Pith. sign in

REVIEW 4 major objections 6 minor 17 references

AURA claims that feeding human-validated diagnoses back into an LLM as retrievable case memory turns generic anomaly reports into precise causal identifications, cutting operator effort by 71%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 23:58 UTC pith:6EOXBDXW

load-bearing objection Real system-building paper with a load-bearing evaluation confound: the learning claim is really a retrieval effect. the 4 major comments →

arxiv 2511.03075 v2 pith:6EOXBDXW submitted 2025-11-04 cs.RO

A Collaborative Reasoning Framework for Anomaly Diagnostics in Underwater Robotics

Markus Buchholz , Niamh Ellis , Rahaf Abu Hara , Ignacio Carlucho , Yvan R. Petillot This is my paper
classification cs.RO
keywords AURAunderwater roboticsanomaly diagnosticshuman-in-the-loop distillationdigital twinlarge language modelsretrieval-augmented generationfault diagnosis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that an underwater robot's fault-diagnosis capability can be made to improve from one encounter to the next by recording each human-validated diagnosis as a retrievable memory. The proposed framework, AURA, pairs a perception agent that turns telemetry into natural-language problem descriptions with a reasoning agent that talks an operator through possible causes. After a diagnosis is confirmed, the episode is distilled into a training example in a vector database; on later encounters, the perception agent retrieves that example and produces sharper, cause-oriented descriptions. The reported experiments show average characterization quality rising from 2.7 to 4.8 on a 5-point scale and dialog turns falling from 6.2 to 1.8. Why it matters: if true, it gives a concrete mechanism for turning operational experience into safer, more autonomous robots without retraining or cloud dependence.

Core claim

The central claim is that a human-AI diagnostic loop can be made self-improving through Human-in-the-Loop Distillation: every resolved anomaly, once the operator validates the root cause, becomes a distilled lesson stored as an embedding in a vector database. When a similar anomaly appears later, the low-level State Anomaly Characterisation Agent retrieves that lesson and uses it to write a problem description that names probable causes instead of merely listing symptoms. This sharper input lets the Diagnostic Reasoning Agent and the operator converge on a diagnosis in a fraction of the time. The paper reports that with a database primed by five validated experiences, the average Characteriz

What carries the argument

The load-bearing mechanism is the two-agent architecture plus its memory loop: Agent A (State Anomaly Characterisation Agent) monitors telemetry against a high-fidelity digital twin, flags deviations via Mahalanobis distance, and translates raw residuals into a structured natural-language problem characterisation; Agent B (Diagnostic Reasoning Agent) grounds that characterisation in external knowledge and runs an interactive dialogue with the operator. The key identity is the training tuple (raw anomaly data → correct problem characterisation) created at the end of each validated session and stored in a local vector database after embedding; retrieval-augmented generation lets Agent A use th

Load-bearing premise

The evaluation assumes that the Phase 2 anomalies are previously unseen and that the Post-Distillation improvement reflects generalization, but Phase 1 and Phase 2 use the same three anomaly classes, so the gain could come from retrieving a nearly identical stored case rather than from a capacity to diagnose truly novel faults.

