Pith. sign in

REVIEW 4 major objections 5 minor 33 references

Multi-agent Reinforcement Learning for Robotized Coral Reef Sample Collection

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A reinforcement-learning policy trained in a game-engine simulator transferred to a physical underwater robot and completed a coral search-and-navigation run without retraining.

desk verdict A useful Unity-based RL testbed with a real BlueROV2, but the 'zero-shot sim-to-real' claim holds only for navigation dynamics under an oracle pose source, not for the perception or collection task. read the letter →

arxiv 2507.16941 v1 pith:74ZWM2S2 submitted 2025-07-22 cs.RO

Daniel Correa , Tero Kaarlela , Jose Fuentes , Paulo Padrao , Alain Duran , Leonardo Bobadilla This is my paper
classification cs.RO
keywords reinforcementlearningzero-shotsim-to-realdigitaltwinunderwaterrobotmotioncapturecoralreefconservationmulti-agentUnity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper aims to establish that a deep reinforcement-learning controller trained entirely inside a game-engine digital twin can run on a physical underwater robot without retraining, in what the authors call an effective zero-shot sim-to-real strategy. The application is coral sample collection: the robot must find a healthy coral, avoid an unhealthy coral, and bring the sample to a collection bucket. The physical test uses a real-time underwater motion-capture system to feed ground-truth positions of the robot, coral, and bucket into the simulator, and the policy then acts on the rendered scene instead of the robot's own camera image. If the approach is sound, it offers a low-cost route from simulated multi-agent reinforcement learning to deployable autonomous underwater vehicles for reef monitoring and restoration.

What carries the argument

The central mechanism is the digital-twin-to-hardware loop. A simulated replica of the test tank, the robot, coral mock-ups, and the bucket is built in Unity, and a real-time underwater motion-capture system supplies ground-truth 3D position and orientation for each object at 10 Hz. Those measured poses update the digital scene, and the trained policy observes the rendered RGB frame plus a vector of noisy sensor readings, namely surge, sway, yaw velocity, and acoustic distance, then emits surge, sway, and yaw commands that are sent to the physical BlueROV2. Because the policy's visual input is rendered from ground-truth poses rather than produced by the robot's camera, the loop tests transfer of the learned dynamics and navigation behavior while setting aside perception.

What would settle it

Run the physical test again while replacing the rendered camera image with the robot's real onboard camera feed but keeping the same motion-capture state information; if the policy can no longer find the coral or reach the bucket, the zero-shot claim applies only when perception is supplied externally.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that combining a general-purpose game engine, deep reinforcement learning, and real-time underwater motion capture delivers an effective zero-shot sim-to-real strategy for autonomous underwater vehicle search and navigation. In simulation, PPO, SAC, and IPPO agents all improve over training, with the multi-agent IPPO run accumulating roughly three times the reward of the single-agent runs. In a single physical verification trial, the trained policy commanded an eight-thruster BlueROV2 from its start point to the coral and then to the collection bucket, with trajectory deviations generally below 70 mm from an ideal path. The learned controller outputs surge, sway, and yaw velocity commands while lower-level PID control stabilizes attitude and terrain-following altitude; the gripper and camera-based recognition are listed as future work, so what transfers is the navigation policy, not the full manipulation loop.

Load-bearing premise

The test assumes that giving the policy a simulated view built from motion-capture positions is a fair stand-in for the robot actually seeing the coral with its own camera.

