Pith. sign in

REVIEW 3 major objections 5 minor 70 references

A benchmark for language-guided navigation tests goals at four semantic levels and claims human-written, contrastively verified descriptions are far more discriminative than earlier auto-generated ones (79.7% vs 55.9% matching accuracy).

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 05:25 UTC pith:WKVEJDFX

load-bearing objection A genuinely useful four-level navigation benchmark whose central quality claims lean on one VLM match metric and an abstract that promises a baseline the paper never delivers. the 3 major comments →

arxiv 2602.02220 v2 pith:WKVEJDFX submitted 2026-02-02 cs.CV cs.RO

LangMap: A Human-Verified Benchmark for Hierarchical Open-Vocabulary Goal Navigation

Bo Miao , Weijia Liu , Jun Luo , Lachlan Shinnick , Jian Liu , Thomas Hamilton-Smith , Yuhe Yang , Zijie Wu
show 3 more authors
Vanja Videnovic Feras Dayoub Anton van den Hengel
This is my paper
classification cs.CV cs.RO
keywords language-conditioned goal navigationbenchmarkhierarchical goalsinstance-level navigationhuman-verified annotationscontrastive annotationopen-vocabulary navigationembodied AI
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.

LangMap is an attempt to make evaluation of language-conditioned navigation more realistic and less noisy. The paper's central claim is that goal instructions written and cross-checked by people, with each same-category region or instance compared against its peers, are substantially more discriminative than instructions generated by vision-language models: 79.7% text-to-view matching accuracy in 5.2 words on average, against 55.9% with more than four times as many words for a previous benchmark. On top of that, the benchmark covers goals at four granularities—scene, room, region, and instance—across 414 object categories and more than 18,000 tasks, including multi-goal episodes. If the claim holds, the field gains a common testbed where an agent's ability to disambiguate 'the cup in the kitchen, not the used ones' is measured, not just its ability to find any cup.

Core claim

On the paper's own terms, LangMap is the first real-world 3D indoor navigation benchmark with human-verified annotations spanning all four goal levels. Its load-bearing result is the contrastive annotation protocol: annotators see every same-category region and every same-category instance in a scene and write one concise and one detailed description that singles out the target; every annotation is then cross-checked. The paper reports that, under a one-to-many text-to-view matching test using a large vision-language model, instance descriptions reach 79.7% accuracy—23.8 percentage points higher than a prior VLM-generated benchmark while using about a quarter of the words. The same protocol

What carries the argument

The contrastive annotation protocol. For each semantic level, annotators are given all same-category candidates (regions or object instances) in a scene, together with their object views and region panoramas, and must write two descriptions—concise and comprehensive—that distinguish the target from every other candidate. A cross-check pass filters out descriptions that are ambiguous or inaccurate. This device is what produces the paper's claimed discriminative advantage; without it, the annotations would be ordinary human labels.

Load-bearing premise

The load-bearing premise is that the text-to-view matching accuracy measured by one large vision-language model is a valid proxy for what a human would find unambiguous; the paper does not report inter-annotator agreement or human verification of the descriptions' discriminative quality.

What would settle it

Ask human raters to match the benchmark's instance descriptions and a prior VLM-generated benchmark's descriptions to candidate object views in a one-to-many protocol, and compare accuracy on the overlapping instances. If humans do not reproduce the claimed 23.8-point margin, the central quality claim fails.

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

If this is right

  • Agents can be compared on a single scale spanning exploration (scene- and room-level) and fine-grained disambiguation (region- and instance-level), with the same success metric.
  • The concise instruction style (5.3 words on average) is shown to be sufficient to uniquely identify instances, suggesting practical robot instructions need not be verbose.
  • The paper's evaluations identify concrete remaining gaps—long-tailed categories, small objects, distant targets, and full multi-goal completion—that future methods can work against.
  • Since region-level descriptions are unique to this benchmark, it provides the first measurement of whether agents can distinguish same-type rooms and their subregions.
  • Multi-goal episodes of five mixed-level instructions make sequence-level reliability (completing all goals in order) a testable quantity, and the paper reports current methods rarely achieve it.

Where Pith is reading between the lines

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

  • Because the headline quality metric is a single vision-language model, not human raters, the transfer to real users is unverified; a human study would likely be needed before trusting the claimed 23.8-point advantage.
  • The contrastive protocol could be adapted to tasks outside navigation, such as referring-expression generation in 3D scenes or instruction tuning for mobile manipulation, where distinguishing similar candidates is the core difficulty.
  • The four-level taxonomy invites a diagnostic approach: evaluating agents independently at each level could separate perception failures from spatial-reasoning failures, which the paper's ablations only begin to address.
  • The benchmark's use of real-world scans means results may depend on the particular sensor distribution; re-testing with a different embodied platform could reveal how much of the reported progress is environment-specific.

