Pith. sign in

REVIEW 5 major objections 6 minor 64 references

AquaChat: An LLM-Guided ROV Framework for Adaptive Inspection of Aquaculture Net Pens

T0 review · 5 major / 6 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 AquaChat claims that a large language model can convert an operator's plain-language instruction into an executable ROV inspection plan, giving aquaculture net-pen inspection greater flexibility and accuracy than pre-programmed or manual…

desk verdict AquaChat is a clean application of the LLM-hierarchical-planner pattern to aquaculture ROV inspection, with reasonable component-level control tests, but the headline claims of adaptive end-to-end inspection are unsupported because no experiment connects the LLM plan to the executed trajectory. read the letter →

arxiv 2507.16841 v1 pith:FEFAU6RS submitted 2025-07-19 cs.RO

Waseem Akram , Muhayy Ud Din , Abdelhaleem Saad , Irfan Hussain This is my paper · ORCID
classification cs.RO
keywords aquaculturenetpeninspectionROVlargelanguagemodelsautonomousnavigationtaskplanningmotioncontrolreplanning
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

AquaChat claims that a language model, placed at the top of a three-layer ROV architecture, lets an operator control net-pen inspection with plain sentences instead of pre-programmed missions. The proposed pipeline converts a natural-language command into a symbolic plan, validates each step against predefined predicates, and executes it with a PID controller, and the reported experiments compare it against a rule-based planner. If the framework works as described, aquaculture staff could ask the robot to 'inspect the net from top to bottom and capture images every meter' and receive a working multi-step route that a fixed template cannot generate. The paper's evidence is a simulation with currents and a pool test with a physical ROV, both showing trajectory tracking with errors that converge over time; the claimed event-triggered replanning, however, is not exercised in these tests.

What carries the argument

The central mechanism is the three-tier decomposition: a prompt that bundles the user command with the net-pen dimensions, ROV specifications, and current observation; an LLM that returns a symbolic action sequence; a mid-level planner that validates each action against Boolean predicates (system_ready, region_detected, environment_stable, navigated(rov), inspected(area), captured(area), replan_needed) and dispatches only when preconditions hold; and a low-level PID controller that converts each symbolic action into linear and angular velocity references. The helical inspection trajectory $x(t)=r \cos(\omega t)$, $y(t)=r \sin(\omega t)$, $z(t)=z_0 - v_z t$ is what the planner must emit for full net coverage, and the feedback loop that triggers the replan action when preconditions fail is the mechanism intended to supply adaptability.

What would settle it

Set up the physical pool experiment with a visible obstacle blocking the zig-zag waypoint path or with a fan-generated current pushing the ROV sideways, and record whether the low-level feedback triggers a replan that resumes and completes the inspection; if the ROV stalls, drifts without correction, or the LLM emits an invalid plan, the adaptive-robustness claim is falsified.

Watch

Extended reading notes

Core claim

The authors claim AquaChat is the first LLM-guided ROV framework specifically designed for aquaculture net-pen inspection. In their account, the LLM planner transforms user instructions into an ordered sequence of {move_to, inspect, capture} actions, the mid-level task manager checks Boolean predicates such as system_ready and navigated(rov) before dispatching each action, and the low-level PID controller turns those actions into velocity commands. The helix that defines a full inspection is $x(t)=r \cos(\omega t)$, $y(t)=r \sin(\omega t)$, $z(t)=z_0 - v_z t$, with the ROV holding a fixed distance from the net. The reported results show the LLM planner producing valid plans for commands a rule-based planner cannot parse, and the ROV tracking reference trajectories and a physical zig-zag depth profile with errors that settle toward zero; the event-triggered replanning path is described but never triggered in the presented experiments.

Load-bearing premise

The entire adaptive-robustness claim rests on the assumption that the replanning loop, which no reported experiment ever triggers, will respond correctly to real currents, obstacles, and visibility loss in the same way the unperturbed simulation and pool tests behave.

Editorial extensions

If this is right

  • Operators can issue unstructured requests such as 'go to the lower part and take pictures' and still receive a valid multi-step plan; the LLM planner shows 65–90% planning success on the unstructured commands in Table 3, where the rule-based planner scores zero.
  • The mid-level task manager is intended to prevent actions from starting in invalid states, so a command that would otherwise be misinterpreted is checked before the ROV moves.
  • At the trajectory level, the PID controller tracks reference 'move-to' paths and a helical inspection path with converging errors, making the generated plans executable in the simulated environment.
  • On a physical ROV in a pool, the framework follows a predefined zig-zag depth profile, suggesting the plan-to-control chain transfers from simulation to hardware.

