Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

SplatSearch claims a robot can find a specific object or person in an unknown indoor environment from a single reference photo taken at any viewpoint, using only a sparse online 3D map built as it moves.

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 21:55 UTC pith:LEOPS543

load-bearing objection SplatSearch is a sensible new integration of online 3DGS, diffusion inpainting, and semantic/visual frontier scoring for instance image-goal navigation, with plausible simulator gains but a load-bearing VCICN whose training is underspecified and an evidence base missing error bars and baselines. the 4 major comments →

arxiv 2511.12972 v2 pith:LEOPS543 submitted 2025-11-17 cs.RO

SplatSearch: Instance Image Goal Navigation for Mobile Robots using 3D Gaussian Splatting and Diffusion Models

Siddarth Narasimhan , Matthew Lisondra , Haitong Wang , Goldie Nejat This is my paper
classification cs.RO
keywords instance image-goal navigation3D Gaussian splattingmulti-view diffusion inpaintingnovel view synthesisfrontier-based explorationmobile robot navigationunknown environmentsfeature matching
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.

SplatSearch claims that a mobile robot can find a specific object or person in an unknown indoor environment from a single reference photo taken at an arbitrary viewpoint, without a pre-built dense map. The method builds a sparse online 3D Gaussian Splatting reconstruction as the robot moves, detects candidate objects of the same class as the goal, renders novel viewpoints around each candidate, and uses a multi-view diffusion inpainting network to fill in the unobserved regions of these renders. The completed views are then matched against the goal image with a feature matcher, and a frontier exploration policy uses both semantic and visual context to choose where to go next. In simulated photorealistic homes and real-world trials, the authors report higher success rates and path efficiencies than the compared methods, including for objects whose images were not in the training data.

Core claim

The paper's central claim is that viewpoint-invariant instance recognition does not require a dense pre-built 3D map. Instead, a sparse online 3D Gaussian Splatting map, combined with a view-consistent diffusion inpainting step, is sufficient to synthesize usable views of candidate objects and to match them to a goal image captured from a different viewpoint. The authors further claim that frontier selection should combine semantic context (an image encoder's similarity between the observation and the goal) with visual context (feature-matching scores at detected object locations), rather than using either alone, and that this fusion improves both success rate and path efficiency.

What carries the argument

The load-bearing mechanism is the synthesis-completion-matching loop. The Novel Viewpoint Synthesis Module renders several camera poses around each detected object's 3D centroid using the sparse 3D Gaussian Splatting map; the View-Consistent Image Completion Network (VCICN), a multi-view diffusion model, inpaints the missing regions across these renders to produce complete, view-consistent images; and a feature-matching module counts correspondences between each completed view and the goal image to decide whether the instance matches. The Exploration Planner then prioritizes frontiers using a weighted utility of distance, the visual context score from feature matching, and a semantic context

Load-bearing premise

The weakest premise is that the diffusion inpainting network, trained on roughly two thousand simulated object instances plus a few hundred people, produces completed novel views that preserve exactly the instance-discriminating details the feature matcher relies on; if it smooths away or fabricates those details, the viewpoint-invariance claim collapses.

What would settle it

A trial set that includes goal objects with high-frequency, instance-unique texture (e.g., a printed logo or patterned book cover) and very sparse initial views would stress the inpainting: if success drops sharply while the no-inpainting baseline also drops, that points to inpainting hallucination rather than preservation. A second check is false-positive rate: give the robot a goal image of an object that is not in the scene and measure how often the matcher still declares a match.

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

If this is right

  • A robot can begin searching immediately in an unknown environment, because the 3D map is built online and only sparsely.
  • Goal images that look very different from the robot's current viewpoint—such as bird's-eye views—can still be matched to what the robot sees from the ground.
  • Because recognition relies on geometric reconstruction plus feature matching rather than a fixed object vocabulary, novel objects outside the training set can be found.
  • Path efficiency improves because the robot can compare candidates without physically navigating to each one, and can steer toward semantically likely frontiers.

