Pith. sign in

REVIEW 4 major objections 4 minor 34 references

SkillNav writes behavioral memory onto a robot's value map to cut navigation detours and stalls in zero-shot object-goal navigation, setting new state-of-the-art SPL on three benchmarks.

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 ↗

A training-free skill layer that modifies the VLM's value map improves zero-shot object-goal navigation SPL by up to 6.0 points on MP3D and HM3D.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection SkillNav's score-map skill operators are a genuinely new idea with real within-backbone evidence, but the SOTA claim is confounded by an uncontrolled VLM switch and under-specified episode filtering. the 4 major comments →

arxiv 2607.15758 v1 pith:DUWD32TN submitted 2026-07-17 cs.RO

SkillNav: Score-Level Skill Intervention for Zero-Shot Object Goal Navigation

Ruijie Sang , Yiqun Duan , Pinhan Fu , Ruilin Wang , Wei Sui , Xianda Guo This is my paper
classification cs.RO
keywords object-goal navigationzero-shot navigationvision-language modelscuriosity value mapscore-level interventionbehavioral memorysuccess weighted by path lengthtraining-free
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

SkillNav argues that VLM-driven robots fail to find objects because they reason frame-by-frame with no memory of where they have been, what failed, or where the target was last seen. Rather than stuffing this history into text prompts, which grows token cost every step, the authors propose writing it directly onto the curiosity value map that already guides direction selection. Three score-level operators—soft scaling, lower-bound boost, and hard override—apply corrective memory at different levels of force, composed in a fixed order that makes conflicts predictable. The paper claims this training-free system sets new state-of-the-art path efficiency (SPL) on MP3D, HM3D v0.1, and HM3D v0.2, beating the prior best by up to 6.0 absolute points, and achieves the highest success rate on HM3D among zero-shot methods. The core insight is that spatial behavioral memory belongs on the map, not in the language model's context.

Core claim

The recurring, stereotyped failures of VLM-based object-goal navigation—getting stuck against obstacles, searching exhausted rooms, and overshooting a detected target—share a structural cause: the VLM sees only the current observation and lacks cross-step behavioral awareness. SkillNav's central discovery is that the curiosity value map already maintained by modern VLM navigators can be repurposed as a writable substrate on which skills inscribe behavioral memory at zero token cost. Three score-level operators act on that map: soft scaling reweights per-direction scores, lower-bound boost guarantees a high score in a radius around a confirmed target, and hard override forces specific directi

What carries the argument

The central object is the curiosity value map, a 2D grid of scores that rank candidate movement directions and is already produced by modern VLM navigators. SkillNav treats this map as a writable substrate: the Refinement Layer applies three composable score-level operators before the next heading is chosen. Soft scaling multiplies each direction's score by a per-direction weight (e.g., suppressing explored directions, amplifying exits). Lower-bound boost sets every map cell within a radius of a confirmed target to the maximum score, ensuring the agent never forgets a verified sighting. Hard override discards the scores entirely for selected directions, forcing the agent to avoid failed head

Load-bearing premise

The HM3D v0.1 results are computed after filtering episodes that the authors say have annotation or navigation-graph issues, and the paper does not demonstrate that the exact same filtered episode set was used for every baseline, so the reported SPL lead could be an artifact of asymmetric filtering.

What would settle it

Run all compared methods on three identical episode sets: the unfiltered HM3D v0.1 set, the authors' filtered set, and a randomly chosen filtered subset of equal size; if SkillNav's SPL advantage over the strongest prior method shrinks to near zero on the unfiltered set, the filtering is responsible. Additionally, sweep each skill threshold (τ_exh, τ_cov, τ_move, W) across a plausible range; if SPL drops rapidly away from the reported values, the gains are overfit to the test episodes.

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

If this is right

  • Behavioral memory for navigation no longer requires prompt expansion: because scores live on the map, memory persists across steps with zero additional VLM tokens.
  • Adding a new navigation behavior becomes a matter of registering a skill with a declared tier, rather than retraining or rewriting prompts, potentially enabling continual refinement of embodied agents.
  • The dual-representation design implies semantic confusions (e.g., mistaking a stool for a chair) are best fixed by bounded category-level prompts, while spatial inefficiencies (stalls, loops, detours) are best fixed by score-level map corrections—empirically decoupled in the ablation.
  • The gains transfer across VLM backbones: both a 7B open model and a larger proprietary model improve through the same refinement layer, with residual failures concentrated in visual grounding on ambiguous categories.
  • If the SPL gains hold under careful evaluation, SkillNav closes the cross-step behavioral-awareness gap for perceptually easy targets, shifting the remaining bottleneck to the VLM's perceptual side.