Reading between the lines

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

  • If the framework scales as claimed, the same prompt-and-predicate pipeline could be reused for nearby routine tasks (water-quality checks, mooring-line surveys) simply by changing the prompt's environment description and the predicate set.
  • The reported 2.7–47.7 s plan-generation latency suggests that real-time deployment would need a caching layer or a smaller domain-tuned model; this is my reading of the timings, not a claim the paper makes.
  • The most direct test of the adaptive claim is to inject an obstacle or current during a live run and observe whether the replanning loop actually fires; since no such experiment is reported, the adaptive part of the framework remains unverified rather than demonstrated.
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

5 major / 6 minor

Summary. The paper proposes AquaChat, a three-layer framework for ROV-based inspection of aquaculture net pens. A high-level LLM planner interprets natural-language user commands and generates symbolic plans, a mid-level task manager is intended to validate preconditions and map actions to ROV commands, and a low-level PID controller executes move-to and inspection trajectories. The authors evaluate the LLM planner by comparing its textual outputs with hand-written expected plans (Section 6.1) and evaluate the controller by tracking reference trajectories in a Gazebo simulation (Figures 9-18) and by following a predefined zig-zag depth profile with a Blueye ROV in a pool (Figures 19-20). The paper claims improved task flexibility, inspection accuracy, and operational efficiency.

Significance. If the integrated pipeline operated as claimed, AquaChat would be a useful domain-specific contribution to natural-language control of underwater inspection robots. The component-level control experiments do provide credible, non-circular evidence that a PID controller can track hand-defined reference paths in simulation and can approximate a zig-zag depth profile on a real ROV; this is a legitimate building block. However, the central contribution asserted in the abstract — the end-to-end translation of natural-language commands into executed, adaptive ROV behavior — is not demonstrated, and the LLM benchmark is internally authored rather than independent. The paper is best regarded as two separate component studies rather than a validated integrated system.

major comments (5)
  1. [Section 6.1, Tables 2-3] The evaluation of the LLM planner is circular: the ground-truth 'expected plans' in Table 2 are hand-authored by the authors, and correctness is scored by matching the LLM output to these author-written strings. No independent task benchmark, user study, or quantitative measure of plan executability is provided. In addition, the 'execution success rate' column in Table 3 is not derived from any actual execution of the generated plans; the paper does not describe how EXESR is computed, so the claim that the plans are executable is unsupported.
  2. [Section 6.2, Figures 9-20] No end-to-end experiment connects the LLM-generated symbolic plan to the executed ROV trajectory. The reference paths in Figures 9 and 14 are hand-crafted 'move to' and spiral trajectories, and the real ROV experiment uses a manually predefined zig-zag waypoint sequence (Figures 19-20). The paper therefore does not demonstrate that a natural-language command is translated into physical ROV motion, which is the central claim of the abstract.
  3. [Section 4.3] The PDDL domain is not well-formed. The :types block declares only ROV and Environment, yet predicates and action parameters use undeclared types Region and aqua-net, and the action definitions use predicates (navigated, trajectory_generated, report_sent) that do not appear in the :predicates declaration. As written, the domain cannot be parsed or validated by a planner, so the claimed precondition-checking mechanism of the mid-level task manager is not supported.
  4. [Section 5 and Section 6.2] The event-triggered replanning loop is never exercised. Although the Gazebo environment is described as including water currents, lighting variations, and visibility constraints, none of the reported experiments injects an obstacle, visibility failure, or current disturbance during a mission. Consequently, the framework's central robustness claim — adaptive replanning in response to unexpected conditions — is untested.
  5. [Abstract and Section 6] The claims of improved inspection accuracy and operational efficiency are not supported by the reported metrics. No experiment measures defect detection accuracy (the paper's own related work cites computer-vision baselines), and there is no comparison of mission time, energy, or coverage against a pre-programmed baseline; the reported results are tracking errors and plan generation times.
minor comments (6)
  1. [Section 4.3, first paragraph] The section is titled 'Mid-Level Task Planner' but the opening sentence refers to 'The Low-Level Planner bridges the gap...'; this appears to be a copy-paste error and should be corrected.
  2. [Section 5, Blueye paragraph] The vehicle is attributed to 'Blue Robotics' in the text, but references [62] and [63] identify the maker as Blueye Robotics; the attribution should be consistent.
  3. [Equation (9)] The PID gains kp, ki, and kd are not given numerical values or a tuning procedure, which limits reproducibility of the control experiments.
  4. [Table 3] The header 'Rules-based Plan' is inconsistent with the 'Rule-Based Planner' terminology used in the text and Table 1; the terminology should be unified.
  5. [Section 3] The action set is defined as {move_to, inspect, capture}, but Section 4.2 and Table 2 introduce additional action names such as inspect_net and defect detection; the notation should be made consistent.
  6. [Page 8, predicates listing] The predicate 'en v i ro n m e nt _ st a b le' contains spacing artifacts; this should be fixed in typesetting.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the internal LLM benchmark and the unvalidated end-to-end chain weaken the paper's central claim, but no experimental result reduces to its own inputs by construction.