Where Pith is reading between the lines

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

  • If the inpainting model's view consistency holds, the same pipeline could extend to partially occluded objects, since the diffusion completion would need to infer occluded structure from the sparse views.
  • The semantic context score could be replaced with an open-vocabulary grounding model, potentially removing the need for a class-specific detector in the goal loop.
  • The visual context score is computed per detected object; a natural extension is to score frontiers by expected information gain about the goal's location, which the current utility function only approximates.

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. SplatSearch tackles Instance Image Goal Navigation (IIN) in unknown environments using a sparse online 3D Gaussian Splatting map. The architecture renders multiple novel viewpoints around candidate object instances, inpaints missing regions with a multi-view diffusion model (VCICN), and matches the completed views against a single goal image using EfficientLoFTR. An exploration planner fuses semantic (BLIP-2) and visual (feature-match) context scores to select frontiers. The paper reports state-of-the-art SR/SPL on HM3D-val and a viewpoint-hard variant, an ablation study of the main components, and a 25-trial real-world deployment. The central claim is that this system achieves viewpoint-invariant instance recognition without a pre-built dense map.

Significance. If the claims hold, this is a meaningful advance for IIN: it removes the pre-built dense map assumption of prior 3DGS methods while addressing arbitrary-viewpoint goal images, which learning-based methods struggle with. The architecture is clearly described, the ablation consistently supports the importance of the NVSM, VCICN, and the two context scores, and the real-world generalization test (including objects outside the training classes) is a valuable addition. The main risk lies in the VCICN training setup and the strength of the experimental evidence for the headline SR/SPL numbers.

major comments (4)
  1. [§V.A, Eq. (16)–(17)] The VCICN training data description is missing the supervision target. The dataset is collected by running SplatSearch without VCICN and storing 8 sparse renders with masks of unconstructed regions. For the diffusion objective in Eq. (17), the model must be trained to denoise latents toward some 'clean' image z_0^i. The paper never states what complete image is used as the target. If the target is the same sparse render (with missing pixels masked), the model can only learn to hallucinate category-typical texture, not to recover the actual instance's discriminative details. This is load-bearing because the FEM's feature matching (Eq. (18)) relies on the inpainted views preserving instance identity. The ablation (Table II) shows that inpainting improves SR, but not that identity is preserved. Please specify the training target and provide a held-out instance-retrieval evaluation (e.g., ma
  2. [§VI.A.2, Table I] The headline claim 'SplatSearch achieved the highest SR (0.7000, 0.6300) and SPL (0.3740, 0.2675) with respect to the SOTA methods' is not supported without variance or statistical testing. The SR margins over UniGoal are 0.01 and 0.04 on the two datasets; with 100 episodes, these differences are within the noise of a binomial proportion. Report standard errors or confidence intervals, the number of episodes per class, and significance tests (e.g., paired bootstrap or McNemar) for the SR and SPL comparisons. This is essential for the central comparative claim.
  3. [§VI.A, HM3D-val-hard construction] The description of the hard dataset is incomplete and contains a likely typo: 'elevations, r ∈ [1.0,2.0]m' uses a length unit for an angle, and the elevation range is not specified. More importantly, the text does not say whether the re-sampled goal images are rendered from the HM3D mesh, whether the goal object instance remains the same, and whether the robot start and goal positions are unchanged. Without this, the 'viewpoint-invariant' evaluation is not reproducible and the difficulty of the dataset cannot be assessed. Specify the full sampling procedure (radial distance, azimuth, elevation, camera roll/pitch perturbations) and confirm that the semantic/instance identity of the goal is preserved.
  4. [§IV.A, §IV.D, §IV.F] Several free parameters that materially affect performance are not reported: the opacity threshold τ_o and depth-error threshold τ_M in Eq. (10), the feature-match threshold τ_m in Eq. (18), the frontier utility weights α, β, δ in Eq. (19), the normalization constant C in Eq. (20), the viewpoint radius r_v, the semantic averaging radius r_val, the re-render frequency f_O, and the loss weights λ_p, λ_g in Eq. (9). These are all said to be 'tuned empirically' or 'determined through simulated experiments,' but no values or sensitivity analysis are given. Without these, the method is not reproducible and the risk of overfitting to the HM3D validation set cannot be assessed. Report the values and, ideally, a sensitivity study over the most critical ones (τ_m and α, β, δ).
minor comments (5)
  1. [§VI.A, paragraph 1] Typo: 'elevations, r ∈ [1.0,2.0]m' — the elevation variable and its unit should be corrected (likely radians or degrees, not meters). Also, the azimuth range '(0, π]' should be clarified as to whether negative azimuths are excluded.
  2. [§IV.A.1, Eq. (8)] In Eq. (8), the depth prediction is denoted D̂_k, but in Eqs. (4)–(6) the rendered images use index t (e.g., D̂_t). Please make the subscript consistent.
  3. [§VI.C] Typo: 'SplatSearch achived an SR' → 'achieved'.
  4. [§V.A] The dataset description says 'six object classes (TV, sofa, chair, table, plant, monitor) [8]' but does not state whether the EgoBody person images are processed in the same way (masks, viewpoints) or whether the text prompt for VCICN is derived from the class label. Please clarify.
  5. [§VI.B, Table II] The ablation rows are informative, but the table would benefit from explicit indication of which components are removed (e.g., checkmarks) to avoid confusion. Also, the differences among some variants (e.g., w/o VCS and full model on HM3D-val) are small; consider reporting effect sizes or episode-level confidence intervals.