What would settle it

Run Phase 2 with the primed VDB on an anomaly class that was not part of the five priming sessions, such as a deliberate IMU sensor failure or a communication dropout. If the average CSS stays near the First-Encounter baseline (about 2.7) and dialog turns remain near 6 instead of falling to 1.8, the claimed generalization from distilled experience is not supported; it would show the benefit depends on case retrieval from near-identical stored experiences.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • With a primed memory of five validated cases, the system's initial problem descriptions jump from symptom-level transcription (average CSS 2.7) to cause-oriented hypotheses (average CSS 4.8).
  • Operator effort, measured in dialog turns to reach a diagnosis with >90% confidence, drops from 6.2 to 1.8 turns, a 71% reduction.
  • The feedback loop is self-contained and runs on locally hosted models, meaning the learning happens without sending vehicle telemetry off-site.
  • Pre-mission knowledge injection lets operators teach the system from historical logs or simulated dialogues before deployment, bootstrapping the memory.
  • Because the human operator remains the final arbiter of every diagnosis, the memory retains only expert-validated lessons, which acts as a verifiability filter against flawed reasoning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The authors call the Phase 2 anomalies 'previously unseen,' but they belong to the same three classes used to prime the memory; the reported gains may partly reflect retrieval of near-identical stored cases rather than reasoning about genuinely novel faults. A fair test would use a held-out anomaly class.
  • The same distillation loop could plausibly transfer to other safety-critical domains — medical device monitoring, autonomous ground vehicles, or industrial robotics — wherever a digital twin or normative model can supply residuals and an operator can validate root causes.
  • The proposed third agent (Knowledge Curation Agent) addresses a real long-term risk: without novelty assessment, the vector database will accumulate redundant or conflicting lessons, degrading retrieval quality over extended deployments.
  • Since Agent B's tool use is deliberately limited to one search function and the system defers to the operator for decisions, AURA is a cognitive-support tool rather than an autonomous decision-maker; the headline improvements are about how quickly human-AI teams converge, not about removing the human.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces AURA, a two-agent architecture for anomaly diagnostics in an underwater vehicle. Agent A (State Anomaly Characterisation Agent) compares live BlueROV2 telemetry with a Stonefish digital twin, detects statistical deviations via Mahalanobis distance, and converts the residual signal into a natural-language problem characterization. Agent B (Diagnostic Reasoning Agent) runs a knowledge-grounded, human-in-the-loop dialogue to reach a root-cause diagnosis. After a human-validated diagnosis, a distilled lesson is embedded and stored in a vector database (VDB); on later encounters Agent A retrieves this context through retrieval-augmented generation. The evaluation compares a First Encounter condition (empty VDB) with a Post-Distillation condition (VDB primed with five human-validated sessions) over three anomaly classes, reporting an increase in Characterization Specificity Score from 2.7 to 4.8 and a reduction in dialog turns from 6.2 to 1.8. The paper interprets this as evidence that Human-in-the-Loop Distillation creates a virtuous learning cycle that improves diagnostic performance on previously unseen anomalies.

Significance. If the central claim were established, AURA would be a useful design pattern for combining LLM-based reasoning, digital twins, and human validation in safety-critical robotics. The paper has concrete strengths: a real-time physical/virtual testbed, a privacy-preserving local LLM stack, a clear separation between perception and reasoning agents, and a human-validation filter for the stored experience. The evaluation, however, does not currently support the headline claim of generalization to unforeseen anomalies. The improvement may largely reflect direct retrieval of near-identical stored cases from the same anomaly classes used for priming, and the CSS metric and single-rater protocol further confound the measurement. The architectural contribution is worth further development, but the empirical evidence requires substantial rework.

