Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Visual habitat context around a sparse coral lets an AUV collect most of the target in half the time of exhaustive coverage.

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 · grok-4.5

2026-07-14 23:49 UTC pith:TPJ6YXHC

load-bearing objection Clean empirical result: one-shot DINOv2 context buffer beats target-only and lawnmower on sparse real reef imagery, but the 75%-in-half-time claim is still offline-only. the 4 major comments →

arxiv 2603.10174 v2 pith:TPJ6YXHC submitted 2026-03-10 cs.RO

Autonomous Search for Sparsely Distributed Visual Phenomena through Environmental Context Modeling

Eric Chen , Travis Manderson , Nare Karapetyan , Peter Edmunds , Nicholas Roy , Yogesh Girdhar This is my paper
classification cs.RO
keywords autonomous underwater vehiclescoral reef surveyone-shot detectionenvironmental contextDINOv2adaptive path planningsparse target search
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.

When a target coral is too rare for its own detections to steer an underwater robot, the denser visual features that usually live next to it can still point the way. From a single labeled photo the method extracts patch embeddings that detect both the coral and its co-occurring habitat on the fly, then uses the denser habitat signal to choose the next grid cell. Simulated trajectories on real AUV imagery from two Caribbean reefs show that following the combined signal gathers up to three-quarters of the target instances in roughly half the time of a full lawnmower path and beats strategies that chase only the coral itself. The result matters because battery-limited vehicles cannot afford to search every square meter, yet scientists still need as many observations of sparse species as possible.

Core claim

One-shot detection of a target coral species combined with online modeling of its visual environmental context enables efficient autonomous surveying: when the target is sparsely distributed the method samples up to 75 percent of the target in roughly half the time required by exhaustive coverage and outperforms search strategies that use only target detections.

What carries the argument

The online environment-context buffer: non-target DINOv2 patch embeddings sampled from images that score high for the target, kept in a fixed-size FIFO queue and used as a denser, smoother reward so the robot can still choose a useful next cell even where no target has yet been seen.

Load-bearing premise

The method assumes that the visual features living near a target coral co-occur strongly enough with it that chasing those features still leads the robot toward more of the target even in empty regions.

What would settle it

Run the same greedy policy on a reef where the target coral is deliberately uncorrelated with surrounding habitat texture (for example randomly transplanted colonies); if context-guided paths no longer outperform target-only or random paths, the co-occurrence premise fails.

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

If this is right

  • Adaptive planners can collect most instances of a sparse target before a full coverage path finishes.
  • A single labeled image is enough to start both target detection and context modeling without retraining.
  • Context signals derived from co-occurring habitat outperform simple substrate segmentations as exploration cues.
  • Online buffer updates help when habitat varies across a site; a fixed buffer still works on small reefs.
  • The approach generalizes across biologically different coral species and two reef sites.

Where Pith is reading between the lines

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

  • The same co-occurrence idea could guide search for other sparse visual targets such as rare plants, debris, or animals where pure detection maps are too patchy.
  • Non-myopic planners able to jump featureless gaps would amplify the context signal on larger unbounded sites.
  • If context drifts strongly with depth or season, the FIFO buffer size and update threshold become critical tuning knobs that may need site-specific priors.
  • Operators could bootstrap multi-species surveys by chaining context models from one labeled image per species.

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 / 5 minor

Summary. The paper proposes one-shot, DINOv2 patch-embedding detection of a target coral species from a single labeled image, together with an online FIFO buffer of co-occurring non-target patch features that models the species’ visual environmental context. A greedy myopic planner then selects neighboring grid cells by a combination of target and context image-level scores (patch counts). Evaluation is offline trajectory replay on MetaShape-registered real AUV imagery from two 20 m × 20 m St. John reef sites and three biologically different coral species, with lawnmower, target-only, context-only, and substrata-segmentation baselines, plus ablations of initial target image and fixed vs. running context buffer. The central claim is that context-guided search samples up to ~75% of a sparsely distributed target in roughly half the exhaustive-coverage time and outperforms target-detection-only search.