Circularity Check

0 steps flagged

No significant circularity: the reported SR/SPL are measured on held-out validation episodes, the VCICN is evaluated through ablations and real-world trials, and the self-citations and tuned thresholds are not load-bearing reductions.

full rationale

Walking the derivation chain, none of the claimed predictions reduce to the paper's inputs by construction. The pipeline is: sparse online 3DGS renders (Eq. 3-6), semantic scores from a pretrained BLIP-2 encoder (Eq. 11), novel-view synthesis around detected object centroids (NVSM), VCICN diffusion inpainting (Eq. 14-17), EfficientLoFTR feature matching (Eq. 18), and frontier selection (Eq. 19-20). Each stage is a computation on RGB-D observations and the goal image; no equation defines a predicted quantity as a fitted parameter. The feature-match threshold tau_m and frontier weights are 'determined through simulated experiments,' but the headline results are success rates over held-out HM3D-val and HM3D-val-hard episodes, not fitted constants renamed as predictions. The VCICN is trained on data generated by running SplatSearch without VCICN, which is a self-generated-data loop; however, this is a training-distribution and supervision concern, not circularity, because the trained model is then applied to new environments and the ablation shows its contribution empirically. The real-world experiments (25 trials, unseen objects) are an external check, albeit small and without baselines. The self-citations ([2], [3]) appear only as related-work references and are not load-bearing. One genuine gap is that Section V.A does not explicitly state the ground-truth complete image used as the diffusion target for VCICN training; that is a missing methodological detail and should be addressed, but it is not an equation-level circularity. Overall, the central viewpoint-invariance claim is supported by independent evaluation and does not reduce to its inputs by definition.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The architecture rests on a chain of pre-trained components (YOLOv7, MobileSAM, BLIP-2, EfficientLoFTR, MVInpainter-style diffusion) and eight unstated hyperparameters; the most fragile premise is that the VCICN, trained on a narrow object/class distribution, preserves instance identity during inpainting for unseen objects.