full rationale

AquaChat's claimed derivation chain has three links: (1) the LLM planner turns a natural-language command into a symbolic action sequence (Sec. 4.2); (2) the mid-level planner validates preconditions using a PDDL domain (Sec. 4.3); (3) the low-level PID controller executes move_to and inspect actions (Sec. 4.4). No link exhibits reduction-by-construction. The LLM evaluation (Sec. 6.1, Tables 2-3) scores GPT-4 against 'Expected Plan' entries written by the same authors; this is an internally authored benchmark, but the LLM is an external system that can and does fail (Table 3 reports PSR 65-90% for unstructured prompts), so its success is not forced by the paper's definitions. The control evaluation (Sec. 6.2, Figs. 9-18) is a genuine tracking test: measured trajectories are compared with fixed hand-crafted references (a move-to path and a 3.5 m helix), the standard PID law [57] is not fitted to the reported errors, and the plotted errors are empirically nonzero. The abstract's headline claims ('improved task flexibility, inspection accuracy, and operational efficiency') are not tautological but are unsupported: no experiment connects an LLM-generated plan to an executed trajectory (the real-ROV test of Figs. 19-20 uses a manually 'predefined zig-zag inspection trajectory'), the PDDL domain of Sec. 4.3 contains undeclared types ('?area - aqua-net' while only 'ROV Environment' are declared as types, and 'trajectory_generated'/'report_sent' are used without declaration), so formal validation was evidently not run, and Table 3's EXESR (execution success rate) is never described as an experiment. These are omitted-proof and validity gaps, flagged per the review rule; they lower confidence in the central claim but are not circular equivalences. Self-citations ([5], [31], [37]-[39]) are background or support the custom net-pen model, and [31] is a published, externally falsifiable prior result, so no self-citation chain forces the paper's conclusion. Overall finding: no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The central result depends on hand-set control and mission parameters and on assumptions about LLM competence, simulator fidelity, and the validity of the PDDL model, none of which are independently verified.

free parameters (3)
  • PID gains kp, ki, kd
    Equation 9 defines the control law but never reports the gain values; all trajectory tracking results depend on them.
  • Inspection trajectory parameters = simulation: 3.5 m spiral radius; real: zig-zag depth between -1.25 m and 0.25 m
    The chosen path shape, radius, and waypoint spacing are hand-picked experimental settings, not derived from a coverage guarantee, so inspection completeness depends on them.
  • LLM prompt context
    The prompt template in Figure 3 embeds mission-specific context such as net pen dimensions, ROV specs, and constraints that is hand-authored for this domain and is not shown to transfer.
assumptions (4)
  • domain assumption GPT-4 will produce correct, executable symbolic plans from natural-language commands
    The high-level planner in Section 4.2 treats the LLM as a black box that returns valid action sequences; no fallback or verification is reported.
  • domain assumption The Gazebo/UUV simulator and custom net pen model are representative of real aquaculture conditions
    Section 5 claims current, lighting, and visibility are included, but no fidelity validation against a real site is provided.
  • ad hoc to paper The PDDL-style predicates and preconditions provide a sound model of ROV mission state
    Section 4.3 presents domain aquachat_inspection with undeclared types and predicates, so the formal model is not well-formed and cannot be executed as written.
  • standard math PID control with manually tuned gains is sufficient for disturbance rejection in net pen inspection
    The low-level planner in Section 4.4 assumes standard PID error feedback converges, but stability is not analyzed for the actual ROV.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AquaChat: An LLM-Guided ROV Framework for Adaptive Inspection of Aquaculture Net Pens." pith.science (2026). https://pith.science/paper/FEFAU6RS

@misc{pith2026250716841,
  author       = {Pith},
  title        = {Pith review of: AquaChat: An LLM-Guided ROV Framework for Adaptive Inspection of Aquaculture Net Pens},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FEFAU6RS}},
  note         = {Machine review of arXiv:2507.16841}
}
read the original abstract