major comments (4)
  1. [Section V-A and V-B] The central empirical claim is that Post-Distillation improves performance on 'new, previously unseen anomalies.' This is not established. Phase 1 priming sessions were explicitly 'varied and covered the different use cases (e.g., two thruster disturbances, two anomalies in commanded rotational motion, and one anomaly in commanded vertical motion),' and Phase 2 tests exactly the same three use-case classes (thruster disturbance, vertical-motion anomaly, rotational-motion anomaly). Therefore every Phase 2 test has at least one stored example of the same class in the VDB. The improvement from CSS 2.7 to 4.8 and turns 6.2 to 1.8 can be explained by near-identical case retrieval rather than by generalization to genuinely novel faults. A held-out fault-class condition, or priming with a disjoint set of classes, is required to support the 'unforeseen anomaly' claim.
  2. [Section V-C, CSS rubric] The CSS rubric itself rewards VDB-based retrieval: score 3 is defined as 'uses historical context from the VDB to identify the single most likely, specific root cause based on past, similar events.' Thus the Post-Distillation condition is partly being graded on whether the agent retrieved the stored example, making the observed improvement partially tautological. Moreover, the CSS is a single expert operator's rating with no blinding and no inter-rater reliability; 'n=5' repetitions of the same protocol do not provide statistical validity without independent raters and agreement statistics. Please report multiple blinded raters, inter-rater agreement, and confidence intervals or effect sizes.
  3. [Table I and Section V-C] The table reports only average CSS and dialog-turn counts. No standard deviations, per-repetition values, or significance tests are provided. With n=5, the differences (e.g., 2.7 vs 4.8) could be within variability; the text's 'significant improvement' and 'direct causal effect' are not supported without appropriate statistical analysis. Please provide the underlying data and a paired test across the five repetitions.
  4. [Sections IV-A, IV-C, VI] The paper repeatedly states that the human-validated diagnosis 'refines the low-level agent's perceptual model' and that expert knowledge is 'distilled into the AI.' In the actual implementation, however, the only update is insertion of a text lesson into the VDB, retrieved later via RAG; Agent A's model weights are never updated. The contribution should be framed as episodic case-based reasoning with retrieval-augmented memory, not as refinement of a perceptual model, unless actual model updating is demonstrated. This overstatement affects the framing of the main contribution and the conclusion.
minor comments (6)
  1. [Fig. 3 caption] The caption uses 'Training Anomaly' and 'Validation Anomaly,' while Section V uses 'First Encounter' and 'Post-Distillation.' Please unify the terminology.
  2. [Section III-C] The models are described only by parameter range ('4-12 billion,' 'Gemma 3 27B'). For reproducibility, specify exact model names, quantization, inference server versions, embedding model, and hyperparameters.
  3. [Section V-B] The physical fault injection ('various external physical forces,' 'physically impeded') is not quantified. Provide magnitudes, durations, and descriptions of how anomalies were generated in the real vehicle.
  4. [Table I] The column header 'CSS / Dialog Turns [Average]' is ambiguous because two very different quantities are reported in one column. Use separate columns with explicit units.
  5. [References] Reference [17] is an arXiv preprint without publication venue while [16] is the same simulator's OCEANS paper; please check and format consistently.
  6. [Section V-A] The phrase 'new, previously unseen anomalies' is misleading because the same three classes were used in Phase 1. Revise this wording throughout, including the abstract, unless a truly held-out class is added.

Circularity Check

2 steps flagged

Post-Distillation improvement reduces to retrieval of same-class stored cases; CSS rubric rewards VDB use by definition

specific steps
  1. fitted input called prediction [Section V-A (Phase 1 and Phase 2) and Section V-B (Use Cases)]
    "Starting with an empty VDB, we conducted five initial diagnostic sessions. These sessions were varied and covered the different use cases (e.g., two thruster disturbances, two anomalies in commanded rotational motion, and one anomaly in commanded vertical motion). ... We measured the system's performance on a set of new, previously unseen anomalies. ... Three classes of anomalies were tested: Thruster Control Under Disturbance; Anomalies in Commanded Vertical Motion; Anomalies in Commanded Rotational Motion."

    Phase 1 populates the VDB with validated diagnoses from the same three classes used as Phase 2 test cases, so the Post-Distillation condition has a stored, human-validated exemplar for every Phase 2 use case. The reported improvement (CSS 2.7→4.8; dialog turns 6.2→1.8) can be obtained by retrieving that near-identical stored lesson, which is exactly the mechanism described in Section IV-A: 'On subsequent encounters with a similar anomaly, Agent A retrieves this distilled lesson.' The control condition with an empty VDB is therefore not a test of generalization to previously unseen fault classes; the 'prediction' of improved performance is largely the re-presentation of the priming input.

  2. self definitional [Section V-C (CSS rubric, score 3)]
    "3: Targeted Hypothesis: The output uses historical context from the VDB to identify the single most likely, specific root cause based on past, similar events."