free parameters (8)
  • lambda_p, lambda_g (photometric/geometric loss weights) = not reported
    Weighting coefficients in Eq. (9) 'determined empirically through expert domain tuning'; values not disclosed, so the 3DGS map quality depends on unstated tuning.
  • tau_o (opacity threshold) = not reported
    Used for map update in Eq. (10), 'tuned empirically from simulated experiments to optimize for reconstruction quality'; no value given.
  • tau_m (feature-match threshold) = not reported
    Threshold for declaring an instance match in FEM (Section IV.E), 'determined through simulated experiments'; central to both success detection and visual-context scoring (Eq. 20).
  • alpha, beta, delta (frontier utility weights) = not reported
    Weights in Eq. (19) for distance, visual context, and semantic context; no values or tuning procedure given.
  • r_v (viewpoint sphere radius) = not reported
    Radius around object centroid for NVSM viewpoint sampling (Section IV.C); affects rendered view coverage.
  • r_val (semantic score averaging radius) = not reported
    Radius around frontier points for averaging value-map scores (Section IV.A.2).
  • f_O (re-render frequency) = not reported
    Number of distinct viewpoints after which NVSM re-renders an object (Section IV.B).
  • C (normalization constant) = not reported
    Normalization constant in Eq. (20) for object score; value unstated.
axioms (6)
  • domain assumption 3D Gaussian Splatting with alpha-compositing (Eqs. 3-6) faithfully reconstructs the scene from sparse RGB-D frames.
    The entire NVSM and map update (Eqs. 7-10) rely on the 3DGS map being accurate enough for novel-view rendering; read from Section IV.A.
  • domain assumption YOLOv7 and MobileSAM detect all goal-relevant objects and segment instances correctly.
    SOIM relies on these off-the-shelf detectors; if an object is missed or mis-segmented, NVSM never evaluates it.
  • domain assumption Cosine similarity of BLIP-2 image embeddings (Eq. 11) is a valid measure of semantic relevance to the goal image.
    The value map and semantic frontier scores depend on this similarity; not validated in the paper.
  • ad hoc to paper EfficientLoFTR feature matches on diffusion-inpainted images are reliable for instance-level identity verification.
    The FEM assumes the VCICN-completed renders preserve instance-discriminative details; this is the core generalization claim and is only tested on 100 sim episodes plus 25 real trials.
  • ad hoc to paper The VCICN trained on ~2,000 HM3D-train objects (six classes) and 400 EgoBody people generalizes to unseen objects and environments.
    Training data (Section V.A) is narrow; real-world generalization is claimed for backpack/laptop/person with only 25 trials.
  • domain assumption Habitat simulator with HM3D-SEM-v0.2 is representative of real-world IIN conditions.
    The main comparison and ablation are performed in this simulator; sim-to-real transfer is assumed rather than measured systematically.

pith-pipeline@v1.3.0-alltime-deepseek · 13493 in / 12354 out tokens · 108592 ms · 2026-08-03T21:55:02.934695+00:00 · methodology

0 comments
read the original abstract

The Instance Image Goal Navigation (IIN) problem requires mobile robots deployed in unknown environments to search for specific objects or people of interest using only a single reference goal image of the target. This problem can be especially challenging when: 1) the reference image is captured from an arbitrary viewpoint, and 2) the robot must operate with sparse-view scene reconstructions. In this paper, we address the IIN problem, by introducing SplatSearch, a novel architecture that leverages sparse-view 3D Gaussian Splatting (3DGS) reconstructions. SplatSearch renders multiple viewpoints around candidate objects using a sparse online 3DGS map and uses a multi-view diffusion model to complete missing regions of the rendered images, enabling robust feature matching against the goal image. A novel frontier exploration policy is introduced which uses visual context from the synthesized viewpoints with semantic context from the goal image to evaluate frontier locations, allowing the robot to prioritize frontiers that are semantically and visually relevant to the goal image. Extensive experiments in photorealistic home and real-world environments validate the higher performance of SplatSearch against current state-of-the-art methods in terms of Success Rate and Success Path Length. An ablation study confirms the design choices of SplatSearch.

Figures

Figures reproduced from arXiv: 2511.12972 by Goldie Nejat, Haitong Wang, Matthew Lisondra, Siddarth Narasimhan.