Inspection of aquaculture net pens is essential for maintaining the structural integrity, biosecurity, and operational efficiency of fish farming systems. Traditional inspection approaches rely on pre-programmed missions or manual control, offering limited adaptability to dynamic underwater conditions and user-specific demands. In this study, we propose AquaChat, a novel Remotely Operated Vehicle (ROV) framework that integrates Large Language Models (LLMs) for intelligent and adaptive net pen inspection. The system features a multi-layered architecture: (1) a high-level planning layer that interprets natural language user commands using an LLM to generate symbolic task plans; (2) a mid-level task manager that translates plans into ROV control sequences; and (3) a low-level motion control layer that executes navigation and inspection tasks with precision. Real-time feedback and event-triggered replanning enhance robustness in challenging aquaculture environments. The framework is validated through experiments in both simulated and controlled aquatic environments representative of aquaculture net pens. Results demonstrate improved task flexibility, inspection accuracy, and operational efficiency. AquaChat illustrates the potential of integrating language-based AI with marine robotics to enable intelligent, user-interactive inspection systems for sustainable aquaculture operations.

Figures

Figures reproduced from arXiv: 2507.16841 by the authors.

Figure 1
Figure 1. AquaChat Project Overview: Using LLM to interpret user commands and guide adaptive ROV navigation and inspection of aquaculture net pens. The system integrates a low-level motion planner to execute action sequences such as navigation, inspection, and data capture. recognition, and defect detection with minimal human in￾tervention. Enhancing ROVs with autonomous capabilities can improve their effectiveness and reduce… view at source ↗
Figure 2
Figure 2. Proposed framework for LLM-guided navigation in aquaculture net pen inspection using ROV. The system comprises three modules: (1) High-level planner: Translates user commands into a symbolic inspection plan using GPT-4; (2) Middle-level planner: Generates a sequence of actions based on the inspection plan; (3) Low-level planner: Executes task sequences adaptively by integrating sensing, path planning, and control mo… view at source ↗
Figure 3
Figure 3. Example of prompt used for aquaculture net pens inspection. This prompt integrates the environment description and the system specification along with the plan constrains [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Output of the LLM Planner for aquaculture net pens inspection based on the user input. The plan consists of a sequences of actions to be executed by the ROV. 4.3. Mid-Level Task Planner The Low-Level Planner bridges the gap between the high-level symbolic plan generate…
Figure 6
Figure 6. Figure 6: Real-time experiment setup for AquaChat: The configuration includes an ROV positioned to face the aqua￾net, which is a rectangular net pen oriented vertically within a pool. • Mesh Detailing: A realistic net mesh to mimic the structural patterns of actual aquaculture p…
Figure 7
Figure 7. Figure 7: Comparison of Metrics Across Methods: The bar chart illustrates the performance differences between Rule￾Based and LLM-based methods across five key metrics: total commands, successful plans, average plan generation time, unrecognized commands, and success rates. 6.2. …
Figure 8
Figure 8. Figure 8: Showcasing plan executing in a simulation for actions in the plan e.g “move to”, “inspect”. relative magnitudes and demonstrates that, despite initial spikes, errors converge towards minimal values, indicating improved accuracy over the observed period [PITH_FULL_IMAG…
Figure 9
Figure 9. Figure 9: Case 1 (“move to” ): ROV 3D position during the plan execution [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Case 1 (“move to” ): ROV position (x, y, z) during the plan execution. Moreover, the errors during the tracking are shown in Figures 17 and 18. The first set of plots shows the errors in 𝑥, 𝑦, 𝑧 positions, and yaw over time. The 𝑥 error (blue) starts around 0.2 and st…
Figure 13
Figure 13. Figure 13: Case 1 (“move to” ): ROV normalized error (x, y, z, yaw) during the plan execution [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]
Figure 14
Figure 14. Figure 14: Case 2 (“inspect”): ROV 3D position during the plan execution. the ROV’s ability to maintain consistent depth and follow the planned trajectory accurately. The data obtained high￾lights the system’s potential for efficient net pen monitoring, with results visualized t…
Figure 15
Figure 15. Figure 15: Case 2 (“inspect” ): ROV position (x, y, z) during the plan execution. pattern (red dashed line and waypoints marked by bold red circles). The reference pattern follows a step-like trajectory, beginning at -1.25m until 100 seconds, sharply rising to 0.25m between 100 …
Figure 18
Figure 18. Figure 18: Case 2 (“inspect” ): ROV normalized error (x, y, z, yaw) during the plan execution [PITH_FULL_IMAGE:figures/full_fig_p019_18.png]
Figure 19
Figure 19. Figure 19: Experiment 1: ROV depth profile over time. 8. Conclusion This work proposed a novel framework named AquaChat, using LLM-guided navigation coupled ROV for the inspec￾tion of aquaculture net pens. By integrating high-level plan￾ning with domain-specific tasks and real-t…
Figure 20
Figure 20. Figure 20: Experiment 2: ROV depth profile over time. and beyond. In future, the framework will be enhanced to address challenges including model adaptation, scalability, and the extension to multi-ROV systems. Addressing these challenges will unlock new opportunities for improv…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 62 canonical work pages