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

3 major / 5 minor

Summary. The paper introduces HieraNav, a multi-granularity open-vocabulary goal navigation task with goals at scene, room, region, and instance levels, and presents LangMap, a benchmark built on 36 HM3D-Sem validation scenes. LangMap provides human-written region labels, region descriptions, and instance descriptions for 414 object categories, along with roughly 18.5K single- and multi-goal navigation tasks. The authors describe a contrastive annotation protocol in which annotators compare same-scene regions and instances, and report that LangMap instance descriptions outperform GOAT-Bench by 23.8% in one-to-many text-to-view matching using Qwen3-VL-235B. They also evaluate five recent navigation methods (PSL, SenseAct-Mono, 3D-Mem, Uni-NaVid, MTU3D) on the new tasks, with ablations over description style, object frequency, object size, and path length.

Significance. If the benchmark and its annotations are as described, LangMap would be a useful contribution to embodied navigation research: it is the first real-world indoor navigation benchmark to combine scene-, room-, region-, and instance-level goals, it covers substantially more object categories than prior benchmarks, and it provides both concise and detailed descriptions. The task-generation rules and evaluation protocols are largely standard, and the experimental study is broad, including multi-goal sequence metrics and several failure-mode analyses. The central value of the paper, however, depends on the claim that the human-written descriptions are discriminative and human-verified. That claim is currently supported by only one automated VLM evaluation, without human-panel or inter-annotator evidence, and the abstract contains a method (PlaNaVid/BDM) that is absent from the full text. With those points addressed, this could become a standard testbed for language-conditioned navigation.

major comments (3)
  1. [§4.2, Table 3; §1 (abstract and introduction)] The headline quantitative claim — that LangMap descriptions 'outperform GOAT-Bench by 23.8% in discriminative accuracy' — rests entirely on one one-to-many text-to-view matching experiment using Qwen3-VL-235B on overlapping instances. No sample size, number of instances tested, variance, or protocol details are given. More importantly, no independent human evaluation is reported: there is no inter-annotator agreement on LangMap's descriptions, no human-panel task in which people select the referred instance from a set of views, and no manual audit of LangMap comparable to the GOAT-Bench audit in Fig. 2. Because the descriptions were written by annotators looking at the same object views used in the VLM matching, the 23.8% margin could reflect stylistic bias of the VLM (e.g., toward short attribute-dense texts) rather than true human discriminability. Since the paper labels the annotation
  2. [Abstract (arXiv metadata) vs. full text] The abstract states 'We further introduce PlaNaVid, a strong RGB-only baseline that combines Bounded Diverse Memory (BDM)...' and claims PlaNaVid achieves top-tier success rates. However, the full text — including §1 contributions, §4 experiments, Table 2, and §5 conclusion — never mentions PlaNaVid or BDM. The submitted abstract is internally inconsistent with the manuscript body. The authors must either remove PlaNaVid/BDM from the abstract or add the method and its experimental results to the paper.
  3. [§3.3, §3.4] The paper repeatedly describes annotations as 'human-verified' and states that 'a second round of cross-checking ensures quality' (§3.3) and 'All annotations are cross-checked' (§3.4), but no quantitative evidence of this verification is provided. How many annotators participated? What was the inter-annotator agreement? How often did the cross-check reject or modify a description? Without such numbers, 'human-verified' is a process description rather than an evaluated property. The lack of any human-side validation is particularly problematic because the only manual audit reported in the paper is of GOAT-Bench (Fig. 2), not of LangMap.
minor comments (5)
  1. [§4.2, Table 3] 'Non-inferiority in 94.7% of cases' is not defined in the text or caption. Please specify the exact criterion used to determine non-inferiority and how it is computed from the matching results.
  2. [Table 1 vs. Table 3] Table 1 reports LangMap concise instance descriptions average 5.3 words, while Table 3 reports an average of 5.2 words for LangMap. If the Table 3 average is over the overlapping-instance subset, this should be stated explicitly to avoid apparent inconsistency.
  3. [§4.3, Table 2] The table caption states that 'Uni-NaVid and MTU3D are evaluated twice, and the results are averaged.' Please specify what the two evaluations were (e.g., different seeds, different episode orders) and report the range or standard deviation; otherwise the averaged numbers cannot be interpreted.
  4. [Fig. 2] The text says 'manual inspection of ten scenes (around 30% of the evaluation set).' Given LangMap uses 36 scenes, 10/36 is 27.8%; if the GOAT evaluation set differs, please clarify the basis for the 30% figure.
  5. [Throughout] The names HieraNav and LangMap are used in adjacent sentences for the task and the benchmark, respectively. Please make the distinction consistent, especially in the introduction and conclusion where 'HieraNav and LangMap establish a rigorous testbed' could be read as two benchmarks.

