implemented to smooth the estimates, and a maximum likely distance heuristic DMAX prevents filter updates if the difference between estimates exceeds this threshold. A control deadzone δdeadzone is also created where the robot would not provide haptic feedback, because the diver was effectively traversing without the need for movement assistance.

Experimental Results and Threshold Sensitivity (TS)

The researchers conducted systematic testing of threshold sensitivity (TS) with eight human scuba diver participants in both closed-water and ocean environments. They demonstrated that relatively low thrust values (≈ 10% of maximum) allow robotic direction of a human’s movement using the physical sensation of the robot’s guidance. Specifically, across four participants and all three configurations, the averaged ascending and descending detection thresholds corresponded to approximately 10% of the maximum available thrust. Furthermore, in simulated wall-following experiments in a closed-water facility, results showed that the RADMCS robot can be used for movement assistance when there is a stable and reliable distance estimate. In open-water tests, the findings indicate that "either the divers felt external disturbances as robotic feedback and could not differentiate between robotic directed force-feedback and subtle environmental factors, or the robot tightly couples to the physical sensing capabilities of the human."

Form, Fit, and Function Testing

Qualitative experiments in open water evaluated form, fit, and function. One participant reported that the device moved out of place during the experiment, creating a painful pinch point at the diver’s head due to forward thrust coupling into equipment. However, the diver reported they could rely on the device for swimming assistance without the need for supplemental movement swimming. The system is demonstrated to be self-contained and does not interfere with the diver’s natural body movements beyond the perturbative feedback provided to the diver.

Limitations and Future Work

The study was limited by logistical complexities, restricting testing to eight participants. Future work is warranted to explore aspects such as deep learning-based depth estimation, e.g., the work of [34] for underwater three-dimensional reconstruction, and to refine control strategies for more complex behaviors like simulated confined-space traversal. The paper concludes that RADMCS shows promise for foundational work in robotic-assisted navigation underwater and establishing a platform for "physical underwater human-robot interaction.

Improvements for AI systems

As a fastidious and diligent researcher, I have analyzed the RADMCS system described in this paper. The core innovation lies in providing a low-thrust, thruster-actuated haptic feedback mechanism to guide a human diver based on real-time monocular depth estimation from cameras.

Here are specific improvements for AI systems derived from or inspired by this research, and what those improved systems could achieve:

The current control algorithm relies on a linear proportional controller with saturation, using an Exponential Moving Average (EMA) filter to smooth distance estimates and a fixed deadzone.
The system uses classical PnP for distance estimation and is vulnerable to underwater distortion, lighting variations, and autofocus instabilities, which can cause unphysical jumps in feedback.

Improvements:

Replace the linear control algorithm with a Reinforcement Learning (RL) based controller (e.g., Proximal Policy Optimization - PPO).
Integrate a Deep Neural Network (DNN) for distance estimation instead of relying solely on classical PnP, utilizing features extracted directly from monocular images to improve robustness against distortion and occlusion.
Implement a more sophisticated state-space model within the RL agent to better predict the diver's movement dynamics and environmental disturbances (currents), allowing the robot to anticipate needs rather than merely reacting to errors.
Develop an adaptive deadzone mechanism where the threshold for haptic feedback dynamically adjusts based on real-time environmental conditions (e.g., detected current speed or turbidity).

Improved AI System Capabilities:

The improved system would enable a diver to maintain a standoff distance with significantly higher precision and robustness in dynamic, unpredictable environments like coral reef canyons or cave systems where camera input quality is inherently poor.
It could proactively guide the diver through complex, multi-stage maneuvers (e.g., navigating around submerged obstacles) by learning optimal control policies that account for the non-linear relationship between thrust perturbations and human swimming dynamics.
The system could perform autonomous obstacle avoidance guidance where, based on predicted trajectory and perceived distance to a structure (using its improved perception), it issues corrective thrust commands long before the diver enters a dangerous proximity zone, effectively acting as an intelligent, low-level navigational co-pilot.

The paper notes that future work could explore deep learning-based depth estimation (e.g., 3D Gaussian Splatting) and model-based control dynamics incorporating human feedback into the robot's model.

Develop a Generative Adversarial Network (GAN) or similar generative model to reconstruct a high-fidelity, persistent 3D map of the environment from sparse camera inputs, improving long-term distance estimation accuracy.
Implement a Human State Estimation module using sensor fusion (e.g., IMU data from the diver's suit/body sensors) coupled with the robot's perception to build a high-fidelity model of the human operator's current state, compensating for occlusions and movement noise.
The system could achieve reliable distance maintenance in environments where visual features are sparse or rapidly changing (e.g., turbulent water), as it would rely less on immediate visual cues and more on learned environmental context.
It could provide personalized guidance tailored to the individual diver's swimming style and fatigue level, adjusting the required thrust sensitivity dynamically to minimize unnecessary physical exertion while maintaining safety margins.

The paper demonstrates that even low thrust (10% of maximum) is sufficient for perceptible directional cues, suggesting a high degree of efficiency in human-robot interaction.

Apply Bayesian inference to quantify the perceptibility threshold across different environmental conditions (e.g., water clarity, ambient noise), allowing the AI to dynamically adjust the required feedback intensity (thrust level) based on perceived sensory quality rather than a fixed percentage of maximum thrust.
The system could operate optimally in varied conditions—from clear open water to murky cave entrances—by intelligently scaling its intervention force, ensuring the guidance remains perceptible and effective regardless of environmental noise or visual degradation.

Abstract

Scuba divers are taught to control their depth to avoid rapid ascents and descents, which could result in serious injuries such as gas embolisms and barotrauma. However, many underwater tasks necessitate lateral control, maintaining distance between subsea structures such as coral reefs, submerged drilling instrumentation, or unexploded ordnance. In this work, we discuss a first-of-its-kind wearable robotic solution providing thruster-actuated directional guidance to a diver, as distinct from prior propulsive-assistance exoskeletons. We introduce ``Robotic Assisted Diver Movement in Confined Spaces'' (RADMCS), a wearable robot that assists divers in maintaining a fixed distance from subsea structures by leveraging perception techniques in monocular depth estimation and force-feedback from submersible thrusters to provide haptic feedback. Its small and compact form factor creates a foundational platform that could be expanded to include more sophisticated control and navigation behaviors. We present results from Institutional Review Board (IRB) in-water studies with eight human scuba diver participants on threshold sensitivity tests in both a closed-water swimming facility and ocean environments; distance-maintaining experiments in a closed-water facility; and form, fit, and function testing in the ocean. We demonstrate that relatively low thrust values (10 percent of maximum) allow robotic direction of a human's movement using the physical sensation of the robot's guidance.