  1. [1]

    Subasinghe, D

    R. Subasinghe, D. Soto, J. Jia, Global aquaculture and its role in sustainable development, Reviews in aquaculture 1 (2009) 2–9

    work page 2009

  2. [2]

    Akram, A

    W. Akram, A. Casavola, A visual control scheme for auv underwater pipeline tracking, in: 2021 IEEE International Conference on Au- tonomous Systems (ICAS), IEEE, 2021, pp. 1–5

    work page 2021

  3. [3]

    Paspalakis, K

    S. Paspalakis, K. Moirogiorgou, N. Papandroulakis, G. Giakos, M. Zervakis, Automated fish cage net inspection using image pro- cessing techniques, IET Image Processing 14 (2020) 2028–2034

    work page 2020

  4. [4]

    M.Sohan,T.SaiRam,R.Reddy,C.Venkata, Areviewonyolov8and its advancements, in: International Conference on Data Intelligence and Cognitive Informatics, Springer, 2024, pp. 529–545

    work page 2024

  5. [5]

    Autonomous Underwater Robotic System for Aquaculture Applications

    W. Akram, M. Ahmed, L. Seneviratne, I. Hussain, Autonomous un- derwater robotic system for aquaculture applications, arXiv preprint arXiv:2308.14762 (2023)

    work page Pith review arXiv 2023

  6. [6]

    K. R. Salin, G. Arome Ataguba, Aquaculture and the environment: towards sustainability, Sustainable Aquaculture (2018) 1–62

    work page 2018

  7. [7]

    H.Lee,D.Jeong,H.Yu,J.Ryu,Autonomousunderwatervehiclecon- trol for fishnet inspection in turbid water environments, International Journal of Control, Automation and Systems 20 (2022) 3383–3392

    work page 2022

  8. [8]

    Almeida, J

    J.Achiam,S.Adler,S.Agarwal,L.Ahmad,I.Akkaya,F.L.Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al., Gpt-4 technical report, arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

Show all 64 references
  1. [9]

    L. Wang, C. Ma, X. Feng, Z. Zhang, H. Yang, J. Zhang, Z. Chen, J. Tang, X. Chen, Y. Lin, et al., A survey on large language model based autonomous agents, Frontiers of Computer Science 18 (2024) 186345

    2024

  2. [10]

    R. Yang, M. Hou, J. Wang, F. Zhang, Oceanchat: Piloting au- tonomous underwater vehicles in natural language, arXiv preprint arXiv:2309.16052 (2023)

    2023 arXiv

  3. [11]

    R. Yang, F. Zhang, M. Hou, Oceanplan: Hierarchical planning and replanningfornaturallanguageauvpilotinginlarge-scaleunexplored ocean environments, arXiv preprint arXiv:2403.15369 (2024)

    2024 arXiv

  4. [12]

    R. Chen, D. Blow, A. Abdullah, M. J. Islam, Word2wave: Language driven mission programming for efficient subsea deployments of marine robots, arXiv preprint arXiv:2409.18405 (2024)

    2024 arXiv

  5. [13]

    Akram, A

    W. Akram, A. Casavola, N. Miškovic, Robust adaptive control allo- cation schemes for overactuated underwater vehicles under actuator faults, IFAC-PapersOnLine 55 (2022) 67–72

    2022

  6. [14]

    K. H. Nguyen, W. Caharija, S. J. Ohrem, J. T. Gravdahl, A. Loria, H. B. Amundsen, Robust control of autonomous remotely operated vehicles at exposed aquaculture sites. (2024)

    2024

  7. [15]

    S. J. Ohrem, H. B. Amundsen, W. Caharija, C. Holden, Robust adaptive backstepping dp control of rovs, Control Engineering Practice 127 (2022) 105282

    2022

  8. [16]

    S. J. Ohrem, L. D. Evjemo, B. O. A. Haugaløkken, H. B. Amundsen, E.Kelasidi, Adaptivespeedcontrolofrovswithexperimentalresults from an aquaculture net pen inspection operation, in: 2023 31st MediterraneanConferenceonControlandAutomation(MED),IEEE, 2023, pp. 868–875

    2023

  9. [17]