    The CSS is the paper's main measure of 'perceptual acuity' and diagnostic quality, yet score 3 is defined as using VDB historical context to identify a root cause based on past similar events. With an empty VDB, the First Encounter condition cannot attain this score by definition, while Post-Distillation can, simply because the VDB was loaded. Thus a substantial part of the measured 2.7→4.8 improvement is an artifact of the rubric rewarding the intervention itself; the metric does not independently measure reasoning ability or perceptual acuity apart from successful retrieval.

full rationale

The central quantitative claim (Table I) is that Post-Distillation improves CSS from 2.7 to 4.8 and reduces dialog turns from 6.2 to 1.8, interpreted as a 'virtuous learning cycle' on 'previously unseen anomalies.' This is not circular in the sense of a mathematical derivation, but the experiment is structured so the improvement is largely the direct retrieval of stored cases. Phase 1 explicitly primes the VDB with five validated sessions covering exactly the three use-case classes later tested in Phase 2, and Phase 2 claims to test 'new, previously unseen anomalies.' Since the Post-Distillation condition gives Agent A access to these stored same-class lessons, its outputs are retrieved 'distilled lessons' rather than evidence of novel fault reasoning. The CSS rubric itself defines score 3 as 'uses historical context from the VDB,' so the First Encounter condition cannot receive that score by design; part of the measured gain is built into the metric. The paper's own Stage 1 text confirms the mechanism: 'On subsequent encounters with a similar anomaly, Agent A retrieves this distilled lesson.' Thus the headline contribution—improved perceptual acuity on unseen faults—is not independently established; the measured improvement reduces to case matching plus a rubric that rewards case matching. There is no load-bearing self-citation or imported uniqueness theorem; those circularity patterns are absent. The score of 6 reflects partial circularity: the claimed 'prediction' of generalization is not forced by an equation, but the reported improvement is by construction tied to the VDB input and the evaluation rubric.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The framework does not introduce new physical entities or novel mathematical objects. The main assumptions are about the fidelity of the digital twin, the reliability of LLM-based semantic grounding, and the validity of the VDB retrieval as a learning mechanism. The free parameters are mostly thresholds and stopping rules that are hand-set and influence the reported metrics.

free parameters (4)
  • Mahalanobis distance threshold (chi-square quantile) = not reported
    Anomaly detection uses MD² compared against a chi-squared threshold; the specific quantile is not given and is a hand-tuned parameter that determines when an anomaly is flagged.
  • Adaptive residual thresholds = not reported
    The text says 'if any residual exceeds adaptive thresholds' in Section III-B; these thresholds are not defined or tuned and influence the anomaly trigger.
  • Dialog termination confidence (90%) = 90% operator confidence
    Section V-C: 'The diagnostic dialog was terminated once a mutually agreed-upon hypothesis was reached with over 90% operator confidence.' This is a hand-set stopping rule that directly affects the dialog-turn metric.
  • CSS rubric anchor points = 5-point Likert rubric
    The characterization specificity score is defined by a rubric (Section V-C) and assigned by an expert; the rubric descriptions are subjective and the assignment is not validated for reliability.
axioms (4)
  • domain assumption The Stonefish digital twin is a 'high-fidelity' and accurate normative model for the real BlueROV2.
    The entire anomaly-detection mechanism (Section III-A, III-B) relies on the twin producing a trustworthy baseline; no validation of the twin's fidelity is provided.
  • domain assumption A moderate LLM (4-12B parameters) can reliably convert raw telemetry differences into meaningful natural-language problem characterizations.
    Stage 1 (Section IV-A) assumes Agent A's signal-to-symbol mapping is accurate enough to support downstream reasoning; the paper provides no quantitative measure of translation correctness.
  • ad hoc to paper Retrieving past human-validated lessons from a vector database improves characterization for new similar anomalies.
    The Post-Distillation condition demonstrates that retrieval helps for the same anomaly classes, but the claim that this generalizes to genuinely new faults is assumed, not established, because the test classes overlap with the stored cases.
  • domain assumption Human-validated diagnoses in the VDB are correct and reliable.
    The feedback loop (Stage 3, Section IV-C) treats the operator's confirmed diagnosis as ground truth for future learning; errors in that judgment would propagate.