Circularity Check

0 steps flagged

No circular reduction found: LangMap's central claims are measured externally, not derived from their own inputs.

full rationale

The paper's main quantitative claims are benchmark statistics and evaluation results, not predictions derived from fitted parameters. The annotation-quality claim (79.7% vs 55.9% discriminative accuracy, Table 3) is obtained by applying Qwen3-VL-235B to fixed, already-written descriptions from LangMap and GOAT-Bench under the same one-to-many text-to-view matching protocol. The descriptions were produced by a contrastive human annotation protocol (§3.3, §3.4) before and independently of this evaluation; they were not optimized against Qwen3-VL, and no parameter was fitted to the measured outcome. Navigation results in Table 2 are also external method evaluations on a fixed benchmark, so none of these steps reduces to its own input by construction. The only self-citation is [15], used in a general motivating statement about VLM spatial-reasoning limitations; it is not load-bearing for any derived result. The absence of an inter-annotator agreement score or an independent human panel weakens the 'human-verified / discriminative' label as a validity claim, but that is an empirical-support limitation, not circularity. Similarly, the abstract introduces PlaNaVid, but the full text and experiments do not report it; this is an internal inconsistency and completeness issue, not a circular derivation. The paper is self-contained against external benchmarks and prior datasets, so the honest finding is no significant circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The claims in this benchmark paper rest on the quality of upstream labels (HM3D-Sem), on the assumption that a VLM matching score approximates human discrimination, and on the internal consistency of the contrastive annotation process. There are no fitted physical parameters; the numbers listed are evaluator design choices. The only abstract-only entity is PlaNaVid/BDM.

free parameters (5)
  • Success radius = 1 m
    Task success is defined as STOP within 1 m of the target (§3.1, Eq. 1). Chosen from prior protocols, not fitted, but it directly sets measured SR.
  • Episode horizon = 500 steps
    Budget for each task (§3.1); affects whether long-horizon goals are feasible; hand-chosen from prior benchmarks.
  • Small-object threshold = 3.3% mean IoU
    Defines "small objects" in Table 6; threshold adopted from prior work and applied to the agent's look-up, forward, and look-down views.
  • Task start constraints = 5–30 m geodesic (relaxed to 1 m)
    Start poses are sampled so the nearest goal is 5–30 m away (§3.5); controls the distance distribution in Fig. 6(b) and hence task difficulty.
  • Head/long-tail split = top 20% vs remaining 80%
    Pareto split in Table 5; an arbitrary grouping that affects the reported head/long-tail performance gap.
axioms (5)
  • domain assumption HM3D-Sem object instances, categories, and region-object mappings are accurate enough to define navigation targets.
    All tasks and success conditions inherit labels from HM3D-Sem [10] (§3.2, §3.5); errors in these labels propagate to the benchmark.
  • domain assumption The Qwen3-VL-235B text-to-view match is a valid proxy for human description discriminability.
    The headline quality comparison (§4.2, Table 3) uses one VLM; no human agreement study is reported.
  • domain assumption Human annotators following the contrastive protocol produce consistent, scene-specific descriptions.
    The "human-verified" claim rests on an internal process (§3.3–3.4); no inter-annotator agreement or verification statistics are given.
  • domain assumption Habitat simulator with the Stretch robot model faithfully represents real indoor navigation.
    Evaluations and success metrics assume the simulator (§3.1); sim-to-real gaps are not assessed in this paper.
  • domain assumption The manual GOAT-Bench error analysis of ten scenes is representative of the whole benchmark.
    Figure 2 reports 39.8% erroneous/ambiguous instructions from ~30% of the evaluation set without sampling details.
invented entities (1)
  • PlaNaVid / Bounded Diverse Memory (BDM) no independent evidence
    purpose: Abstract-only RGB baseline claimed to reach top-tier success without depth, maps, or object masks.
    Named in the abstract but absent from §4.3, Table 2, and the reference list; no architecture, training, or result is given in the full text, so there is no falsifiable handle.

pith-pipeline@v1.3.0-alltime-deepseek · 16240 in / 17509 out tokens · 158401 ms · 2026-08-03T05:25:15.332601+00:00 · methodology

0 comments
read the original abstract

Language-conditioned goal navigation (LGN) requires agents to locate user-specified targets without step-by-step guidance. However, existing benchmarks largely focus on category-level goals or rely on instance descriptions generated by vision-language models (VLMs), which often contain ambiguities and semantic errors, limiting systematic and reliable evaluation. We introduce HieraNav, an open-vocabulary LGN task with goals specified at four hierarchical semantic levels: scene, room, region, and instance. To this end, we present Language as a Map (LangMap), to our knowledge the first real-world 3D indoor navigation benchmark with human-verified semantic annotations to support tasks across all four goal levels. LangMap provides region labels and discriminative region and instance descriptions covering 414 object categories, produced through a rigorous contrastive annotation protocol comparing same-scene regions and instances, and contains over 18K tasks. Each target is paired with concise and detailed descriptions, enabling evaluation across instruction styles. Quantitative and qualitative analyses validate our annotation quality; notably, our instance descriptions outperform GOAT-Bench annotations by 23 percentage points in text-to-view matching. We further introduce PlaNaVid, a strong RGB-only baseline that combines Bounded Diverse Memory (BDM) with high-level planning to prime a reactive policy for multi-goal navigation. PlaNaVid achieves top-tier success rates without depth, 3D scene representations, or object masks. Further analysis shows that memory and richer context boost performance, while long-tailed categories, small objects, distant targets, and multi-goal completion remain open challenges. The benchmark is available at https://bo-miao.github.io/LangMap

Figures

Figures reproduced from arXiv: 2602.02220 by Anton van den Hengel, Bo Miao, Feras Dayoub, Jian Liu, Jun Luo, Lachlan Shinnick, Thomas Hamilton-Smith, Vanja Videnovic, Weijia Liu, Yuhe Yang, Zijie Wu.