Editorial extensions

If this is right

  • An RL policy trained entirely in simulation can command a physical underwater vehicle on a search-and-navigation task without any fine-tuning, as long as live pose feedback is available.
  • Reducing the learned action space to surge, sway, and yaw while assigning attitude and altitude stabilization to a classical PID controller makes the navigation problem tractable without losing full 2D mobility over the seafloor.
  • Multi-agent training with independent PPO scales rewards roughly in proportion to the number of agents and objects, suggesting the same simulator can train fleets of collection vehicles.
  • The MOCAP-synchronized digital twin provides a reusable hardware-in-the-loop testbed for evaluating other underwater controllers in the same tank.
  • Replacing motion capture with an underwater positioning system that returns x-y-z coordinates would be the next step toward open-sea deployment.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The test isolates dynamics transfer, not perception transfer: because the policy sees a scene rendered from motion-capture ground truth, a reader should not conclude that the same policy would work on the robot's real camera feed.
  • A true end-to-end sim-to-real test would feed the physical camera image into the policy; until then, the authors' claim is best read as zero-shot transfer of learned navigation under externally supplied state.
  • The roughly threefold reward gain of IPPO matches the threefold increase in agents and objects, so it is evidence of independent scaling rather than emergent coordination among agents.
  • Domain randomization of the rendered scene, such as changing lighting, turbidity, and coral textures during training, would be a cheap way to test whether the transfer robustness extends to perception.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. This manuscript presents a Unity-based RL training environment for a BlueROV2 performing coral sample collection. The authors train PPO, SAC, and IPPO agents in simulation (SIL) and then validate one trained policy by using an underwater motion-capture system to write ground-truth poses of the vehicle, corals, and bucket into Unity while sending the policy's surge/sway/yaw commands to the physical BlueROV2 (HIL). Training curves show increasing accumulated reward, and a single physical run is compared with an ideal trajectory. The paper claims this constitutes an effective zero-shot sim-to-real strategy.

Significance. The proposed integration of a general-purpose game engine, deep RL, and real-time underwater MOCAP is a worthwhile engineering contribution if the claims are scoped correctly. The reward structure is hand-coded rather than fitted to a target result, and the vehicle parameters are taken from datasheets and published literature, so there is no obvious circularity. The paper is also candid about its limitations: Section IV.F states that ground-truth pose is used during HIL, and the final Discussion paragraph identifies image recognition and gripper functionality as future work. These admissions help the reader, but they directly limit the scope of the 'zero-shot sim-to-real' claim to low-level navigation under an oracle state source.

major comments (4)
  1. [IV.F, Fig. 5] The HIL validation feeds the policy a Unity-rendered RGB observation generated from MOCAP ground-truth poses, not the physical camera feed. Section IV.F states that the agent's 3D pose 'is not estimated using simulation as in training but is instead obtained using ground truth data,' and Fig. 5 shows MOCAP positions for the ROUV, corals, and bucket being written into Unity. The policy therefore sees the same observation distribution as in training, and the real perception path is outside the loop. This means the abstract's claim of 'an effective zero-shot sim-to-real strategy' is supported only for dynamics/control transfer under idealized state feedback, not for the full robotized coral sample collection pipeline. I recommend either re-running the validation with the real camera image as the policy observation (using MOCAP only for ground-truth comparison) or explicitly narrowing the title, abstract, and claims to 'zero-shot transfer of a learned navigation policy with external MOCAP state feedback.'
  2. [V, Fig. 7] The physical validation consists of a single run with no stated success criteria, no repeated trials, and no quantitative error statistics. The text says deviations from an ideal trajectory are 'usually less than 70 mm,' but Fig. 7 does not show the ideal trajectory's path definition, how the 70 mm bound was computed, or whether the run ended in a successful coral approach and bucket deposit. Add a clear task-success definition (e.g., distance thresholds to coral and bucket within a time limit), report the number of runs and per-run metrics such as RMSE, max deviation, and success/failure, and if only one run was performed, present it as a feasibility demonstration rather than as evidence of robust zero-shot transfer.
  3. [VI, final paragraph; Fig. 8] The manuscript itself notes that 'AI-image recognition and gripper functionality' are future work, and the physical setup contained no diseased coral and no physical collection/deposit action by a gripper. Consequently, the demonstrated task is AUV search and navigation in a coral-like environment, not coral sample collection. This is a load-bearing gap with respect to the paper's stated contributions; please either add a physical demonstration of the full collection loop (perception, approach, gripper collection, deposit) or revise the title, abstract, and contribution statements to describe a navigation task only.
  4. [Table I, Fig. 6] The training comparison in Fig. 6 is not apples-to-apples: IPPO runs three agents with three buckets and fifteen corals, so its larger accumulated reward is expected from more environmental features and agents. The claim in Section VI that the IPPO system 'can be successfully scaled up' is therefore not supported by the raw reward curves. Report per-agent normalized rewards, success rates, or episode-length metrics, and specify which model was selected for HIL testing and on what criterion.
minor comments (5)
  1. [Table I] In Table I, 'trail' should be 'trial' in both the caption and the text; the header 'TABLE OF INDEPENDENT VARIABLES FOR EACH EXPERIMENTAL TRAIL' contains a typo.
  2. [Fig. 7 caption] The caption says 'with an accuracy relative to ± 100mm,' but it is unclear what quantity is accurate and what it is relative to; please rephrase, for example 'coral and bucket positions are marked to within ±100 mm.'
  3. [Fig. 8] The time stamps are formatted inconsistently (0:00s, 5:03s, 10:01s, 21:12s) and the 'coral collection' frame is ambiguous because no gripper operation is described; use a single consistent time format and clarify what action is shown.
  4. [IV.C] Section IV.C says sensor noise is added 'based on hardware specifications' and can be determined empirically, but the simulated noise parameters are not reported; include the numerical values or a reference so the experiments can be reproduced.
  5. [General] The paper does not state whether code, trained models, or training configurations are available; for reproducibility, please include a link to a repository or supplementary material.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the MOCAP-based HIL validation narrows the sim-to-real scope but does not make the demonstrated transfer a tautology.