Figure 1
Figure 1. Figure 1: An overview of SplatSearch. SplatSearch builds a sparse-view 3DGS map and uses semantic and visual context scores to select frontier locations during exploration. To recognize goal images provided from arbitrary viewpoints, SplatSearch renders novel viewpoints around the objects and inpaints incomplete regions using a multi-view diffusion model [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The SplatSearch architecture, consisting of: 1) Map Generation Module (MGM) which generates a 3DGS map for photorealistic mapping and occupancy map for frontier exploration, 2) Semantic Object Identification Module (SOIM) detects if a current RGB image contains an object with the same class as the goal image, 3) Novel Viewpoint Synthesis Module (NVSM) generates multiple viewpoints around the object, 4) Vie… view at source ↗

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. AnyImageNav: Any-View Geometry for Precise Last-Meter Image-Goal Navigation

    cs.RO 2026-04 unverdicted novelty 7.0

    AnyImageNav uses a semantic-to-geometric cascade with 3D multi-view foundation models to recover precise 6-DoF poses from goal images, achieving 0.27m position error and state-of-the-art success rates on Gibson and HM...

Reference graph

Works this paper leans on

43 extracted references · 24 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Robots in Healthcare: a Scoping Review,

    A. A. Morgan et al., “Robots in Healthcare: a Scoping Review,” Curr Robot Rep, vol. 3, no. 4, pp. 271–280, Oct. 2022

    2022

  2. [2]

    MLLM - Search: A Zero -Shot Approach to Finding People Using Multimodal Large Language Models,

    A. Fung, A. H. Tan, H. Wang, B. Benhabib, and G. Nejat, “MLLM - Search: A Zero -Shot Approach to Finding People Using Multimodal Large Language Models,” Robotics, vol. 14, no. 8, p. 102, Jul. 2025

    2025

  3. [3]

    OLiVia -Nav: An Online Lifelong Vision Language Approach for Mobile Robot Social Navigation,

    S. Narasimhan, A. H. Tan, D. Choi, and G. Nejat, “OLiVia -Nav: An Online Lifelong Vision Language Approach for Mobile Robot Social Navigation,” in 2025 IEEE International Conference on Robotics and Automation (ICRA), IEEE, May 2025, pp. 9130–9137

    2025

  4. [4]

    Casper: An Assistive Kitchen Robot to Promote Aging in Place1,

    P. Bovbel and G. Nejat, “Casper: An Assistive Kitchen Robot to Promote Aging in Place1,” Journal of Medical Devices, vol. 8, no. 3, p. 030945, Sep. 2014

    2014

  5. [5]

    $π_0$: A Vision -Language-Action Flow Model for General Robot Control,

    K. Black et al., “$π_0$: A Vision -Language-Action Flow Model for General Robot Control,” Nov. 13, 2024, arXiv: arXiv:2410.24164

    Pith/arXiv arXiv 2024

  6. [6]

    TidyBot: Personalized Robot Assistance with Large Language Models,

    J. Wu et al. , “TidyBot: Personalized Robot Assistance with Large Language Models,” Auton Robot, vol. 47, no. 8, pp. 1087 –1102, Dec. 2023

    2023

  7. [7]

    NavFormer: A Transformer Architecture for Robot Target -Driven Navigation in Unknown and Dynamic Environments,

    H. Wang, A. H. Tan, and G. Nejat, “NavFormer: A Transformer Architecture for Robot Target -Driven Navigation in Unknown and Dynamic Environments,” IEEE Robot. Autom. Lett., vol. 9, no. 8, pp. 6808–6815, Aug. 2024

    2024

  8. [8]

    Instance - Specific Image Goal Navigation: Training Embodied Agents to Find Object Instances,

    J. Krantz, S. Lee, J. Malik, D. Batra, and D. S. Chaplot, “Instance - Specific Image Goal Navigation: Training Embodied Agents to Find Object Instances,” Nov. 29, 2022, arXiv: arXiv:2211.15876

    Pith/arXiv arXiv 2022

  9. [9]