Figure 1
Figure 1. Figure 1: HieraNav requires agents to interpret natural language [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Analysis of instance-level instructions in GOAT-Bench. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: HieraNav requires agents to interpret natural language instructions and navigate to scene-, room-, region-, and instance-level [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of region labels in descending frequency. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Contrastive region annotation. Annotators are provided with region panoramas, corresponding labeled object views, and 3D [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Distribution analysis of LangMap. (a) Instance count [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Contrastive instance annotation. Using object views, region panoramas, verified region descriptions, and the 3D scene, annotators [PITH_FULL_IMAGE:figures/full_fig_p006_7.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

70 extracted references · 15 linked inside Pith

  1. [1]

    Hm3d-ovon: A dataset and benchmark for open- vocabulary object goal navigation,

    N. Yokoyama, R. Ramrakhya, A. Das, D. Batra, and S. Ha, “Hm3d-ovon: A dataset and benchmark for open- vocabulary object goal navigation,” in2024 IEEE/RSJ In- ternational Conference on Intelligent Robots and Systems (IROS), pp. 5543–5550, IEEE, 2024. 1, 2, 3, 4, 6

    2024

  2. [2]

    Habitat challenge 2023

    K. Yadav, J. Krantz, R. Ramrakhya, S. K. Ramakrish- nan, J. Yang,et al., “Habitat challenge 2023.”https: //aihabitat.org/challenge/2023/, 2023. 1, 2, 3, 4, 6, 7

    2023

  3. [3]

    Instance-specific image goal navigation: Training em- bodied agents to find object instances,

    J. Krantz, S. Lee, J. Malik, D. Batra, and D. S. Chap- lot, “Instance-specific image goal navigation: Training em- bodied agents to find object instances,”arXiv preprint arXiv:2211.15876, 2022. 1, 2, 7

    Pith/arXiv arXiv 2022

  4. [4]

    Nav- igating to objects specified by images,

    J. Krantz, T. Gervet, K. Yadav, A. Wang, C. Paxton, R. Mot- taghi, D. Batra, J. Malik, S. Lee, and D. S. Chaplot, “Nav- igating to objects specified by images,” inProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pp. 10916–10925, 2023. 1, 2

    2023

  5. [5]

    Matterport3d: Learning from rgb-d data in indoor environments,

    A. Chang, A. Dai, T. Funkhouser, M. Halber, M. Niessner, M. Savva, S. Song, A. Zeng, and Y . Zhang, “Matterport3d: Learning from rgb-d data in indoor environments,”arXiv preprint arXiv:1709.06158, 2017. 1, 2, 3, 4

    Pith/arXiv arXiv 2017

  6. [6]

    To- wards long-horizon vision-language navigation: Platform, benchmark and method,

    X. Song, W. Chen, Y . Liu, W. Chen, G. Li, and L. Lin, “To- wards long-horizon vision-language navigation: Platform, benchmark and method,” inProceedings of the Computer Vi- sion and Pattern Recognition Conference, pp. 12078–12088,

  7. [7]

    Goat-bench: A benchmark for multi- modal lifelong navigation,

    M. Khanna, R. Ramrakhya, G. Chhablani, S. Yenaman- dra, T. Gervet, M. Chang, Z. Kira, D. S. Chaplot, D. Ba- tra, and R. Mottaghi, “Goat-bench: A benchmark for multi- modal lifelong navigation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 16373–16383, 2024. 1, 2, 3, 4, 5, 6, 7

    2024

  8. [8]

    Priori- tized semantic learning for zero-shot instance navigation,

    X. Sun, L. Liu, H. Zhi, R. Qiu, and J. Liang, “Priori- tized semantic learning for zero-shot instance navigation,” inEuropean Conference on Computer Vision, pp. 161–178, Springer, 2024. 1, 2, 3, 4, 6, 7

    2024

  9. [9]

    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,” in Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track...