full rationale

The paper's derivation chain is: train an RL policy entirely in a Unity digital twin, select the best training model, then run a hardware-in-the-loop test in which real-time MOCAP feeds the physical vehicle's pose back into Unity so the policy observes a rendered scene and commands the physical BlueROV2. No parameter is fitted to the physical test to force its outcome; the training is self-contained and the physical test is an independent evaluation. The physical model parameters come from external datasheets and published literature, and the cited prior work is not by the present authors. The reward shaping is hand-coded and does not encode the testing trajectory. The comparison to an 'ideal trajectory' is an evaluation metric, not a fitted prediction. The only substantial limitation is that the HIL architecture uses MOCAP ground-truth positions to render the policy's observations, so the camera/perception path is not exercised in the real loop; this is a scope constraint on the 'zero-shot sim-to-real' claim, not a circular reduction. The paper explicitly states that 'the agent's 3D pose in the testing environment is not estimated using simulation as in training but is instead obtained using ground truth data,' which is transparent about the substitution. Thus, while the full coral-collection pipeline is not yet demonstrated, the paper's central result—that a simulation-trained policy can transfer to physical vehicle dynamics in this setup—is not constructed from its own conclusion.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the fidelity of the simulation model, the reward design, and the assumption that validating against MOCAP-driven digital twins is a fair zero-shot test. The vehicle model and environmental assumptions are borrowed from literature or stated without validation, and the reward shaping parameters are hand-selected. No new entities are postulated.

free parameters (3)
  • Reward +1 for collecting good coral = 1
    Hand-coded reward in the DFSM in Fig. 3 to drive collection behavior; arbitrary magnitude.
  • Reward -1 for bad coral or failed states, -0.1 for other transitions = -1, -0.1
    Hand-coded penalties in the DFSM; not fit to data but chosen ad hoc to shape policy.
  • Reward ln(1 + bucket distance) = ln(1+d)
    Hand-coded shaping term in Fig. 3; scale not justified relative to other rewards.
assumptions (5)
  • domain assumption Fossen 6-DoF vehicle model (Eq. 1) accurately describes BlueROV2 dynamics.
    Stated in Section IV.C and attributed to [30]; no validation against measured dynamics is shown.
  • domain assumption Hydrodynamic disturbances are negligible except linear currents.
    Stated in Section IV.B: 'negligible hydrodynamic disturbances due to surface waves... currents are assumed to follow a linear pattern'.
  • domain assumption Terrain-following PID and roll/pitch stabilization can be treated as reliable low-level controllers, so only surge, sway, yaw need RL.
    Stated in Section IV.C; no failure-mode analysis of the PID layer is provided.
  • ad hoc to paper The reward DFSM is a sufficient proxy for the real coral sampling goal.
    The reward structure in Fig. 3 encodes the task; the physical robot never actuates the gripper or performs vision-based detection.
  • domain assumption MOCAP ground truth can substitute for on-board perception in validation.
    Section IV.F uses MOCAP poses to drive the digital twin, so the policy's observation is rendered from ground truth, not real camera data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-agent Reinforcement Learning for Robotized Coral Reef Sample Collection." pith.science (2026). https://pith.science/paper/74ZWM2S2