    GaussNav: Gaussian Splatting for Visual Navigation,

    X. Lei, M. Wang, W. Zhou, and H. Li, “GaussNav: Gaussian Splatting for Visual Navigation,” IEEE Trans. Pattern Anal. Mach. Intell. , pp. 1–14, 2025

    2025

  10. [10]

    GOAT: GO to Any Thing,

    M. Chang et al., “GOAT: GO to Any Thing,” Nov. 10, 2023, arXiv: arXiv:2311.06430

    Pith/arXiv arXiv 2023

  11. [11]

    VLFM: Vision-Language Frontier Maps for Zero-Shot Semantic Navigation,

    N. Yokoyama, S. Ha, D. Batra, J. Wang, and B. Bucher, “VLFM: Vision-Language Frontier Maps for Zero-Shot Semantic Navigation,” in 2024 IEEE International Conference on Robotics and Automation (ICRA), Yokohama, Japan: IEEE, May 2024, pp. 42–48

    2024

  12. [12]

    Memory-Augmented Reinforcement Learning for Image-Goal Navigation,

    L. Mezghan et al., “Memory-Augmented Reinforcement Learning for Image-Goal Navigation,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , Kyoto, Japan: IEEE, Oct. 2022, pp. 3316–3323

    2022

  13. [13]

    REGNav: Room Expert Guided Image-Goal Navigation,

    P. Li, K. Wu, J. Fu, and S. Zhou, “REGNav: Room Expert Guided Image-Goal Navigation,” AAAI, vol. 39, no. 5, pp. 4860 –4868, Apr. 2025

    2025

  14. [14]

    Towards Generalization in Target-Driven Visual Navigation by Using Deep Reinforcement Learning,

    A. Devo, G. Mezzetti, G. Costante, M. L. Fravolini, and P. Valigi, “Towards Generalization in Target-Driven Visual Navigation by Using Deep Reinforcement Learning,” IEEE Trans. Robot., vol. 36, no. 5, pp. 1546–1561, Oct. 2020

    2020

  15. [15]

    Target-driven Visual Navigation in Indoor Scenes using Deep Reinforcement Learning,

    Y. Zhu et al., “Target-driven Visual Navigation in Indoor Scenes using Deep Reinforcement Learning,” 2016, arXiv

    2016

  16. [16]

    MemoNav: Working Memory Model for Visual Navigation,

    H. Li et al. , “MemoNav: Working Memory Model for Visual Navigation,” Mar. 28, 2024, arXiv: arXiv:2402.19161

    Pith/arXiv arXiv 2024

  17. [17]

    NoMaD: Goal Masked Diffusion Policies for Navigation and Exploration,

    A. Sridhar, D. Shah, C. Glossop, and S. Levine, “NoMaD: Goal Masked Diffusion Policies for Navigation and Exploration,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) , Yokohama, Japan: IEEE, May 2024, pp. 63–70

    2024

  18. [18]

    Rapid Exploration for Open -World Navigation with Latent Goal Models,

    D. Shah, B. Eysenbach, G. Kahn, N. Rhinehart, and S. Levine, “Rapid Exploration for Open -World Navigation with Latent Goal Models,” Oct. 11, 2023, arXiv: arXiv:2104.05859

    Pith/arXiv arXiv 2023

  19. [19]

    OVRL -V2: A simple state -of-art baseline for ImageNav and ObjectNav,

    K. Yadav et al. , “OVRL -V2: A simple state -of-art baseline for ImageNav and ObjectNav,” Mar. 14, 2023, arXiv: arXiv:2303.07798

    Pith/arXiv arXiv 2023

  20. [20]

    Towards Target -Driven Visual Navigation in Indoor Scenes via Generative Imitation Learning,

    Q. Wu et al. , “Towards Target -Driven Visual Navigation in Indoor Scenes via Generative Imitation Learning,” IEEE Robot. Autom. Lett., vol. 6, no. 1, pp. 175–182, Jan. 2021

    2021

  21. [21]