    2021

  10. [10]

    Habitat-matterport 3d semantics dataset,

    K. Yadav, R. Ramrakhya, S. K. Ramakrishnan, T. Gervet, J. Turner, A. Gokaslan, N. Maestre, A. X. Chang, D. Batra, M. Savva,et al., “Habitat-matterport 3d semantics dataset,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4927–4936, 2023. 1, 2, 3, 4, 5, 6

    2023

  11. [11]

    Evaluating object hallucination in large vision-language models,

    Y . Li, Y . Du, K. Zhou, J. Wang, W. X. Zhao, and J.-R. Wen, “Evaluating object hallucination in large vision-language models,”arXiv preprint arXiv:2305.10355, 2023. 1, 3

    Pith/arXiv arXiv 2023

  12. [12]

    Spatialbot: Precise spatial understanding with vision language models,

    W. Cai, I. Ponomarenko, J. Yuan, X. Li, W. Yang, H. Dong, and B. Zhao, “Spatialbot: Precise spatial understanding with vision language models,” in2025 IEEE International Con- ference on Robotics and Automation (ICRA), pp. 9490–9498, IEEE, 2025. 1, 3

    2025

  13. [13]

    Spatialvlm: Endowing vision-language mod- els with spatial reasoning capabilities,

    B. Chen, Z. Xu, S. Kirmani, B. Ichter, D. Sadigh, L. Guibas, and F. Xia, “Spatialvlm: Endowing vision-language mod- els with spatial reasoning capabilities,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 14455–14465, 2024

    2024

  14. [14]

    Regiongpt: Towards region understanding vision language model,

    Q. Guo, S. De Mello, H. Yin, W. Byeon, K. C. Cheung, Y . Yu, P. Luo, and S. Liu, “Regiongpt: Towards region understanding vision language model,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13796–13806, 2024. 3

    2024

  15. [15]

    Referring human pose and mask estimation in the wild,

    B. Miao, M. Feng, Z. Wu, M. Bennamoun, Y . Gao, and A. Mian, “Referring human pose and mask estimation in the wild,”Advances in Neural Information Processing Systems, vol. 37, pp. 44791–44813, 2024. 1

    2024

  16. [16]

    Vision-and-language navigation: Interpreting visually- grounded navigation instructions in real environments,

    P. Anderson, Q. Wu, D. Teney, J. Bruce, M. Johnson, N. S ¨underhauf, I. Reid, S. Gould, and A. Van Den Hen- gel, “Vision-and-language navigation: Interpreting visually- grounded navigation instructions in real environments,” in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 3674–3683, 2018. 2

    2018

  17. [17]

    Beyond the nav-graph: Vision-and-language navigation in continuous environments,

    J. Krantz, E. Wijmans, A. Majumdar, D. Batra, and S. Lee, “Beyond the nav-graph: Vision-and-language navigation in continuous environments,” inEuropean Conference on Com- puter Vision, pp. 104–120, Springer, 2020

    2020

  18. [18]

    Room-across-room: Multilingual vision-and-language nav- igation with dense spatiotemporal grounding,

    A. Ku, P. Anderson, R. Patel, E. Ie, and J. Baldridge, “Room-across-room: Multilingual vision-and-language nav- igation with dense spatiotemporal grounding,”arXiv preprint arXiv:2010.07954, 2020. 2

    Pith/arXiv arXiv 2010

  19. [19]

    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. Savva,et al., “On evaluation of embodied navigation agents,”arXiv preprint arXiv:1807.06757, 2018. 2

    Pith/arXiv arXiv 2018

  20. [20]

    Learning to explore using active neural slam,

    D. S. Chaplot, D. Gandhi, S. Gupta, A. Gupta, and R. Salakhutdinov, “Learning to explore using active neural slam,”ICLR, 2020. 2

    2020

  21. [21]

    The surprising effectiveness of visual odometry techniques for embodied pointgoal navigation,

    X. Zhao, H. Agrawal, D. Batra, and A. G. Schwing, “The surprising effectiveness of visual odometry techniques for embodied pointgoal navigation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 16127–16136, 2021

    2021

  22. [22]

    Is mapping necessary for real- istic pointgoal navigation?,

    R. Partsey, E. Wijmans, N. Yokoyama, O. Dobosevych, D. Batra, and O. Maksymets, “Is mapping necessary for real- istic pointgoal navigation?,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 17232–17241, 2022. 2

    2022

  23. [23]

    Object goal navigation using goal-oriented seman- tic exploration,

    D. S. Chaplot, D. P. Gandhi, A. Gupta, and R. R. Salakhut- dinov, “Object goal navigation using goal-oriented seman- tic exploration,”Advances in Neural Information Processing Systems, vol. 33, pp. 4247–4258, 2020. 2

    2020

  24. [24]

    Object- nav revisited: On evaluation of embodied agents navigating to objects,

    D. Batra, A. Gokaslan, A. Kembhavi, O. Maksymets, R. Mottaghi, M. Savva, A. Toshev, and E. Wijmans, “Object- nav revisited: On evaluation of embodied agents navigating to objects,”arXiv preprint arXiv:2006.13171, 2020. 9

    Pith/arXiv arXiv 2006

  25. [25]

    3d-aware object goal navigation via simul- taneous exploration and identification,

    J. Zhang, L. Dai, F. Meng, Q. Fan, X. Chen, K. Xu, and H. Wang, “3d-aware object goal navigation via simul- taneous exploration and identification,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 6672–6682, 2023