Significance. If the result holds under realistic closed-loop conditions, the work is practically significant for AUV ecological surveys: it removes the need for large labeled training sets or continuous human supervision, and it supplies a denser directional signal when target detections are too sparse to guide motion. Strengths that should be credited include evaluation on real registered AUV imagery across two sites and three dissimilar species, 100-seed aggregation of greedy trials, explicit comparison to a substrata segmentation baseline (showing that learned context is not merely rock detection), and ablations of initial exemplar and fixed vs. running buffer. The one-shot DINOv2 operationalization is a concrete, deployable contribution relative to prior goal-conditioned reef survey work that relies on trained segmenters or continuous human exemplars.

major comments (4)
  1. [§III Experimental Evaluation; Fig. 5; Fig. 7] §III and Fig. 5: All reported trajectories are offline replays on a fixed, pre-collected, MetaShape-registered image lattice. The vehicle never selects a viewpoint that was not already imaged, never experiences localization/altitude error, and never updates the context buffer under closed-loop motion. The headline claim (≈75% of sparse targets in half lawnmower time; outperforming target-only) therefore rests on an experimental regime that under-tests the online adaptive buffer the abstract credits. Fig. 7 already shows Fixed vs. Running Context nearly identical at Tektite and only modestly different at Yawzi, which is consistent with this concern. Either (a) scope the claim explicitly to offline/open-loop replay and temper “autonomous surveying” language, or (b) add closed-loop hardware or at least viewpoint-selection experiments that can leave the pre-imaged lattice.
  2. [§I; §II.C; Fig. 4] Introduction formalization p(c|x) ∝ p(c|E)p(E|x) vs. §II.C operationalization: the method never estimates p(c|E); it greedily maximizes a patch-count context score from a FIFO buffer of non-target embeddings updated only when Φ_c(I) ≥ τ_c. Fig. 4 reports only a modest positive regression between target and context scores. The load-bearing hypothesis—that context supplies useful directional signal in cells with zero target detections so far—is therefore only indirectly supported by aggregate reward curves. A cell-level analysis (e.g., fraction of greedy steps taken from zero-target neighborhoods that later yield targets, vs. random) would make the causal role of context falsifiable rather than inferred from overall curves.
  3. [§IV.A; Fig. 5] §IV.A and Fig. 5 (Tektite panels): when the target distribution is multi-modal/uniform, the myopic greedy policy is trapped by local maxima and the lawnmower eventually overtakes total yield. The paper notes this but still frames the method as generally enabling efficient autonomous surveying. The central claim should be conditioned on sparsity/isolation of the target (as at Yawzi) rather than stated as a general result across sites; otherwise the Tektite outcome undercuts the abstract’s unqualified efficiency claim.
  4. [§II.A–C; §IV.B–C] §II.A–C free parameters (σ_c target/context, M, K, τ_c) are chosen by inspecting cosine-similarity distributions on the first reef site and held fixed. Sensitivity is reported only for the initial target image (Fig. 6) and fixed vs. running buffer (Fig. 7), not for thresholds or buffer size. Because the context signal is defined by these thresholds and the update gate τ_c, a brief sensitivity sweep (or cross-site re-selection) is needed to show that the Fig. 5 gains are not an artifact of site-specific tuning.
minor comments (5)
  1. [Fig. 1] Fig. 1 caption has a line break mid-word (“t arget”); clean for camera-ready.
  2. [§II.A] Notation reuses σ_c for both target and context cosine thresholds with different numeric values (0.3 vs 0.1); introduce distinct symbols (e.g., σ_t, σ_e) to avoid confusion in §II.A.
  3. [Abstract; §I] Abstract and intro say “sampling up to 75% of the target in roughly half the time”; specify that this is the Yawzi Gorgonia sparse case so readers do not over-generalize before reading §IV.A.
  4. [§V.B] Related work cites overlapping adaptive-survey papers appropriately; a short explicit contrast table (supervision required, online vs offline, species-specific vs general exploration) would help position against Karapetyan, Manderson, Todd, and Jamieson.
  5. [§VI; Abstract] Conclusion correctly flags stronger information-driven planners and non-myopic planning as future work; consider elevating one sentence of that caveat into the abstract so the evaluation scope is clear up front.