@misc{pith2026250716941,
  author       = {Pith},
  title        = {Pith review of: Multi-agent Reinforcement Learning for Robotized Coral Reef Sample Collection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/74ZWM2S2}},
  note         = {Machine review of arXiv:2507.16941}
}
read the original abstract

This paper presents a reinforcement learning (RL) environment for developing an autonomous underwater robotic coral sampling agent, a crucial coral reef conservation and research task. Using software-in-the-loop (SIL) and hardware-in-the-loop (HIL), an RL-trained artificial intelligence (AI) controller is developed using a digital twin (DT) in simulation and subsequently verified in physical experiments. An underwater motion capture (MOCAP) system provides real-time 3D position and orientation feedback during verification testing for precise synchronization between the digital and physical domains. A key novelty of this approach is the combined use of a general-purpose game engine for simulation, deep RL, and real-time underwater motion capture for an effective zero-shot sim-to-real strategy.

Figures

Figures reproduced from arXiv: 2507.16941 by the authors.

Figure 1
Figure 1. Illustration of the coral sample collection task, (a) Sample col [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Materials utilized in this work (a) healthy coral mock-up, (b) [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Network architecture used in the PPO experiment. A Critic network [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: Reward model for the coral collection agent [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]
Figure 5
Figure 5. Figure 5: The testing environment HIL architecture. (a) The underwater motion capture cameras, (b) ROUV, (c) Coral mockups, (d) Sample collection [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 7
Figure 7. Figure 7: ROUV trajectory compared to an ideal trajectory. [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: The coral sample collection task: (a) starting point, (b) detection [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 24 canonical work pages

  1. [1]

    Introduction to coral reef ecosystems, threats, and solutions,

    The Coral Reef Alliance, “Introduction to coral reef ecosystems, threats, and solutions,” Coral Parks Program Education Series , 2003

    work page 2003

  2. [2]

    Microbial nitrogen removal in reef-building corals: A light-sensitive process,

    Q. Yang, J. Ling, Y . Zhang, W. Zhou, Z. Wei, J. Li, Y . Zhang, J. Dong, and P. Qian, “Microbial nitrogen removal in reef-building corals: A light-sensitive process,” Chemosphere, vol. 359, p. 142394, 2024

    work page 2024

  3. [3]

    Coral reefs and climate change: Examining two institutional approaches to managing a novel marine ecosystem,

    K. Wally, R. Williamson, and K. Dunning, “Coral reefs and climate change: Examining two institutional approaches to managing a novel marine ecosystem,” Marine Policy, vol. 163, p. 106107, 2024

    work page 2024

  4. [4]

    Func- tion and stability of mesophotic coral reefs,

    M. Slattery, M. Lesser, L. Rocha, H. Spalding, and T. Smith, “Func- tion and stability of mesophotic coral reefs,” Trends in Ecology & Evolution, vol. 39, pp. 585–598, 2024

    work page 2024

  5. [5]

    From marketready ROVs to lowcost AUVs,

    J. Scharff Willners, I. Carlucho, S. Katagiri, C. Lemoine, J. Roe, D. Stephens, T. Łuczynski, S. Xu, Y . Carreno, E. Pairet, C. Barbalata, Y . Petillot, and S. Wang, “From marketready ROVs to lowcost AUVs,” in OCEANS 2021: San Diego – Porto , 09 2021, pp. 1–7

    work page 2021

  6. [7]