    Kelasidi, B

    E. Kelasidi, B. Su, W. Caharija, M. Føre, M. O. Pedersen, K. Frank, Autonomous monitoring and inspection operations with uuvs in fish farms, IFAC-PapersOnLine 55 (2022) 401–408

    2022

  10. [18]

    S. J. Ohrem, B. O. A. Haugaløkken, C. Holden, Application of modified model reference adaptive controller and observer (mraco) for speed control of an unmanned underwater vehicle, IFAC- PapersOnLine 58 (2024) 196–202

    2024

  11. [19]

    Botta, L

    D. Botta, L. Ebner, A. Studer, V. Reijgwart, R. Siegwart, E. Kelasidi, Framework for robust localization of uuvs and mapping of net pens,

  12. [20]

    J.Xia,T.Ma,Y.Li,S.Xu,H.Qi, Ascale-awaremonocularodometry for fishnet inspection with both repeated and weak features, IEEE Transactions on Instrumentation and Measurement (2023)

    2023

  13. [21]

    M.Bjerkeng,E.I.Grøtli,T.Kirkhus,J.T.Thielemann,H.B.Amund- sen, B. Su, S. Ohrem, Absolute localization of an rov in a fish pen usinglasertriangulation, in:202331stMediterraneanConferenceon Control and Automation (MED), IEEE, 2023, pp. 182–188

    2023

  14. [22]

    Bjerkeng, T

    M. Bjerkeng, T. Kirkhus, W. Caharija, J. T. Thielemann, H. B. Amundsen, S. Johan Ohrem, E. Ingar Grøtli, Rov navigation in a fish cage with laser-camera triangulation, Journal of Marine Science and Engineering 9 (2021) 79

    2021

  15. [23]

    Cardaillac, H

    A. Cardaillac, H. B. Amundsen, E. Kelasidi, M. Ludvigsen, Ap- plication of maneuvering based control for autonomous inspection of aquaculture net pens, in: 2023 8th Asia-Pacific Conference on Intelligent Robot Systems (ACIRS), IEEE, 2023, pp. 44–51

    2023

  16. [24]

    Akram et al.:Preprint submitted to Elsevier Page 20 of 22 Short Title of the Article of Intelligent & Robotic Systems 110 (2024) 59

    A.Cardaillac,R.Skjetne,M.Ludvigsen, Rov-basedautonomousma- neuvering for ship hull inspection with coverage monitoring, Journal W. Akram et al.:Preprint submitted to Elsevier Page 20 of 22 Short Title of the Article of Intelligent & Robotic Systems 110 (2024) 59

    2024

  17. [25]

    Skaldebø, C

    M. Skaldebø, C. Schellewald, L. D. Evjemo, H. B. Amundsen, M. Xanthidis, E. Kelasidi, Approaches enabling underwater auton- omy and sensing in sea-based aquaculture settings, in: 2024 32nd MediterraneanConferenceonControlandAutomation(MED),IEEE, 2024, pp. 197–202

    2024

  18. [26]

    Schellewald, A

    C. Schellewald, A. Stahl, E. Kelasidi, Vision-based pose estima- tion for autonomous operations in aquacultural fish farms, IFAC- PapersOnLine 54 (2021) 438–443

    2021

  19. [27]

    Y.Wu,J.Liu,Y.Wei,D.An,Y.Duan,W.Li,B.Li,Y.Chen,Q.Wei, Intelligent control method of underwater inspection robot in netcage, Aquaculture Research 53 (2022) 1928–1938

    2022

  20. [28]

    H. B. Amundsen, W. Caharija, K. Y. Pettersen, Autonomous rov inspections of aquaculture net pens using dvl, IEEE Journal of Oceanic Engineering 47 (2021) 1–19

    2021

  21. [29]

    D. Rosa, D. Cabecinhas, F. Ferreira, Forward-looking sonar based autonomous aquaculture inspection, in: OCEANS 2024-Singapore, IEEE, 2024, pp. 1–8

    2024

  22. [30]

    T. T. Tun, L. Huang, M. A. Preece, Development and high-fidelity simulation of trajectory tracking control schemes of a uuv for fish net-pen visual inspection in offshore aquaculture, IEEE Access 11 (2023) 135764–135787

    2023

  23. [31]

    Akram, A

    W. Akram, A. Casavola, N. Kapetanovic, N. Miskovic, A visual servoing scheme for autonomous aquaculture net pens inspection using rov, Sensors 22 (2022) 3525

    2022

  24. [32]

    B.O.Haugaløkken,O.Nissen,M.B.Skaldebø,S.J.Ohrem,E.Kela- sidi, Low-cost sensor technologies for underwater vehicle navigation in aquaculture net pens, IFAC-PapersOnLine 58 (2024) 87–94