Circularity Check

0 steps flagged

Empirical robotics paper with independent ground-truth evaluation; no derivation step reduces the reported sampling gains to inputs by construction.

full rationale

The paper's load-bearing claim is experimental: one-shot DINOv2 target+context scoring plus a greedy myopic policy samples up to 75% of sparse targets in roughly half the lawnmower time, outperforming target-only and substrata baselines (Fig. 5). That claim is measured against hand-labeled ground-truth segmentation pixel area on real AUV imagery, not against the context buffer scores themselves. The modeling step p(c|x) ∝ p(c|E)p(E|x) is an explicit hypothesis operationalized by a FIFO buffer of non-target patches (Eqs. 9–11); it is not a definition that forces the reported reward curves. Thresholds (σ_c, τ_c) and buffer sizes (M=200, K=25) are fixed hyperparameters chosen on held-out images from one site, not fitted parameters renamed as predictions of the same quantity. Related-work citations to prior adaptive-survey papers by overlapping authors (Girdhar, Karapetyan, Manderson, Roy) situate the contribution but do not supply a uniqueness theorem or ansatz that forces the result; the central comparison (Target vs EC vs Target+EC vs Segmentation) is an independent ablation on the same orthomosaics. Offline replay limitations affect external validity, not circularity of the derivation. No equation equates the reported sampling fraction to a fitted input. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central efficiency claim rests on a small set of free thresholds chosen by inspection, the modeling assumption that context co-occurrence supplies a usable gradient, and the experimental scaffold of offline orthomosaic replay. No new physical entities are postulated; the invented constructs are algorithmic (context buffer, patch-count score).

free parameters (4)
  • target cosine threshold σ_c = 0.3
    Set to 0.3 by inspecting similarity distributions on held-out images from the first reef site; directly controls which patches count as target detections and therefore the image-level score Φ_c.
  • context cosine threshold σ_c = 0.1
    Set to 0.1 by the same inspection procedure; admits broader features into the context buffer.
  • context buffer size M and sample size K = M=200, K=25
    Chosen as M=200, K=25 for the 20 m × 20 m sites to avoid catastrophic forgetting; larger sites would require retuning.
  • target presence threshold τ_c for buffer update
    User-defined gate that decides when an image is strong enough to inject new context features; value not numerically reported beyond 'user-defined'.
axioms (4)
  • domain assumption Different coral species occupy distinct environmental sub-niches at meter-to-tens-of-meters scales, so visual context E co-occurs with target c sufficiently that p(c|x) ∝ p(c|E)p(E|x).
    Stated in the introduction and used to justify maximizing p(E|x) instead of p(c|x) directly; load-bearing for why context helps in zero-detection regions.
  • domain assumption The AUV maintains approximately constant altitude and the environment is approximately planar, so patch-count Φ_c(I) correlates with spatial coverage of the target.
    Section II; required for the image-level score to be a valid reward proxy.
  • domain assumption DINOv2 patch embeddings support reliable one-shot correspondence for the three coral species without fine-tuning.
    Empirically asserted after noting that off-the-shelf segmentation models fail on these species; underpins the entire detector.
  • ad hoc to paper Offline trajectory replay on MetaShape-registered orthomosaics is a faithful proxy for online AUV decision-making.
    Experimental evaluation section; all quantitative claims rest on this simulation scaffold rather than closed-loop vehicle runs.
invented entities (2)
  • online environment-context feature buffer E_c no independent evidence
    purpose: Maintains a fixed-size FIFO of non-target DINOv2 patches that co-occur with strong target detections, supplying a dense exploration signal.
    Algorithmic construct introduced in §II-C; independent evidence is only the empirical reward curves inside this paper.
  • patch-count image-level score Φ_c(I) no independent evidence
    purpose: Heuristic abundance proxy obtained by counting patches whose max cosine similarity exceeds threshold and wins the argmax class assignment.
    Defined in Eq. 8; used both as reward and as gate for buffer updates. No external validation beyond correlation plots in Fig. 4.