    Offline Visual Representation Learning for Embodied Navigation,

    K. Yadav et al., “Offline Visual Representation Learning for Embodied Navigation,” Apr. 27, 2022, arXiv: arXiv:2204.13226

    Pith/arXiv arXiv 2022

  22. [22]

    Neural Topological SLAM for Visual Navigation,

    D. S. Chaplot, R. Salakhutdinov, A. Gupta, and S. Gupta, “Neural Topological SLAM for Visual Navigation,” May 28, 2020, arXiv: arXiv:2005.12256

    Pith/arXiv arXiv 2020

  23. [23]

    Instance -aware Exploration-Verification-Exploitation for Instance ImageGoal Navigation,

    X. Lei, M. Wang, W. Zhou, L. Li, and H. Li, “Instance -aware Exploration-Verification-Exploitation for Instance ImageGoal Navigation,” Mar. 22, 2024, arXiv: arXiv:2402.17587

    Pith/arXiv arXiv 2024

  24. [24]

    ZSON: Zero -Shot Object -Goal Navigation using Multimodal Goal Embeddings,

    A. Majumdar, G. Aggarwal, B. Devnani, J. Hoffman, and D. Batra, “ZSON: Zero -Shot Object -Goal Navigation using Multimodal Goal Embeddings,” Oct. 13, 2023, arXiv: arXiv:2206.12403

    Pith/arXiv arXiv 2023

  25. [25]

    OpenIN: Open -Vocabulary Instance -Oriented Navigation in Dynamic Domestic Environments,

    Y. Tang et al. , “OpenIN: Open -Vocabulary Instance -Oriented Navigation in Dynamic Domestic Environments,” IEEE Robot. Autom. Lett., vol. 10, no. 9, pp. 9256–9263, Sep. 2025

    2025

  26. [26]

    UniGoal: Towards Universal Zero -shot Goal-oriented Navigation,

    H. Yin et al., “UniGoal: Towards Universal Zero -shot Goal-oriented Navigation,” Mar. 18, 2025, arXiv: arXiv:2503.10630

    Pith/arXiv arXiv 2025

  27. [27]

    3D Gaussian Splatting for Real -Time Radiance Field Rendering,

    B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis, “3D Gaussian Splatting for Real -Time Radiance Field Rendering,” Aug. 08, 2023, arXiv: arXiv:2308.04079

    Pith/arXiv arXiv 2023

  28. [28]

    GSplatVNM: Point-of-View Synthesis for Visual Navigation Models Using Gaussian Splatting,

    K. Honda, T. Ishita, Y. Yoshimura, and R. Yonetani, “GSplatVNM: Point-of-View Synthesis for Visual Navigation Models Using Gaussian Splatting,” Mar. 10, 2025, arXiv: arXiv:2503.05152

    Pith/arXiv arXiv 2025

  29. [29]

    Hierarchical Scoring with 3D Gaussian Splatting for Instance Image -Goal Navigation,

    Y. Deng et al., “Hierarchical Scoring with 3D Gaussian Splatting for Instance Image -Goal Navigation,” Jun. 10, 2025, arXiv: arXiv:2506.07338

    Pith/arXiv arXiv 2025

  30. [30]

    BEINGS: Bayesian Embodied Image-Goal Navigation With Gaussian Splatting,

    W. Meng, T. Wu, H. Yin, and F. Zhang, “BEINGS: Bayesian Embodied Image-Goal Navigation With Gaussian Splatting,” in 2025 IEEE International Conference on Robotics and Automation (ICRA) , Atlanta, GA, USA: IEEE, May 2025, pp. 5252–5258

    2025

  31. [31]

    SLGaussian: Fast Language Gaussian Splatting in Sparse Views,

    K. Chen et al. , “SLGaussian: Fast Language Gaussian Splatting in Sparse Views,” Dec. 11, 2024, arXiv: arXiv:2412.08331

    Pith/arXiv arXiv 2024

  32. [32]

