Pith. sign in

REVIEW 4 major objections 4 minor 36 references

This paper claims that zero-shot object goal navigation improves when an LLM first infers room types from observed objects, and reports 53.0% success on HM3D and 84.0% on Gibson for its hierarchical HRO framework.

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 ↗

Hierarchical LLM room-type inference plus object-room affinity scoring yields 84.0% Gibson and 53.0% HM3D success for zero-shot object-goal navigation.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A coherent hierarchical room-to-object extension of L3MVN with a plausible Gibson gain, but the headline HM3D win over L3MVN is a ~52-episode gap without error bars, so the 'significant' claim needs proper stats and code before it is believable. the 4 major comments →

arxiv 2607.13072 v1 pith:BWWCDVMR submitted 2026-07-12 cs.RO cs.AIeess.SP

HRO: Hierarchical Room-to-Object Framework for Zero-Shot Object Goal Navigation with Large Language Models

Luyuan Jia , Yinfeng Yu This is my paper
classification cs.RO cs.AIeess.SP
keywords zero-shot object goal navigationlarge language modelshierarchical reasoningroom type inferencesemantic affinity scoringfrontier-based explorationvisual navigationcommonsense reasoning
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.

The reading

HRO is an attempt to show that an agent can find objects it has never been trained on by reasoning the way a person would: first infer what kind of room it is looking at, then go to the kind of room where the target usually lives. The method uses a language model to convert objects seen near an unexplored frontier into a probability distribution over seven room types, and scores each frontier by the highest probability over the rooms associated with the target object. On the Gibson and HM3D benchmarks, this 'room-to-object' hierarchy beats flat object-to-object LLM navigation, raising success rate by about 8 to 11 points over L3MVN while keeping path efficiency competitive. The central claim is that the semantic bridge of room type is what unlocks the commonsense already stored in the LLM.

Core claim

The paper proposes to treat 'room type' as an explicit intermediate representation. For each candidate frontier, objects detected nearby are aggregated and a small language model (GPT-2) is prompted with 'A room containing [objects] is likely a [room type]' for each of seven predefined room types; the language-modeling loss is negated and softmaxed to give P(r|O). The frontier's semantic score is then max_{r in M(g)} P(r|O), where M(g) maps each target category to the set of rooms it typically occupies (e.g., bed to bedroom, tv to living room or bedroom). The best frontier is chosen by a weighted sum of this semantic score and a geometric frontier score, and path planning is executed by the

What carries the argument

The 'room-type semantic bridge' — a fixed set R of seven indoor room types (living room, bedroom, bathroom, kitchen, dining room, office, hallway) plus a hand-authored object-to-room mapping M(g). For each candidate frontier, the language model (GPT-2) scores each room type by the negative language-modeling loss of a prompt listing the observed objects, and softmax turns those scores into P(r|O). Equation (7) assigns the frontier the maximum of P(r|O) over M(g); that number is combined with a geometric frontier score to choose where to go next.

Load-bearing premise

The framework assumes that a fixed list of seven room types and a hand-authored object-to-room mapping M(g) are sufficient to localize any target, and that this mapping holds in every new environment the agent encounters.

What would settle it

Run the unmodified agent on episodes where the target's true room lies outside M(g) (for example, a toaster in a hallway or a bookshelf in a bathroom) and compare success against episodes where the room is inside M(g); if success does not drop sharply on the outside cases, then the room-mapping assumption is not the load-bearing component and the measured gains must come from elsewhere.

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

If this is right

  • Zero-shot object goal navigation can reach success rates comparable to or better than trained baselines without target-specific training or fine-tuning.
  • Structuring reasoning hierarchically—room first, object second—matters more than simply scaling the language model; HRO with GPT-2 outperforms flat methods using GPT-3.5.
  • The method's ceiling is largely set by the upstream semantic segmentation: substituting ground-truth segmentation raises HM3D success from 53.0% to 64.9%.
  • A single seven-room commonsense ontology transfers across two different scene datasets, indicating that room-level priors are a reusable navigation asset.