pith-pipeline@v1.1.0-grok45 · 16396 in / 3528 out tokens · 28772 ms · 2026-07-14T23:49:21.248995+00:00 · methodology

0 comments
read the original abstract

Autonomous underwater vehicles (AUVs) are increasingly used to survey coral reefs, yet efficiently locating specific coral species of interest remains difficult: target species are often sparsely distributed across the reef, and an AUV with limited battery life cannot afford to search everywhere. When detections of the target itself are too sparse to provide directional guidance, the robot benefits from an additional signal to decide where to look next. We propose using the visual environmental context -- the habitat features that tend to co-occur with a target species -- as that signal. Because context features are spatially denser and often vary more smoothly than target detections, we hypothesize that a reward function targeted at broader environmental context will enable adaptive planners to make better decisions on where to go next, even in regions where no target has yet been observed. Starting from a single labeled image, our method uses patch-level DINOv2 embeddings to perform one-shot detections of both the target species and its surrounding context online. We validate our approach using real imagery collected by an AUV at two reef sites in St. John, U.S. Virgin Islands, simulating the robot's motion offline. Our results demonstrate that one-shot detection combined with adaptive context modeling enables efficient autonomous surveying, sampling up to 75$\%$ of the target in roughly half the time required by exhaustive coverage when the target is sparsely distributed, and outperforming search strategies that only use target detections.

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. DINO-Explorer: Active Underwater Discovery via Ego-Motion Compensated Semantic Predictive Coding

    cs.RO 2026-04 unverdicted novelty 5.0

    DINO-Explorer uses ego-motion compensated semantic surprise from a frozen DINOv3 model and action-conditioned predictor to triage underwater events, retaining 78.8% of human-consensus events while cutting telemetry ba...

Reference graph

Works this paper leans on

31 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khali- dov, P. Fernandez, D. Haziza, F. Massa, A. El-Noubyet al., “Dinov2: Learning robust visual features without supervision,”arXiv preprint arXiv:2304.07193, 2023

    Pith/arXiv arXiv 2023

  2. [2]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Loet al., “Segment anything,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 4015–4026

    2023

  3. [3]

    The coralscapes dataset: Semantic scene understanding in coral reefs,

    J. Sauder, V . Domazetoski, G. Banc-Prandi, G. Perna, A. Meibom, and D. Tuia, “The coralscapes dataset: Semantic scene understanding in coral reefs,”arXiv preprint arXiv:2503.20000, 2025

    arXiv 2025

  4. [4]

    Coralscop: Segment any coral image on this planet,

    Z. Zheng, H. Liang, B.-S. Hua, Y . H. Wong, P. Ang, A. P. Y . Chui, and S.-K. Yeung, “Coralscop: Segment any coral image on this planet,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 28 170–28 180

    2024

  5. [5]

    Agisoft Metashape Professional (version 2.2.0),

    AgiSoft LLC, “Agisoft Metashape Professional (version 2.2.0),” 2025. [Online]. Available: http://www.agisoft.com/downloads/installer/

    2025

  6. [6]

    The hugin 1000 autonomous underwater vehicle for military applications,

    P. E. Hagen, N. Storkersen, K. Vestgard, and P. Kartvedt, “The hugin 1000 autonomous underwater vehicle for military applications,” in Oceans 2003. Celebrating the Past... Teaming Toward the Future (IEEE Cat. No. 03CH37492), vol. 2. IEEE, 2003, pp. 1141–1145

    2003

  7. [7]

    Exploring the aurora vent field: 4000 m under ice with the nui hybrid remotely operated vehicle,

    M. V . Jakuba, M. Curran, V . Naklicki, T. Isler, A. Klesh, H. J. Buenger, A. Dalpe, L. Lindzey, M. Silvia, R. Loeret al., “Exploring the aurora vent field: 4000 m under ice with the nui hybrid remotely operated vehicle,” in2024 IEEE/OES Autonomous Underwater V ehicles Sym- posium (AUV). IEEE, 2024, pp. 1–6