    2023

  26. [26]

    Cows on pasture: Baselines and benchmarks for language-driven zero-shot object navigation,

    S. Y . Gadre, M. Wortsman, G. Ilharco, L. Schmidt, and S. Song, “Cows on pasture: Baselines and benchmarks for language-driven zero-shot object navigation,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 23171–23181, 2023. 3

    2023

  27. [27]

    How to not train your dragon: Training-free embodied object goal navi- gation 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 navi- gation with semantic frontiers,” inProceedings of Robotics: Science and Systems (RSS), 2023

    2023

  28. [28]

    Mul- tion: Benchmarking semantic map memory using multi- object navigation,

    S. Wani, S. Patel, U. Jain, A. X. Chang, and M. Savva, “Mul- tion: Benchmarking semantic map memory using multi- object navigation,” inAdvances in Neural Information Pro- cessing Systems (NeurIPS), 2020. 2

    2020

  29. [29]

    Target-driven visual navigation in in- door scenes using deep reinforcement learning,

    Y . Zhu, R. Mottaghi, E. Kolve, J. J. Lim, A. Gupta, L. Fei- Fei, and A. Farhadi, “Target-driven visual navigation in in- door scenes using deep reinforcement learning,” in2017 IEEE international conference on robotics and automation (ICRA), pp. 3357–3364, IEEE, 2017. 2

    2017

  30. [30]

    Topological semantic graph memory for image-goal nav- igation,

    N. Kim, O. Kwon, H. Yoo, Y . Choi, J. Park, and S. Oh, “Topological semantic graph memory for image-goal nav- igation,” inConference on Robot Learning, pp. 393–402, PMLR, 2023

    2023

  31. [31]

    Fg- prompt: fine-grained goal prompting for image-goal naviga- tion,

    X. Sun, P. Chen, J. Fan, J. Chen, T. Li, and M. Tan, “Fg- prompt: fine-grained goal prompting for image-goal naviga- tion,”Advances in Neural Information Processing Systems, vol. 36, pp. 12054–12073, 2023. 2

    2023

  32. [32]

    Dd-ppo: Learning near- perfect pointgoal navigators from 2.5 billion frames,

    E. Wijmans, A. Kadian, A. Morcos, S. Lee, I. Essa, D. Parikh, M. Savva, and D. Batra, “Dd-ppo: Learning near- perfect pointgoal navigators from 2.5 billion frames,” inIn- ternational Conference on Learning Representations (ICLR),

  33. [33]

    Visual representations for semantic target driven navigation,

    A. Mousavian, A. Toshev, M. Fi ˇser, J. Ko ˇseck´a, A. Wahid, and J. Davidson, “Visual representations for semantic target driven navigation,” inInternational Conference on Robotics and Automation (ICRA), pp. 8846–8852, 2019

    2019

  34. [34]

    Auxiliary tasks and exploration enable objectgoal navigation,

    J. Ye, D. Batra, A. Das, and E. Wijmans, “Auxiliary tasks and exploration enable objectgoal navigation,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 16117–16126, IEEE, 2021

    2021

  35. [35]

    Vln-bert: A recurrent vision-and-language bert for naviga- tion,

    Y . Hong, Q. Wu, Y . Qi, C. Rodriguez-Opazo, and S. Gould, “Vln-bert: A recurrent vision-and-language bert for naviga- tion,” inProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pp. 1643– 1653, 2021

    2021

  36. [36]

    Pirlnav: Pretraining with imitation and rl finetuning for objectnav,

    R. Ramrakhya, D. Batra, E. Wijmans, and A. Das, “Pirlnav: Pretraining with imitation and rl finetuning for objectnav,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 17896–17906, 2023

    2023

  37. [37]

    Ver: Scaling on-policy rl leads to the emergence of navigation in embodied rear- rangement,

    E. Wijmans, I. Essa, and D. Batra, “Ver: Scaling on-policy rl leads to the emergence of navigation in embodied rear- rangement,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 35, pp. 7727–7740, 2022

    2022

  38. [38]

    Poliformer: Scaling on-policy rl with transformers results in masterful navigators,

    K.-H. Zeng, Z. Zhang, K. Ehsani, R. Hendrix, J. Salvador, A. Herrasti, R. B. Girshick, A. Kembhavi, and L. Weihs, “Poliformer: Scaling on-policy rl with transformers results in masterful navigators,”arXiv preprint arXiv:2406.20083, June 2024. 2

    Pith/arXiv arXiv 2024

  39. [39]

    Frontier semantic explo- ration for visual target navigation,

    B. Yu, H. Kasaei, and M. Cao, “Frontier semantic explo- ration for visual target navigation,” inIEEE International Conference on Robotics and Automation (ICRA), pp. 4099– 4105, 2023. 2

    2023

  40. [40]

    Vlfm: Vision-language frontier maps for zero-shot se- mantic navigation,

    N. Yokoyama, S. Ha, D. Batra, J. Wang, and B. Bucher, “Vlfm: Vision-language frontier maps for zero-shot se- mantic navigation,” inProceedings of the IEEE/RSJ Inter- national Conference on Robotics and Automation (ICRA), 2024