Where Pith is reading between the lines

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

  • Because Eq. (7) takes a maximum over a fixed room set M(g), the method cannot reward a frontier that leads to the target in an atypical room (e.g., a plant in a bathroom); replacing the max with a distribution or allowing a fallback 'anywhere with high room evidence' would test how much this bias matters.
  • The category-wise gain over L3MVN is concentrated on the six HM3D object types; extending the evaluation to open-set targets (e.g., 'cat-shaped mug') would show whether the room bridge generalizes beyond canonical home-object placement.
  • A testable extension is to let the room ontology and M(g) be generated on the fly by an LLM from the observed objects, rather than fixed in advance, which could adapt HRO to non-home indoor environments like offices or warehouses.
Share X Bluesky LinkedIn Reddit HN

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 / 4 minor

Summary. The paper proposes HRO, a training-free hierarchical framework for zero-shot object-goal navigation. HRO uses a three-level pipeline: (1) GPT-2 computes a room-type probability distribution from objects observed near each candidate frontier (Section III-D); (2) a target-oriented affinity score, Eq. (7), combines this distribution with a hand-authored object-room mapping M(g) to rank frontiers; (3) the selected frontier is reached via Fast Marching path planning. The authors report success rates of 53.0% on HM3D and 84.0% on Gibson (Table I), claiming significant improvement over existing LLM-based baselines such as L3MVN, and provide an ablation study (Table II) decomposing the contribution of room-type inference, affinity scoring, and ground-truth segmentation.

Significance. The idea of using room semantics as an intermediate 'bridge' between LLM common sense and low-level navigation is timely and plausible. The method is zero-shot and training-free, and the paper evaluates on two standard benchmarks. The hierarchical decomposition is clearly described and the inclusion of a ground-truth segmentation ablation is useful for understanding the bottleneck. However, the central empirical claim rests on single-run comparisons without error bars or statistical tests, and the key semantic mapping M(g) is only partially disclosed. If the reported gains are reproducible and statistically robust, the contribution would be a meaningful step forward for LLM-driven zero-shot navigation. The current manuscript does not yet provide sufficient evidence for that claim.

major comments (4)
  1. [Section IV-E, Table I] The headline claim that HRO 'significantly outperforms all baseline methods' is not supported by any statistical evidence. Results are reported as single runs with no error bars, number of seeds, or episode-level counts. On HM3D, HRO exceeds L3MVN by 2.6 SR points (53.0 vs 50.4), which on a 2000-episode validation set corresponds to roughly 52 episodes. This gap is within the range of run-to-run and episode-sampling noise for embodied navigation benchmarks. The authors should provide multiple seeds (or episode bootstraps), confidence intervals, and a significance test for the primary comparisons, particularly against L3MVN on both datasets.
  2. [Section IV-F, Table II] The ablation study is the only direct evidence that the proposed hierarchical modules, rather than the base LLM reasoning, drive the improvement. The reported increments are small: the addition of room-type inference changes SR from 50.2% to 51.5% (+1.3), and the semantic affinity scoring adds another 1.5 points to reach 53.0%. Without repeated runs, these deltas are statistically indistinguishable from noise. The large jump to 64.9% with ground-truth segmentation also shows that the method is highly sensitive to upstream perception, which should be discussed more carefully. The ablation needs repeated runs and significance testing before the hierarchy's specific contribution can be accepted.
  3. [Section III-E, Eq. (7)] Eq. (7) defines the affinity score as the maximum room probability over M(g), a hand-constructed object-room mapping that is only partially specified (two examples are given). If the target's actual location lies outside M(g) -- e.g., a TV in a hallway or a plant in a bathroom -- the semantic score is zero by construction, and the correct frontier is never rewarded. This is a load-bearing component of the claimed 'room-to-object' reasoning. The authors should release the full mapping for all six target categories, justify its coverage with respect to the evaluation scenes, and ideally perform a sensitivity analysis showing how performance changes with perturbations to M(g).
  4. [Section IV-B, Eqs. (1), (5), (8)] Several parameters required for independent reproduction are not disclosed: the weighting coefficient alpha in Eq. (8), the coefficients alpha and beta in Eq. (1), the window dimensions w and h in Eq. (5), the map size M, the semantic category set C, and the exact computation of the language modeling loss in Eq. (3). The paper states that the implementation is built on L3MVN but does not provide code or a full parameter listing. Given that the reported advantage over L3MVN is small, the inability to reproduce the exact setup is a substantial concern. These details should be included in the final version.
minor comments (4)
  1. [Eq. (1) and Eq. (8)] The symbol alpha is used both for the frontier geometric weights in Eq. (1) and for the semantic-geometric trade-off in Eq. (8). This is confusing; please use different symbols or explicitly state that they are distinct and report both values.
  2. [Section II and Table I] The related work discusses L-ZSON and PixNav as LLM-based zero-shot methods, but neither appears in the comparison table. If they are not compared for practical reasons, this should be stated. The phrase 'all baseline methods' in Section IV-E should be qualified accordingly.
  3. [Eq. (3)] The paper says the score is the negative language modeling loss of the prompt sequence, but it does not specify whether the loss is summed over all tokens, whether the room token is scored, or how the prompt is tokenized. Please define the loss precisely.
  4. [Section IV-A] For the Gibson dataset, the paper states that 1,000 episodes were selected from 5 scenes in the Gibson tiny split, but it does not describe the episode selection protocol or the per-scene distribution. This matters for comparability with published baseline numbers.