    2024

  8. [8]

    Coverage of known spaces: The boustrophedon cellular decomposition,

    H. Choset, “Coverage of known spaces: The boustrophedon cellular decomposition,”Autonomous Robots, vol. 9, no. 3, pp. 247–253, 2000

    2000

  9. [9]

    Automated annotation of coral reef survey images,

    O. Beijbom, P. J. Edmunds, D. I. Kline, B. G. Mitchell, and D. Krieg- man, “Automated annotation of coral reef survey images,” in2012 IEEE conference on computer vision and pattern recognition. IEEE, 2012, pp. 1170–1177

    2012

  10. [10]

    Deep learning for coral classification,

    A. Mahmood, M. Bennamoun, S. An, F. Sohel, F. Boussaid, R. Hovey, G. Kendrick, and R. B. Fisher, “Deep learning for coral classification,” inHandbook of neural computation. Elsevier, 2017, pp. 383–401

    2017

  11. [11]

    Monitoring of coral reefs using artificial intelligence: A feasible and cost-effective ap- proach,

    M. Gonzalez-Rivero, O. Beijbom, A. Rodriguez-Ramirez, D. E. Bryant, A. Ganase, Y . Gonzalez-Marrero, A. Herrera-Reveles, E. V . Kennedy, C. J. Kim, S. Lopez-Marcanoet al., “Monitoring of coral reefs using artificial intelligence: A feasible and cost-effective ap- proach,”Remote Sensing, vol. 12, no. 3, p. 489, 2020

    2020

  12. [12]

    A new deep learning engine for coralnet,

    Q. Chen, O. Beijbom, S. Chan, J. Bouwmeester, and D. Kriegman, “A new deep learning engine for coralnet,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 3693–3702

    2021

  13. [13]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PmLR, 2021, pp. 8748–8763

    2021

  14. [14]

    Emerging properties in self-supervised vision trans- formers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision trans- formers,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 9650–9660

    2021

  15. [15]

    Upsampling dinov2 features for unsupervised vision tasks and weakly supervised materials segmentation,

    R. Docherty, A. Vamvakeros, and S. J. Cooper, “Upsampling dinov2 features for unsupervised vision tasks and weakly supervised materials segmentation,”arXiv preprint arXiv:2410.19836, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Deep vit features as dense visual descriptors,

    S. Amir, Y . Gandelsman, S. Bagon, and T. Dekel, “Deep vit features as dense visual descriptors,”arXiv preprint arXiv:2112.05814, vol. 2, no. 3, p. 4, 2021

    Pith/arXiv arXiv 2021

  17. [17]

    Dinov2 based self supervised learning for few shot medical image segmentation,

    L. Ayzenberg, R. Giryes, and H. Greenspan, “Dinov2 based self supervised learning for few shot medical image segmentation,” in2024 IEEE International Symposium on Biomedical Imaging (ISBI). IEEE, 2024, pp. 1–5

    2024

  18. [18]

    An evaluation of a pre-trained transformer-based self-distillation model (dinov2) for cross-domain plant species identification,

    C. A. Ong, F. S. Tay, Y . L. Then, and C. McCarthy, “An evaluation of a pre-trained transformer-based self-distillation model (dinov2) for cross-domain plant species identification,”Neural Computing and Applications, pp. 1–27, 2025

    2025

  19. [19]

    Multi-label classification for multi-temporal, multi-spatial coral reef condition monitoring using vision foundation model with adapter learning,

    X. Shao, H. Chen, F. Zhao, K. Magson, J. Chen, P. Li, J. Wang, and J. Sasaki, “Multi-label classification for multi-temporal, multi-spatial coral reef condition monitoring using vision foundation model with adapter learning,”arXiv preprint arXiv:2503.23012, 2025

    arXiv 2025

  20. [20]