    2024

  41. [41]

    osmag-llm: Zero- shot open-vocabulary object navigation via semantic maps and large language models reasoning,

    F. Xie, S. Schwertfeger, and H. Blum, “osmag-llm: Zero- shot open-vocabulary object navigation via semantic maps and large language models reasoning,”IEEE Robotics and Automation Letters, vol. 11, no. 3, pp. 2426–2433, 2026

    2026

  42. [42]

    Frontiernet: Learning visual cues to explore,

    B. Sun, H. Chen, S. Leutenegger, C. Cadena, M. Pollefeys, and H. Blum, “Frontiernet: Learning visual cues to explore,” IEEE Robotics and Automation Letters, 2025. 2

    2025

  43. [43]

    Learning hierar- chical relationships for object-goal navigation,

    A. Pal, Y . Qiu, and H. I. Christensen, “Learning hierar- chical relationships for object-goal navigation,” inProceed- ings of the Conference on Robot Learning (CoRL), vol. 164, pp. 517–528, PMLR, 2021. 3

    2021

  44. [44]

    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 (CVPR), pp. 18890–18900, 2022. 3

    2022

  45. [45]

    Learning to map for active semantic goal nav- igation,

    G. Georgakis, B. Bucher, K. Schmeckpeper, S. Singh, and K. Daniilidis, “Learning to map for active semantic goal nav- igation,” inInternational Conference on Learning Represen- tations (ICLR), 2022. 3

    2022

  46. [46]

    Learning transferable vi- sual 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. Clark, G. Krueger, and I. Sutskever, “Learning transferable vi- sual models from natural language supervision,” inProceed- ings of the International Conference on Machine Learning (ICML), vol. 139, pp. 8748–8763, 2021. 3, 7

    2021

  47. [47]

    Gpt-4o system card,

    OpenAI, “Gpt-4o system card,”arXiv preprint arXiv:2410.21276, 2024

    Pith/arXiv arXiv 2024

  48. [48]

    Visual instruction tun- ing: Large language and vision assistant,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tun- ing: Large language and vision assistant,” inAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  49. [49]

    Qwen2.5-vl technical report,

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, and et al., “Qwen2.5-vl technical report,”arXiv preprint arXiv:2502.13923, 2025. 7

    Pith/arXiv arXiv 2025

  50. [50]

    Vi- cuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality,

    W.-L. Chiang, Z. Li, Z. Lin, Y . Sheng, Z. Wu, H. Zhang, L. Zheng, S. Zhuang, Y . Zhuang, J. E. Gonzalez,et al., “Vi- cuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality,”See https://vicuna. lmsys. org (accessed 14 April 2023), vol. 2, no. 3, p. 6, 2023. 3

    2023

  51. [51]

    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,” inAdvances in Neural Infor- mation Processing Systems (NeurIPS), vol. 35, 2022. 3 10

    2022

  52. [52]

    Open-vocabulary queryable scene representations for real world planning,

    B. Chen, F. Xia, B. Ichter, K. Rao, K. Gopalakrishnan, M. S. Ryoo, A. Stone, and D. Kappler, “Open-vocabulary queryable scene representations for real world planning,” arXiv preprint arXiv:2209.09874, September 2022

    Pith/arXiv arXiv 2022

  53. [53]

    Esc: Exploration with soft commonsense con- straints 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 con- straints for zero-shot object navigation,” inProceedings of the International Conference on Machine Learning (ICML), vol. 202 ofProceedings of Machine Learning Research, pp. 42829–42842, 2023

    2023

  54. [54]

    Instruct- nav: Zero-shot system for generic instruction navigation in unexplored environments,

    Y . Long, W. Cai, H. Wang, G. Zhan, and H. Dong, “Instruct- nav: Zero-shot system for generic instruction navigation in unexplored environments,” inProceedings of the Conference on Robot Learning (CoRL), vol. 270, pp. 2049–2060, 2025

    2049

  55. [55]

    Uni-navid: A video-based vision- language-action model for unifying embodied navigation tasks,

    J. Zhang, K. Wang, S. Wang, M. Li, H. Liu, S. Wei, Z. Wang, Z. Zhang, and H. Wang, “Uni-navid: A video-based vision- language-action model for unifying embodied navigation tasks,”Robotics: Science and Systems, 2025. 7, 8

    2025

  56. [56]

    Unigoal: Towards universal zero-shot goal-oriented navi- gation,

    H. Yin, X. Xu, L. Zhao, Z. Wang, J. Zhou, and J. Lu, “Unigoal: Towards universal zero-shot goal-oriented navi- gation,” inProceedings of the Computer Vision and Pattern Recognition Conference, pp. 19057–19066, 2025. 3

    2025

  57. [57]