Circularity Check

0 steps flagged

No circularity found: HRO's performance claim is an empirical benchmark comparison; the scoring equations are design definitions, not derivations that reduce to their inputs.

full rationale

The paper's central claim is empirical: HRO achieves 53.0% SR on HM3D and 84.0% on Gibson against L3MVN and other baselines (Table I, Section IV-E). This is not a derived prediction that is logically forced by its own inputs. The method combines a hand-authored object-to-room mapping M(g) (Eq. 6), GPT-2 room-type probabilities (Eqs. 2-4), and a frontier-scoring rule (Eq. 7). These equations define a navigation heuristic; they are not fit to the reported success rates, and the empirical result could in principle contradict the heuristic's usefulness. The absence of error bars, seeds, or a statistical test is a reproducibility/statistical-robustness concern, not an instance of circularity. The reference list includes many self-citations by the authors, but they are introductory context (e.g., refs. [1]-[14]) and unrelated to the load-bearing navigation mechanism; none is invoked as a uniqueness theorem or as the justification for the main result. The method is compared against external benchmarks and prior systems rather than being validated by an internal self-citation chain. Therefore no specific step reduces to its inputs by construction, and the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

Central method depends on several undisclosed or hand-set hyperparameters (alpha, beta, M, window size) and on the assumption that seven room categories plus GPT-2 loss capture room semantics. No code or parameter files shipped. No new physical entities proposed.

free parameters (7)
  • affinity-geometric weight alpha (Eq 8)
    Weight balancing semantic affinity and frontier score in Eq 8; no value reported, and if tuned on validation it could inflate Table I.
  • Mixed Criteria weights alpha, beta (Eq 1)
    Weights balancing frontier area and distance in Eq 1; values not disclosed.
  • Room-type set R = {living room, bedroom, bathroom, kitchen, dining room, office, hallway}
    Fixed seven-category room set chosen by authors; coverage claim unverified; choice may favor these benchmarks.
  • Object-room mapping M(g) = examples: bed->{bedroom}, tv->{living room, bedroom}
    Hand-constructed commonsense knowledge base used in Eq 7; full contents not provided; governs affinity scores.
  • Frontier window size w x h
    Local window in Eq 5 not specified; affects which objects are summarized into the prompt.
  • candidate frontier count n = 4
    Set to 4; no sensitivity analysis.
  • map size M and semantic categories C
    Semantic map dimensions K x M x M; M and C not reported; needed to reproduce.
axioms (5)
  • domain assumption Indoor environments can be represented by the seven room types in R, and room type is inferable from observed objects.
    Section III-D defines R and builds prompts from objects; no evidence of coverage across HM3D/Gibson scenes.
  • domain assumption GPT-2 LM loss is a reliable zero-shot compatibility score for room type classification.
    Eq 3 defines score(ri|O) = -L_LLM(Prompt(O,ri)); no calibration or accuracy measurement for the room-inference step is reported.
  • ad hoc to paper Hand-built object-room mapping M(g) covers typical co-occurrences for all six target categories.
    Section III-E only gives M(bed)={bedroom}, M(tv)={living room, bedroom}; full mapping not listed, and Eq 7's max over M(g) fails if target is missing from M.
  • domain assumption Mask R-CNN semantic segmentation is accurate enough for the semantic map.
    Section III-C relies on Mask R-CNN predictions; Table II Row 4 shows that replacing with GT segmentation raises SR from 53.0 to 64.9, so the method's success depends heavily on this upstream assumption.
  • standard math Frontier clustering and Fast Marching Method work as cited.
    Section III-F uses FMM [28] and Mixed Criteria scoring [27] without modification; standard robotics primitives.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of HRO: Hierarchical Room-to-Object Framework for Zero-Shot Object Goal Navigation with Large Language Models." pith.science (2026). https://pith.science/paper/BWWCDVMR