    Unsupervised semantic segmentation by distilling feature correspon- dences,

    M. Hamilton, Z. Zhang, B. Hariharan, N. Snavely, and W. T. Freeman, “Unsupervised semantic segmentation by distilling feature correspon- dences,”arXiv preprint arXiv:2203.08414, 2022

    Pith/arXiv arXiv 2022

  21. [21]

    Human-in-the-loop segmentation of multi-species coral imagery,

    S. Raine, R. Marchant, B. Kusy, F. Maire, N. Sunderhauf, and T. Fischer, “Human-in-the-loop segmentation of multi-species coral imagery,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 2723–2732

    2024

  22. [22]

    Information-guided robotic maximum seek-and-sample in partially observable continuous environments,

    G. Flaspohler, V . Preston, A. P. Michel, Y . Girdhar, and N. Roy, “Information-guided robotic maximum seek-and-sample in partially observable continuous environments,”IEEE Robotics and Automation Letters, vol. 4, no. 4, pp. 3782–3789, 2019

    2019

  23. [23]

    Informative path planning for an autonomous underwater vehicle,

    J. Binney, A. Krause, and G. S. Sukhatme, “Informative path planning for an autonomous underwater vehicle,” in2010 IEEE International Conference on Robotics and Automation. IEEE, 2010, pp. 4791–4796

    2010

  24. [24]

    Large area 3d reconstructions from underwater surveys,

    O. Pizarro, R. Eustice, and H. Singh, “Large area 3d reconstructions from underwater surveys,” inOceans’ 04 MTS/IEEE Techno-Ocean’04 (IEEE Cat. No. 04CH37600), vol. 2. IEEE, 2004, pp. 678–687

    2004

  25. [25]

    Efficient 3d real-time adaptive auv sampling of a river plume front,

    M. O. Berild, Y . Ge, J. Eidsvik, G.-A. Fuglstad, and I. Ellingsen, “Efficient 3d real-time adaptive auv sampling of a river plume front,” Frontiers in Marine Science, vol. 10, p. 1319719, 2024

    2024

  26. [26]

    Autonomous adaptive ex- ploration using realtime online spatiotemporal topic modeling,

    Y . Girdhar, P. Giguere, and G. Dudek, “Autonomous adaptive ex- ploration using realtime online spatiotemporal topic modeling,”The International Journal of Robotics Research, vol. 33, no. 4, pp. 645– 657, 2014

    2014

  27. [27]

    Human diver-inspired visual navigation: Towards coverage path planning of shipwrecks,

    N. Karapetyan, J. V . Johnson, and I. Rekleitis, “Human diver-inspired visual navigation: Towards coverage path planning of shipwrecks,” Marine Technology Society Journal, vol. 55, no. 4, pp. 24–32, 2021

    2021

  28. [28]

    Uivnav: Underwater information-driven vision- based navigation via imitation learning,

    X. Lin, N. Karapetyan, K. Joshi, T. Liu, N. Chopra, M. Yu, P. Tokekar, and Y . Aloimonos, “Uivnav: Underwater information-driven vision- based navigation via imitation learning,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 5250–5256

    2024

  29. [29]

    Vision-based goal-conditioned policies for underwater navigation in the presence of obstacles,

    T. Manderson, J. C. G. Higuera, S. Wapnick, J.-F. Tremblay, F. Shkurti, D. Meger, and G. Dudek, “Vision-based goal-conditioned policies for underwater navigation in the presence of obstacles,”arXiv preprint arXiv:2006.16235, 2020

    Pith/arXiv arXiv 2006

  30. [30]

    Adaptive multi-altitude search and sampling of sparsely distributed natural phenomena,

    J. E. Todd, S. McCammon, Y . Girdhar, N. Roy, and D. Yoerger, “Adaptive multi-altitude search and sampling of sparsely distributed natural phenomena,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 8650–8657

    2024

  31. [31]

    Active reward learning for co- robotic vision based exploration in bandwidth limited environments,

    S. Jamieson, J. P. How, and Y . Girdhar, “Active reward learning for co- robotic vision based exploration in bandwidth limited environments,” in2020 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2020, pp. 1806–1812

    2020