pith-pipeline@v1.3.0-alltime-deepseek · 8678 in / 8032 out tokens · 71576 ms · 2026-08-03T23:58:59.725472+00:00 · methodology

0 comments
read the original abstract

The safe deployment of autonomous systems in safety-critical settings requires a paradigm that combines human expertise with AI-driven analysis, especially when anomalies are unforeseen. We introduce AURA (Autonomous Resilience Agent), a collaborative framework for anomaly and fault diagnostics in robotics. AURA integrates large language models (LLMs), a high-fidelity digital twin (DT), and human-in-the-loop interaction to detect and respond to anomalous behavior in real time. The architecture uses two agents with clear roles: (i) a low-level State Anomaly Characterization Agent that monitors telemetry and converts signals into a structured natural-language problem description, and (ii) a high-level Diagnostic Reasoning Agent that conducts a knowledge-grounded dialogue with an operator to identify root causes, drawing on external sources. Human-validated diagnoses are then converted into new training examples that refine the low-level perceptual model. This feedback loop progressively distills expert knowledge into the AI, transforming it from a static tool into an adaptive partner. We describe the framework's operating principles and provide a concrete implementation, establishing a pattern for trustworthy, continually improving human-robot teams.

Figures

Figures reproduced from arXiv: 2511.03075 by Ignacio Carlucho, Markus Buchholz, Niamh Ellis, Rahaf Abu Hara, Yvan R. Petillot.

Figure 1
Figure 1. Figure 1: High-level concept of AURA. We use a DT (top [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The AURA Collaborative Reasoning Architecture. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: An illustrative example of the Human-in-the-Loop Distillation process, designed to produce a quantifiable improvement [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The experimental validation platform for the AURA framework. The system simultaneously processes state telemetry [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of AURA’s performance, contrasting its response to a [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 8 linked inside Pith

  1. [1]

    Fault diagnosis of complex systems using a large language model-based agent,

    E. Cui, A. Khattak, W. Du, Y .-J. Li, and K.-M. Lee, “Fault diagnosis of complex systems using a large language model-based agent,”IEEE Transactions on Instrumentation and Measurement, vol. 73, pp. 1–11, 2024

    2024

  2. [2]

    An empirical characterization of outages and incidents in public services for large language models,

    X. Chu, S. Talluri, Q. Lu, and A. Iosup, “An empirical characterization of outages and incidents in public services for large language models,”arXiv preprint arXiv:2501.12469, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    On the vulnerability of llm/vlm-controlled robotics,

    X. Wu, S. Chakraborty, R. Xian, J. Liang, T. Guan, F. Liu, B. M. Sadler, D. Manocha, and A. S. Bedi, “On the vulnerability of llm/vlm-controlled robotics,”arXiv preprint arXiv:2402.10340, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    T. I. Fossen,Handbook of Marine Craft Hydrodynamics and Motion Control. Chichester, UK: Wiley, 2011

    2011

  5. [5]

    Digital twin-based out- of-distribution detection in autonomous vessels,

    E. Isaku, H. Sartaj, and S. Ali, “Digital twin-based out- of-distribution detection in autonomous vessels,”arXiv preprint arXiv:2504.19816, 2025

    arXiv 2025

  6. [6]

    Digital twin-assisted fault diagnosis system for robot joints with insufficient data,

    Z. Song, H. Shi, X. Bai, and G. Li, “Digital twin-assisted fault diagnosis system for robot joints with insufficient data,”Journal of Field Robotics, vol. 40, no. 2, pp. 258–271, 2023

    2023

  7. [7]