    2024

  25. [33]

    M. N. A. Rahim, D. T. A. Hamid, M. F. M. Said, M. N. Jamaludin, H.H.Rozalan, Designanddevelopmentofremotelyoperatedvehicle to assess the water quality in aquaculture area, Asian People Journal (APJ) 5 (2022) 50–60

    2022

  26. [34]

    P.Tarwadi,Y.Shiraki,O.Ganoni,S.Wei,H.S.Ahn,B.MacDonald, Designanddevelopmentofaroboticvehicleforshallow-watermarine inspections, arXiv preprint arXiv:2007.04563 (2020)

    2020 arXiv

  27. [35]

    Vasileiou, N

    M. Vasileiou, N. Manos, E. Kavallieratou, A low-cost 3d printed mini underwater vehicle: Design and fabrication, in: 2021 20th InternationalConferenceonAdvancedRobotics(ICAR),IEEE,2021, pp. 390–395

    2021

  28. [36]

    G.Xu,D.Zhou,L.Yuan,W.Guo,Z.Huang,Y.Zhang, Vision-based underwater target real-time detection for autonomous underwater vehicle subsea exploration, Frontiers in Marine Science 10 (2023) 1112310

    2023

  29. [37]

    Akram, A

    W. Akram, A. Baidar Bakht, M. Ud Din, L. Seneviratne, I. Hussain, Enhancing aquaculture net pen inspection: A benchmark study on detection and semantic segmentation, IEEE Access 13 (2025) 3453– 3474

    2025

  30. [38]

    Hussain, Aquaculture defects recognition via multi-scale semanticsegmentation, Expertsystemswithapplications237(2024) 121197

    W.Akram,T.Hassan,H.Toubar,M.Ahmed,N.Miškovic,L.Senevi- ratne, I. Hussain, Aquaculture defects recognition via multi-scale semanticsegmentation, Expertsystemswithapplications237(2024) 121197

    2024

  31. [39]

    Akram, M

    W. Akram, M. Ahmed, L. Seneviratne, I. Hussain, Evaluating deep learningassistedautomatedaquaculturenetpensinspectionusingrov, arXiv preprint arXiv:2308.13826 (2023)

    2023 arXiv

  32. [40]

    Madshaven, C

    A. Madshaven, C. Schellewald, A. Stahl, Hole detection in aqua- culture net cages from video footage, in: Fourteenth International Conference on Machine Vision (ICMV 2021), volume 12084, SPIE, 2022, pp. 258–267

    2021

  33. [41]

    C.Schellewald,A.Stahl, Irregularitydetectioninnetpensexploiting computer vision, IFAC-PapersOnLine 55 (2022) 415–420

    2022

  34. [42]

    Ronneberger, P

    O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation, in: Medical image comput- ing and computer-assisted intervention–MICCAI 2015: 18th interna- tionalconference,Munich,Germany,October5-9,2015,proceedings, part III 18, Springer,...

    2015

  35. [43]

    Zacheilas, K

    T. Zacheilas, K. Moirogiorgou, N. Papandroulakis, E. Sotiriades, M. Zervakis, A. Dollas, An fpga-based system for video processing to detect holes in aquaculture nets, in: 2021 IEEE 21st International Conference on Bioinformatics and Bioengineering (BIBE), IEEE, 2021, pp. 1–6

    2021

  36. [44]

    Jocher, A

    G. Jocher, A. Chaurasia, A. Stoken, et al., Yolov5 - you only look once, https://github.com/ultralytics/yolov5,2020.Accessed:2024- 09-23

    2020

  37. [45]

    Paraskevas, E

    K. Paraskevas, E. Kavallieratou, Detecting holes in fish farming nets: A two–method approach, in: 2023 International Conference on Control, Automation and Diagnosis (ICCAD), IEEE, 2023, pp. 1–7

    2023

  38. [46]

    K.Paraskevas,E.Kavallieratou, Detectingholesinfisherynetsusing an rov, in: 2022 International Conference on Electrical, Computer, CommunicationsandMechatronicsEngineering(ICECCME),IEEE, 2022, pp. 1–5

    2022

  39. [47]

    J.-H. Kang, T. Keruzel, U.-J. Baek, K.-C. Lee, Detection of fish cage net damage using image processing with mesh-hole grouping, in: 2023 IEEE Region 10 Symposium (TENSYMP), IEEE, 2023, pp. 1– 3

    2023

  40. [48]

    W.Qiu,V.Pakrashi,B.Ghosh, Fishingnethealthstateestimationus- ing underwater imaging, Journal of Marine Science and Engineering 8 (2020) 707

    2020

  41. [49]

    Zhang, F

    Z. Zhang, F. Gui, X. Qu, D. Feng, Netting damage detection for marine aquaculture facilities based on improved mask r-cnn, Journal of Marine Science and Engineering 10 (2022) 996

    2022

  42. [50]

    López-Barajas, P

    S. López-Barajas, P. J. Sanz, R. Marín-Prades, A. Gómez-Espinosa, J. González-García, J. Echagüe, Inspection operations and hole detection in fish net cages through a hybrid underwater intervention system using deep learning techniques, Journal of Marine Science and Engineerin...