@misc{pith2026260713072,
  author       = {Pith},
  title        = {Pith review of: HRO: Hierarchical Room-to-Object Framework for Zero-Shot Object Goal Navigation with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BWWCDVMR}},
  note         = {Machine review of arXiv:2607.13072}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Zero-shot object-goal navigation aims to enable an intelligent agent to explore and navigate to objects of unknown categories in an unfamiliar environment without specific target training. In zero-shot navigation tasks, pre-trained large models are usually employed to leverage their prior knowledge for guiding the agent's navigation. However, existing zero-shot object-goal navigation methods based on large language models (LLMs) merely utilize LLMs as flat reasoning tools to directly associate objects or regions. They lack the hierarchical spatial cognition modeling of human-like room semantics to object localization, which leads to strong blindness in exploration, insufficient accuracy in semantic association, and failure to fully unleash the common-sense reasoning potential of LLMs. This paper proposes an LLM-driven hierarchical room-to-object (HRO) framework for zero-shot object-goal navigation, which guides the agent to explore and navigate to the target object in a coarse-to-fine manner. Experiments on Gibson and HM3D datasets verify that our HRO framework achieves superior success rate and generalization over existing LLM-based methods, underscoring LLMs' strong potential for zero-shot object-goal navigation.

Figures

Figures reproduced from arXiv: 2607.13072 by Luyuan Jia, Yinfeng Yu.

Figure 1
Figure 1. Figure 1: The figure illustrates the HRO framework. Taking “finding plant” [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The architecture of the HRO Framework. This framework demonstrates a hierarchical “room-to-object” navigation process. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The object goal navigation experiment process for finding a chair on the Habitat platform. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of the success rate of each goal category on HM3D [PITH_FULL_IMAGE:figures/full_fig_p005_4.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

36 extracted references · 5 linked inside Pith

  1. [1]

    Sound adversarial audio-visual navigation,

    Y . Yu, W. Huang, F. Sun, C. Chen, Y . Wang, and X. Liu, “Sound adversarial audio-visual navigation,” inThe Tenth International Con- ference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022

    2022

  2. [2]

    Measuring acoustics with collaborative multiple agents,

    Y . Yu, C. Chen, L. Cao, F. Yang, and F. Sun, “Measuring acoustics with collaborative multiple agents,” inProceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, 2023, pp. 335–343

    2023

  3. [3]

    Beyond textual knowledge: Leverag- ing multimodal knowledge bases for enhancing vision-and-language navigation,

    D. Yang, Y . Yu, and L. Wang, “Beyond textual knowledge: Leverag- ing multimodal knowledge bases for enhancing vision-and-language navigation,”Information Processing & Management, vol. 63, no. 6, p. 104766, 2026

    2026

  4. [4]

    Dope: Dual object perception-enhancement network for vision-and-language navigation,

    Y . Yu and D. Yang, “Dope: Dual object perception-enhancement network for vision-and-language navigation,” inProceedings of the 2025 International Conference on Multimedia Retrieval, 2025, pp. 1739–1748

    2025

  5. [5]

    Dgfnet: End-to-end audio-visual source separation based on dynamic gating fusion,

    Y . Yu and S. Sun, “Dgfnet: End-to-end audio-visual source separation based on dynamic gating fusion,” inProceedings of the 2025 Interna- tional Conference on Multimedia Retrieval, 2025, pp. 1730–1738

    2025

  6. [6]

    Fsdenet: A frequency and spatial do- mains based detail enhancement network for remote sensing semantic segmentation,

    J. Fu, Y . Yu, and L. Wang, “Fsdenet: A frequency and spatial do- mains based detail enhancement network for remote sensing semantic segmentation,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2025

    2025

  7. [7]

    Bss-cffma: cross-domain feature fusion and multi-attention speech enhancement network based on self- supervised embedding,

    A. Mattursun, L. Wang, and Y . Yu, “Bss-cffma: cross-domain feature fusion and multi-attention speech enhancement network based on self- supervised embedding,” in2024 IEEE International Conference on Systems, Man, and Cybernetics (SMC). IEEE, 2024, pp. 3589–3594

    2024

  8. [8]

    Nonlinear regularization decoding method for speech recognition,

    J. Zhang, L. Wang, Y . Yu, and M. Xu, “Nonlinear regularization decoding method for speech recognition,”Sensors, vol. 24, no. 12, p. 3846, 2024