    Dynamic water physics 2,

    N. P. World, “Dynamic water physics 2,” https://dynamicwaterphysics.com/doku.php/index, 2025, accessed 28 Febuary 2025

    work page 2025

  7. [8]

    Uncrewed maritime sys- tems for coral reef conservation,

    M. Dunbabin, J. Manley, and P. L. Harrison, “Uncrewed maritime sys- tems for coral reef conservation,” in Global Oceans 2020: Singapore – U.S. Gulf Coast , 2020, pp. 1–6

    work page 2020

  8. [9]

    Motion control of autonomous underwater vehicle based on physics-informed offline reinforcement learning,

    X. Li, L. Geng, K. Liu, Y . Zhao, and W. Du, “Motion control of autonomous underwater vehicle based on physics-informed offline reinforcement learning,” Ocean Engineering , vol. 313, p. 119432, 2024

    work page 2024

Show all 33 references
  1. [10]

    The NDRE-AUV flight control system,

    B. Jalving, “The NDRE-AUV flight control system,” IEEE Journal of Oceanic Engineering , vol. 19, no. 4, pp. 497–501, 1994

    1994

  2. [11]

    Trajectory tracking control of AUVs via adaptive fast nonsingular integral terminal sliding mode control,

    L. Qiao and W. Zhang, “Trajectory tracking control of AUVs via adaptive fast nonsingular integral terminal sliding mode control,”IEEE Transactions on Industrial Informatics , vol. 16, no. 2, pp. 1248–1258, 2020

    2020

  3. [12]

    Trajectory tracking control of an autonomous underwater vehicle using Lyapunov-based model predic- tive control,

    C. Shen, Y . Shi, and B. Buckham, “Trajectory tracking control of an autonomous underwater vehicle using Lyapunov-based model predic- tive control,” IEEE Transactions on Industrial Electronics , vol. 65, no. 7, pp. 5796–5805, 2018

    2018

  4. [13]

    Angler: An autonomy framework for intervention tasks with lightweight underwater vehicle manipulator systems,

    E. Palmer, C. Holm, and G. Hollinger, “Angler: An autonomy framework for intervention tasks with lightweight underwater vehicle manipulator systems,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) , 2024, pp. 6126–6132

    2024

  5. [14]

    Disturbance observer-based fuzzy adapted s-surface controller for spatial trajectory tracking of autonomous underwater vehicle,

    G. V . Lakhekar, L. M. Waghmare, and R. G. Roy, “Disturbance observer-based fuzzy adapted s-surface controller for spatial trajectory tracking of autonomous underwater vehicle,” IEEE Transactions on Intelligent V ehicles, vol. 4, no. 4, pp. 622–636, 2019

    2019

  6. [15]

    Adaptive neural network control of a marine vessel with constraints using the asymmetric barrier Lyapunov function,

    W. He, Z. Yin, and C. Sun, “Adaptive neural network control of a marine vessel with constraints using the asymmetric barrier Lyapunov function,” IEEE Transactions on Cybernetics, vol. 47, no. 7, pp. 1641– 1651, 2017

    2017

  7. [16]

    A general motion control architecture for an autonomous underwater vehicle with actuator faults and unknown disturbances through deep reinforcement learning,

    F. Huang, J. Xu, L. Yin, D. Wu, Y . Cui, Z. Yan, and T. Chen, “A general motion control architecture for an autonomous underwater vehicle with actuator faults and unknown disturbances through deep reinforcement learning,” Ocean Engineering , vol. 263, p. 112424, 2022

    2022

  8. [17]

    Learning to swim: Reinforcement learning for 6-DoF control of thruster-driven autonomous underwater vehicles,

    L. Cai, K. Chang, and Y . Girdhar, “Learning to swim: Reinforcement learning for 6-DoF control of thruster-driven autonomous underwater vehicles,” 2024. [Online]. Available: https://arxiv.org/abs/2410.00120

    2024 arXiv

  9. [18]