    Sparks of artificial general intelligence: Early experiments with gpt-4,

    S. Bubeck, V . Chandrasekaran, R. Eldan, J. Gehrke, E. Horvitz, E. Kamar, P. Lee, Y . T. Lee, Y . Li, S. Lundberg, H. Nori, H. Palangi, M. T. Ribeiro, and Y . Zhang, “Sparks of artificial general intelligence: Early experiments with gpt-4,”arXiv preprint arXiv:2303.12712, 2023, accessed: March 2025. [Online]. Available: https://arxiv.org/abs/2303.12712

    Pith/arXiv arXiv 2023

  8. [8]

    Chatgpt for robotics: Design principles and model abilities,

    S. Vemprala, R. Bonatti, A. Bucker, and A. Kapoor, “Chatgpt for robotics: Design principles and model abilities,”arXiv preprint arXiv:2306.17582, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Do as i can, not as i say: Grounding language in robotic affordances,

    M. Ahn, A. Brohan, N. Brown, Y . Chebotar, O. Cortes, C. Finn, K. Hausman, A. Herzoget al., “Do as i can, not as i say: Grounding language in robotic affordances,” inProc. of the Conference on Robot Learning (CoRL), 2022

    2022

  10. [10]

    Palm- e: An embodied multimodal language model,

    D. Driess, F. Xia, M. S. M. Sajjadi, C. Lynch, A. Chowd- hery, B. Ichter, A. Wahid, J. Tompson, Q. Vuong, T. Yu, W. Huang, Y . Chebotar, P. Sermanet, D. Duckworth, S. Levine, V . Vanhoucke, K. Hausman, M. Toussaint, K. Greff, A. Zeng, I. Mordatch, and P. Florence, “Palm- e: An embodied multimodal language model,”arXiv preprint arXiv:2303.03378, 2023

    Pith/arXiv arXiv 2023

  11. [11]

    Embodied large language models enable robots to complete complex tasks in unpredictable environments,

    R. Mon-Williams, G. Li, R. Long, W. Du, and C. G. Lu- cas, “Embodied large language models enable robots to complete complex tasks in unpredictable environments,” Nature Machine Intelligence, vol. 7, pp. 592–601, 2025

    2025

  12. [12]

    J. E. Laird,The Soar Cognitive Architecture. MIT Press, 2012

    2012

  13. [13]

    Training language models to follow instructions with human feedback,

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. L. Wain- wright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, J. Schulman, J. Hilton, F. Kelton, L. Miller, M. Simens, A. Askell, P. Welinder, P. Christiano, J. Leike, and R. Lowe, “Training language models to follow instructions with human feedback,” inAdvances in Neural Information Processing Systems (Ne...

    2022

  14. [14]

    Robot learning on the job: Human-in-the-loop auton- omy and learning during deployment,

    H. Liu, S. Nasiriany, L. Zhang, Z. Bao, and Y . Zhu, “Robot learning on the job: Human-in-the-loop auton- omy and learning during deployment,”arXiv preprint arXiv:2211.08416, 2022

    Pith/arXiv arXiv 2022

  15. [15]

    Retrieval- augmented generation for knowledge-intensive nlp tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. K ¨uttler, M. Lewis, W. tau Yih, T. Rockt ¨aschel, S. Riedel, and D. Kiela, “Retrieval- augmented generation for knowledge-intensive nlp tasks,” arXiv preprint arXiv:2005.11401, 2020

    Pith/arXiv arXiv 2005

  16. [16]

    Stonefish: An advanced open-source sim- ulation tool designed for marine robotics, with a ros interface,

    P. Cie´slak, “Stonefish: An advanced open-source sim- ulation tool designed for marine robotics, with a ros interface,” inOCEANS 2019 - Marseille, 2019, pp. 1–6

    2019

  17. [17]

    Stonefish: Supporting machine learning research in marine robotics,

    M. Grimaldi, P. Cieslak, E. Ochoa, V . Bharti, H. Rajani, I. Carlucho, M. Koskinopoulou, Y . R. Petillot, and N. Gracias, “Stonefish: Supporting machine learning research in marine robotics,” 2025. [Online]. Available: https://arxiv.org/abs/2502.11887

    Pith/arXiv arXiv 2025