    2023

  43. [51]

    Betancourt, W

    J. Betancourt, W. Coral, J. Colorado, An integrated rov solution for underwater net-cage inspection in fish farms using computer vision, SN Applied Sciences 2 (2020) 1946

    2020

  44. [52]

    Z. Bi, N. Zhang, Y. Xue, Y. Ou, D. Ji, G. Zheng, H. Chen, Oceangpt: A large language model for ocean science tasks, arXiv preprint arXiv:2310.02031 (2023)

    2023 arXiv

  45. [53]

    D. J. Samuel, Y. Sermet, D. Cwiertny, I. Demir, Integrating vision- based ai and large language models for real-time water pollution surveillance, Water Environment Research 96 (2024) e11092

    2024

  46. [54]

    Khanal, C

    N. Khanal, C. M. Yu, J.-C. Chiu, A. Chaudhary, Z. Zhang, K. Katija, A. G. Forbes, Fathomgpt: A natural language interface for interac- tively exploring ocean science data, in: Proceedings of the 37th An- nual ACM Symposium on User Interface Software and Technology, 2024, pp. 1–15

    2024

  47. [55]

    S. Lian, H. Li, Evaluation of segment anything model 2: The role of sam2 in the underwater environment, arXiv preprint arXiv:2408.02924 (2024)

    2024 arXiv

  48. [56]

    M. N. Cahyadi, T. Asfihani, R. Mardiyanto, R. Erfianti, Performance ofgpsandimusensorfusionusingunscentedkalmanfilterforprecise i-boat navigation in infinite wide waters, Geodesy and Geodynamics 14 (2023) 265–274

    2023

  49. [57]

    K. H. Ang, G. Chong, Y. Li, Pid control system analysis, design, and technology, IEEE transactions on control systems technology 13 (2005) 559–576

    2005

  50. [58]

    U. S. D. Team, Uuv simulator: Gazebo/ros packages for underwater vehicle simulation, urlhttps://uuvsimulator.github.io, 2018. Accessed: 2025-01-16

    2018

  51. [59]

    Robotics, Contributors, Bluerov2 simulator: Simulation package for bluerov2 in ros/gazebo, urlhttps://github.com/HKPolyU-UAV/bluerov2, 2018

    B. Robotics, Contributors, Bluerov2 simulator: Simulation package for bluerov2 in ros/gazebo, urlhttps://github.com/HKPolyU-UAV/bluerov2, 2018. Accessed: 2025-01-16

    2018

  52. [60]

    URL: https://bluerobotics.com/store/rov/bluerov2/, accessed: 2025-01-16

    B.Robotics,Bluerov2-affordableandcapableunderwaterrov,2025. URL: https://bluerobotics.com/store/rov/bluerov2/, accessed: 2025-01-16

    2025

  53. [61]

    M.vonBenzon,F.F.Sørensen,E.Uth,J.Jouffroy,J.Liniger,S.Ped- ersen, An open-source benchmark simulator: Control of a bluerov2 underwater robot, Journal of Marine Science and Engineering 10 (2022) 1898

    2022

  54. [62]

    Robotics, Blueye x3 rov, https://www.blueyerobotics.com/ products/x3, 2025

    B. Robotics, Blueye x3 rov, https://www.blueyerobotics.com/ products/x3, 2025. Accessed: 2025-01-23. W. Akram et al.:Preprint submitted to Elsevier Page 21 of 22 Short Title of the Article

    2025

  55. [63]

    Robotics, Blueye robotics sdk, https://github.com/ BluEye-Robotics/blueye.sdk, 2025

    B. Robotics, Blueye robotics sdk, https://github.com/ BluEye-Robotics/blueye.sdk, 2025. Accessed: 2025-01-23. W. Akram et al.:Preprint submitted to Elsevier Page 22 of 22

    2025

  56. [2024]

    arXiv:2409.15475

    URL:https://arxiv.org/abs/2409.15475. arXiv:2409.15475

    arXiv

Pith tools

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