    Design and use paradigms for Gazebo, an open-source multi-robot simulator,

    N. Koenig and A. Howard, “Design and use paradigms for Gazebo, an open-source multi-robot simulator,” in 2004 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (IEEE Cat. No.04CH37566), vol. 3, 2004, pp. 2149–2154 vol.3

    2004

  10. [19]

    Mujoco: A physics engine for model-based control,

    E. Todorov, T. Erez, and Y . Tassa, “Mujoco: A physics engine for model-based control,” in 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems , 2012, pp. 5026–5033

    2012

  11. [20]

    Dave aquatic virtual environment: Toward a general underwater robotics simulator,

    M. M. Zhang, W.-S. Choi, J. Herman, D. Davis, C. V ogt, M. McCarrin, Y . Vijay, D. Dutia, W. Lew, S. Peters, and B. Bingham, “Dave aquatic virtual environment: Toward a general underwater robotics simulator,”

  12. [21]

    Overcoming the sim- to-real gap in autonomous robots,

    P. Trentsios, M. Wolf, and D. Gerhard, “Overcoming the sim- to-real gap in autonomous robots,” Procedia CIRP , vol. 109, pp. 287–292, 2022, 32nd CIRP Design Conference (CIRP Design

    2022

  13. [22]

    BlueROV2,

    Bluerobotics Inc., “BlueROV2,” https://bluerobotics.com/store/rov/bluerov2/, 2025, accessed 2 February 2025

    2025

  14. [23]

    6-DoF modelling and control of a remotely operated vehicle,

    C.-J. Wu, “6-DoF modelling and control of a remotely operated vehicle,” Master’s thesis, Flinders University, 2018

    2018

  15. [24]

    [Online]

    - Design in a changing world. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S2212827122007004

  16. [25]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” 2017. [Online]. Available: https://arxiv.org/abs/1707.06347

    2017 arXiv

  17. [26]

    Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,” 2018. [Online]. Available: https://arxiv.org/abs/1801.01290

    2018 arXiv

  18. [27]

    Bluerov2 datasheet,

    B. R. Inc, “Bluerov2 datasheet,” 2020, accessed 1 March 2025. [Online]. Available: https://bluerobotics.com/wp \-content/uploads/ 2020/02/br\ bluerov2\ datasheet\ rev6.pdf

    2020

  19. [28]

    The surprising effectiveness of ppo in cooperative, multi-agent games,

    C. Yu, A. Velu, E. Vinitsky, J. Gao, Y . Wang, A. Bayen, and Y . Wu, “The surprising effectiveness of ppo in cooperative, multi-agent games,” 2022. [Online]. Available: https://arxiv.org/abs/2103.01955

    2022 arXiv

  20. [29]

    A survey and comparative evaluation of actor-critic methods in process control,

    D. Dutta and S. R. Upreti, “A survey and comparative evaluation of actor-critic methods in process control,” The Canadian Journal of Chemical Engineering, vol. 100, no. 9, pp. 2028–2056, 2022. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/cjce.24508

    2022 doi

  21. [31]

    Available: https://arxiv.org/abs/2011.09533

    [Online]. Available: https://arxiv.org/abs/2011.09533

    2011 arXiv

  22. [34]

    Handbook of marine craft hydrodynamics and motion control [bookshelf],

    L. McCue, “Handbook of marine craft hydrodynamics and motion control [bookshelf],” IEEE Control Systems Magazine , vol. 36, no. 1, pp. 78–79, 2016

    2016

  23. [35]

    Water linked BlueROV2 integration,

    Bluerobotics Inc., “Water linked BlueROV2 integration,” https://bluerobotics.com/learn/water-linked-bluerov2-integration- r2-integration-kit/, 2025, accessed 7 March 2025

    2025

  24. [2020]

    Available: https://arxiv.org/pdf/1809.02627.pdf

    [Online]. Available: https://arxiv.org/pdf/1809.02627.pdf

    arXiv

  25. [2022]

    Available: https://arxiv.org/abs/2209.02862

    [Online]. Available: https://arxiv.org/abs/2209.02862

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.