    IGL-Nav: Incremental 3D Gaussian Localization for Image-goal Navigation,

    W. Guo et al., “IGL-Nav: Incremental 3D Gaussian Localization for Image-goal Navigation,” Aug. 01, 2025, arXiv: arXiv:2508.00823

    Pith/arXiv arXiv 2025

  33. [33]

    Learning Transferable Visual Models From Natural Language Supervision,

    A. Radford et al., “Learning Transferable Visual Models From Natural Language Supervision,” Feb. 26, 2021, arXiv: arXiv:2103.00020

    Pith/arXiv arXiv 2021

  34. [34]

    3D Gaussian Splatting in Robotics: A Survey,

    S. Zhu, G. Wang, X. Kong, D. Kong, and H. Wang, “3D Gaussian Splatting in Robotics: A Survey,” Dec. 19, 2024, arXiv: arXiv:2410.12262

    Pith/arXiv arXiv 2024

  35. [35]

    ActiveSplat: High-Fidelity Scene Reconstruction through Active Gaussian Splatting,

    Y. Li et al., “ActiveSplat: High-Fidelity Scene Reconstruction through Active Gaussian Splatting,” IEEE Robot. Autom. Lett., vol. 10, no. 8, pp. 8099–8106, Aug. 2025

    2025

  36. [36]

    SplaTAM: Splat, Track & Map 3D Gaussians for Dense RGB-D SLAM,

    N. Keetha et al., “SplaTAM: Splat, Track & Map 3D Gaussians for Dense RGB-D SLAM,” Apr. 16, 2024, arXiv: arXiv:2312.02126

    Pith/arXiv arXiv 2024

  37. [37]

    BLIP -2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models,

    J. Li, D. Li, S. Savarese, and S. Hoi, “BLIP -2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models,” Jun. 15, 2023, arXiv: arXiv:2301.12597

    Pith/arXiv arXiv 2023

  38. [38]

    YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,

    C.-Y. Wang, A. Bochkovskiy, and H. -Y. M. Liao, “YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” Jul. 06, 2022, arXiv: arXiv:2207.02696

    Pith/arXiv arXiv 2022

  39. [39]

    Faster Segment Anything: Towards Lightweight SAM for Mobile Applications,

    C. Zhang et al., “Faster Segment Anything: Towards Lightweight SAM for Mobile Applications,” Jul. 01, 2023, arXiv: arXiv:2306.14289

    Pith/arXiv arXiv 2023

  40. [40]

    MVInpainter: Learning Multi-View Consistent Inpainting to Bridge 2D and 3D Editing,

    C. Cao, C. Yu, F. Wang, X. Xue, and Y. Fu, “MVInpainter: Learning Multi-View Consistent Inpainting to Bridge 2D and 3D Editing,” Nov. 19, 2024, arXiv: arXiv:2408.08000

    Pith/arXiv arXiv 2024

  41. [41]

    Exploration of a cluttered environment using Voronoi Transform and Fast Marching,

    S. Garrido, L. Moreno, and D. Blanco, “Exploration of a cluttered environment using Voronoi Transform and Fast Marching,” Robotics and Autonomous Systems, vol. 56, no. 12, pp. 1069–1081, Dec. 2008

    2008

  42. [42]

    Habitat-Matterport 3D Semantics Dataset,

    K. Yadav et al., “Habitat-Matterport 3D Semantics Dataset,” Oct. 12, 2023, arXiv: arXiv:2210.05633

    Pith/arXiv arXiv 2023

  43. [43]

    EgoBody: Human Body Shape and Motion of Interacting People from Head-Mounted Devices,

    S. Zhang et al. , “EgoBody: Human Body Shape and Motion of Interacting People from Head-Mounted Devices,” in Computer Vision – ECCV 2022, vol. 13666, S. Avidan, G. Brostow, M. Cissé, G. M. Farinella, and T. Hassner, Eds., in Lecture Notes in Computer Science, vol. 13666. , Cham: Springer Nature Switzerland, 2022, pp. 180–200

    2022