    2024

  9. [9]

    Audio-guided dynamic modality fusion with stereo-aware attention for audio-visual naviga- tion,

    J. Li, Y . Yu, L. Wang, F. Sun, and W. Zheng, “Audio-guided dynamic modality fusion with stereo-aware attention for audio-visual naviga- tion,” inInternational Conference on Neural Information Processing. Springer, 2025, pp. 346–359

    2025

  10. [10]

    Advancing audio- visual navigation through multi-agent collaboration in 3d environ- ments,

    H. Zhang, Y . Yu, L. Wang, F. Sun, and W. Zheng, “Advancing audio- visual navigation through multi-agent collaboration in 3d environ- ments,” inInternational Conference on Neural Information Process- ing. Springer, 2025, pp. 502–516

    2025

  11. [11]

    Iterative residual cross-attention mechanism: An inte- grated approach for audio-visual navigation tasks,

    ——, “Iterative residual cross-attention mechanism: An inte- grated approach for audio-visual navigation tasks,”arXiv preprint arXiv:2509.25652, 2025

    arXiv 2025

  12. [12]

    Dynamic multi-target fusion for efficient audio-visual navigation,

    Y . Yu, H. Zhang, and M. Zhu, “Dynamic multi-target fusion for efficient audio-visual navigation,”arXiv preprint arXiv:2509.21377, 2025

    arXiv 2025

  13. [13]

    Modality-invariant bidirec- tional temporal representation distillation network for missing multi- modal sentiment analysis,

    X. Wang, L. Wang, Y . Yu, and X. Jiao, “Modality-invariant bidirec- tional temporal representation distillation network for missing multi- modal sentiment analysis,” inICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  14. [14]

    Vnet: A gan-based multi-tier discriminator network for speech synthesis vocoders,

    Y . Cao, Y . Li, L. Wang, and Y . Yu, “Vnet: A gan-based multi-tier discriminator network for speech synthesis vocoders,” in2024 IEEE International Conference on Systems, Man, and Cybernetics (SMC). IEEE, 2024, pp. 4384–4389

    2024

  15. [15]

    Esc: Exploration with soft commonsense constraints for zero-shot object navigation,

    K. Zhou, K. Zheng, C. Pryor, Y . Shen, H. Jin, L. Getoor, and X. E. Wang, “Esc: Exploration with soft commonsense constraints for zero-shot object navigation,” inInternational Conference on Machine Learning. PMLR, 2023, pp. 42 829–42 842

    2023

  16. [16]

    L3mvn: Leveraging large language models for visual target navigation,

    B. Yu, H. Kasaei, and M. Cao, “L3mvn: Leveraging large language models for visual target navigation,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 3554–3560

    2023

  17. [17]

    Enhancing multimodal-input object goal navigation by leveraging large language models for inferring room–object relationship knowledge,

    L. Sun, A. Kanezaki, G. Caron, and Y . Yoshiyasu, “Enhancing multimodal-input object goal navigation by leveraging large language models for inferring room–object relationship knowledge,”Advanced Engineering Informatics, vol. 65, p. 103135, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S147403462500028X

    2025

  18. [18]

    Instructnav: Zero-shot system for generic instruction navigation in unexplored environment,

    Y . Long, W. Cai, H. Wang, G. Zhan, and H. Dong, “Instructnav: Zero-shot system for generic instruction navigation in unexplored environment,”arXiv preprint arXiv:2406.04882, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    Map-based modular approach for zero-shot embodied question an- swering,

    K. Sakamoto, D. Azuma, T. Miyanishi, S. Kurita, and M. Kawanabe, “Map-based modular approach for zero-shot embodied question an- swering,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 10 013–10 019

    2024

  20. [20]

    Get-zero: Graph embodiment transformer for zero-shot embodiment generalization,

    A. Patel and S. Song, “Get-zero: Graph embodiment transformer for zero-shot embodiment generalization,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 14 262–14 269

    2025

  21. [21]

    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,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 42–48

    2024

  22. [22]

    GAMap: Zero-shot object goal navigation with multi-scale geometric- affordance guidance,

    S. Yuan, H. Huang, Y . Hao, C. Wen, A. Tzes, and Y . Fang, “GAMap: Zero-shot object goal navigation with multi-scale geometric- affordance guidance,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. [Online]. Available: https://openreview.net/forum?id=IjHrALdQNP