    Robothor: An open simulation-to- real embodied ai platform,

    M. Deitke, W. Han, A. Herrasti, A. Kembhavi, E. Kolve, R. Mottaghi, J. Salvador, D. Schwenk, E. VanderBilt, M. Wallingford,et al., “Robothor: An open simulation-to- real embodied ai platform,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 3164–3174, 2020. 4

    2020

  58. [58]

    Procthor: Large-scale embodied ai using procedu- ral generation,

    M. Deitke, E. VanderBilt, A. Herrasti, L. Weihs, K. Ehsani, J. Salvador, W. Han, E. Kolve, A. Kembhavi, and R. Mot- taghi, “Procthor: Large-scale embodied ai using procedu- ral generation,”Advances in Neural Information Processing Systems, vol. 35, pp. 5982–5994, 2022. 3

    2022

  59. [59]

    Homerobot: Open-vocabulary mobile ma- nipulation,

    S. Yenamandra, A. Ramachandran, K. Yadav, A. Wang, M. Khanna, T. Gervet, T.-Y . Yang, V . Jain, A. W. Clegg, J. Turner,et al., “Homerobot: Open-vocabulary mobile ma- nipulation,”arXiv preprint arXiv:2306.11565, 2023. 3

    Pith/arXiv arXiv 2023

  60. [60]

    Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal naviga- tion,

    M. Khanna, Y . Mao, H. Jiang, S. Haresh, B. Shacklett, D. Ba- tra, A. Clegg, E. Undersander, A. X. Chang, and M. Savva, “Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal naviga- tion,” inProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pp. 16384– 1639...

    2024

  61. [61]

    The design of stretch: A compact, lightweight mobile manipulator for indoor human environments,

    C. C. Kemp, A. Edsinger, H. M. Clever, and B. Matule- vich, “The design of stretch: A compact, lightweight mobile manipulator for indoor human environments,” in2022 In- ternational Conference on Robotics and Automation (ICRA), pp. 3150–3157, IEEE, 2022. 4

    2022

  62. [62]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anad- kat, and et al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023. 5

    Pith/arXiv arXiv 2023

  63. [63]

    Sentence-bert: Sentence embeddings using siamese bert-networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,”arXiv preprint arXiv:1908.10084, 2019. 6

    Pith/arXiv arXiv 1908

  64. [64]

    3d-mem: 3d scene memory for embodied explo- ration and reasoning,

    Y . Yang, H. Yang, J. Zhou, P. Chen, H. Zhang, Y . Du, and C. Gan, “3d-mem: 3d scene memory for embodied explo- ration and reasoning,” inProceedings of the Computer Vi- sion and Pattern Recognition Conference, pp. 17294–17303,

  65. [65]

    Move to understand a 3d scene: Bridging visual grounding and exploration for ef- ficient and versatile embodied navigation,

    Z. Zhu, X. Wang, Y . Li, Z. Zhang, X. Ma, Y . Chen, B. Jia, W. Liang, Q. Yu, Z. Deng,et al., “Move to understand a 3d scene: Bridging visual grounding and exploration for ef- ficient and versatile embodied navigation,” inProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pp. 8120–8132, 2025. 6, 7, 8

    2025

  66. [66]

    Qwen3-vl technical report,

    S. Bai, Y . Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, W. Ge, Z. Guo, Q. Huang, J. Huang, F. Huang, B. Hui, S. Jiang, Z. Li, M. Li, M. Li, K. Li, Z. Lin, J. Lin, X. Liu, J. Liu, C. Liu, Y . Liu, D. Liu, S. Liu, D. Lu, R. Luo, C. Lv, R. Men, L. Meng, X. Ren, X. Ren, S. Song, Y . Sun, J. Tang, J. Tu, J. Wan, P. Wang, P. Wang,...

    Pith/arXiv arXiv 2025

  67. [67]

    Winoground: Probing vision and lan- guage models for visio-linguistic compositionality,

    T. Thrush, R. Jiang, M. Bartolo, A. Singh, A. Williams, D. Kiela, and C. Ross, “Winoground: Probing vision and lan- guage models for visio-linguistic compositionality,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5238–5248, 2022. 7

    2022

  68. [68]

    Text encoders bottleneck compositionality in contrastive vision-language models,

    A. Kamath, J. Hessel, and K.-W. Chang, “Text encoders bottleneck compositionality in contrastive vision-language models,” inProceedings of the 2023 Conference on Empir- ical Methods in Natural Language Processing, pp. 4933– 4944, 2023. 7

    2023

  69. [69]

    When and why vision-language models behave like bags-of-words, and what to do about it?,

    M. Yuksekgonul, F. Bianchi, P. Kalluri, D. Jurafsky, and J. Zou, “When and why vision-language models behave like bags-of-words, and what to do about it?,”arXiv preprint arXiv:2210.01936, 2022. 7

    Pith/arXiv arXiv 2022

  70. [70]

    Pareto,Cours d’ ´economie politique, vol

    V . Pareto,Cours d’ ´economie politique, vol. 1. Librairie Droz, 1964. 8 11

    1964