    2024

  23. [23]

    Can an embodied agent find your “cat-shaped mug

    V . S. Dorbala, J. F. Mullen, and D. Manocha, “Can an embodied agent find your “cat-shaped mug”? llm-based zero-shot object navigation,” IEEE Robotics and Automation Letters, vol. 9, no. 5, pp. 4083–4090, 2023

    2023

  24. [24]

    Bridging zero-shot object navigation and foundation models through pixel-guided navigation skill,

    W. Cai, S. Huang, G. Cheng, Y . Long, P. Gao, C. Sun, and H. Dong, “Bridging zero-shot object navigation and foundation models through pixel-guided navigation skill,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 5228–5234

    2024

  25. [25]

    Object goal navigation using goal-oriented semantic exploration,

    D. S. Chaplot, D. P. Gandhi, A. Gupta, and R. R. Salakhutdinov, “Object goal navigation using goal-oriented semantic exploration,” Advances in Neural Information Processing Systems, vol. 33, pp. 4247–4258, 2020

    2020

  26. [26]

    Mask r-cnn,

    K. He, G. Gkioxari, P. Doll ´ar, and R. Girshick, “Mask r-cnn,” in Proceedings of the IEEE international conference on computer vision, 2017, pp. 2961–2969

    2017

  27. [27]

    His- togram based frontier exploration,

    A. Mobarhani, S. Nazari, A. H. Tamjidi, and H. D. Taghirad, “His- togram based frontier exploration,” in2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2011, pp. 1128–1133

    2011

  28. [28]

    A fast marching level set method for monotonically advancing fronts

    J. A. Sethian, “A fast marching level set method for monotonically advancing fronts.”proceedings of the National Academy of Sciences, vol. 93, no. 4, pp. 1591–1595, 1996

    1996

  29. [29]

    Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI,

    S. K. Ramakrishnan, A. Gokaslan, E. Wijmans, O. Maksymets, A. Clegg, J. M. Turner, E. Undersander, W. Galuba, A. Westbury, A. X. Chang, M. Savva, Y . Zhao, and D. Batra, “Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI,” inThirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track,...

    Pith/arXiv arXiv 2021

  30. [30]

    Gibson env: Real-world perception for embodied agents,

    F. Xia, A. R. Zamir, Z. He, A. Sax, J. Malik, and S. Savarese, “Gibson env: Real-world perception for embodied agents,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 9068–9079

    2018

  31. [31]

    3d scene graph: A structure for unified semantics, 3d space, and camera,

    I. Armeni, Z.-Y . He, J. Gwak, A. R. Zamir, M. Fischer, J. Malik, and S. Savarese, “3d scene graph: A structure for unified semantics, 3d space, and camera,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 5664–5673

    2019

  32. [32]

    Habitat: A platform for embodied ai research,

    M. Savva, A. Kadian, O. Maksymets, Y . Zhao, E. Wijmans, B. Jain, J. Straub, J. Liu, V . Koltun, J. Maliket al., “Habitat: A platform for embodied ai research,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 9339–9347

    2019

  33. [33]

    On evaluation of embodied navigation agents,

    P. Anderson, A. Chang, D. S. Chaplot, A. Dosovitskiy, S. Gupta, V . Koltun, J. Kosecka, J. Malik, R. Mottaghi, M. Savvaet al., “On evaluation of embodied navigation agents,”arXiv preprint arXiv:1807.06757, 2018

    Pith/arXiv arXiv 2018

  34. [34]

    How to not train your dragon: Training-free embodied object goal navigation with semantic frontiers,

    J. Chen, G. Li, S. Kumar, B. Ghanem, and F. Yu, “How to not train your dragon: Training-free embodied object goal navigation with semantic frontiers,”arXiv preprint arXiv:2305.16925, 2023

    Pith/arXiv arXiv 2023

  35. [35]

    V oronav: V oronoi-based zero-shot object navigation with large language model,

    P. Wu, Y . Mu, B. Wu, Y . Hou, J. Ma, S. Zhang, and C. Liu, “V oronav: V oronoi-based zero-shot object navigation with large language model,” arXiv preprint arXiv:2401.02695, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Poni: Potential functions for objectgoal navigation with interaction-free learning,

    S. K. Ramakrishnan, D. S. Chaplot, Z. Al-Halah, J. Malik, and K. Grauman, “Poni: Potential functions for objectgoal navigation with interaction-free learning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 18 890–18 900

    2022

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.