Where Pith is reading between the lines

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

  • The score-map substrate is a general idea: any VLM agent maintaining a scoring surface—frontier maps, affordance grids, candidate-action scores—could adopt the same three-tier intervention pattern, not just object-goal navigation.
  • The reported HM3D v0.1 filtering raises a comparability question that the paper should resolve by releasing the exact episode filter and evaluating all baselines on the identical subset; until then, the +2.4 SPL lead over the closest prior method on that benchmark is the least robust result.
  • The skill thresholds (τ_exh, τ_cov, τ_move, W) are given without a sensitivity analysis; measuring SPL as these parameters are swept would show whether the gains are stable or a tuning artifact.
  • If the skill-registration view holds, it points toward a maintenance schedule for embodied agents where new failure modes are patched by registering corrective skills, potentially transferring across embodiments and scenes.
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 identifies three recurrent failure modes in zero-shot VLM object-goal navigation agents (stuck against obstacles, in-room invalid search, and overshooting a detected target) and attributes them to the lack of cross-step behavioral awareness. To address this, it proposes SkillNav, a training-free framework that inserts a Refinement Layer between VLM-based score evaluation and action selection. The layer writes behavioral memory directly onto the existing curiosity value map using three score-level operators: soft scaling (Sec. 3.2, Eq. 4), lower-bound boost (Eq. 5), and hard override (Eq. 6), applied in a fixed composition order (Eq. 8). A bounded experience-prompt channel injects category-level semantic hints at constant token cost (Sec. 3.4). The paper reports new SPL state of the art on MP3D (25.5), HM3D v0.1 (39.3), and HM3D v0.2 (43.2) and the highest SR on HM3D, using Gemini-3-Flash-Preview as the backbone (Table 1). Table 2 ablates the Memory and Skill components on HM3D v0.2 with Qwen2.5-VL-7B-Instruct, and Table 3 gives a per-category breakdown for two backbones.

Significance. If the reported results hold, SkillNav makes a conceptually useful contribution: treating the curiosity value map as a writable substrate for zero-token behavioral memory is an elegant and potentially extensible mechanism, and the three-tier operator hierarchy gives a clean way to compose new skills. The paper includes useful ablations that separate semantic-memory and spatial-skill effects on the Qwen backbone, and the per-category analysis helps localize remaining bottlenecks. However, the headline SOTA claims are currently conditional: the ablation suite lacks a Gemini-3-Flash-Preview control without SkillNav, the HM3D v0.1 comparison is based on an undisclosed episode filter that may not be applied to baselines, and most of the method's thresholds are not reported or sensitivity-tested. With the missing controls and experimental details supplied, the work could be a solid systems contribution to zero-shot ObjectNav.

major comments (4)
  1. [Section 4.3, Table 2] The ablation does not include a Gemini-3-Flash-Preview baseline without the SkillNav components. Rows (a)-(c) all use Qwen2.5-VL-7B-Instruct; row (d) switches to Gemini-3-Flash-Preview while simultaneously enabling the full Memory+Skill stack. There is no row with Gemini alone, nor Gemini with Memory but without Skill. The +11.3 SPL jump between rows (c) and (d) therefore conflates the backbone change with the skill layer. Since the SOTA numbers in Table 1 are obtained with Gemini while the VLM backbones of the prior baselines (e.g., MerNav) are not reported, the central claim that the skill framework, rather than the base VLM, drives the SPL lead is unverified. A minimal control row with Gemini and no SkillNav, and ideally a Gemini+Memory row, is required.
  2. [Section 4.2, HM3D v0.1] The paper states that a subset of HM3D v0.1 episodes contain annotation or navigation-graph issues and that reported results are 'computed after filtering these problematic instances,' but it does not specify the filter rule, the number/percentage of episodes removed, or whether the same filtered episode set was used for every baseline in Table 1. The reported SPL advantage over MerNav on v0.1 is only +2.4 absolute; an asymmetric filter could easily create or inflate this gap. The filtering criterion must be described and applied uniformly to all methods, or the v0.1 SOTA claim should be withdrawn.
  3. [Sections 3.2-3.4] The method's behavior depends on many manually fixed constants and thresholds: alpha_soft=7.0, alpha_max=10, tau_exh, tau_cov, tau_move, W=2, r, L_max, and the contents of the category-hint library. The paper gives values for only a few of these, provides no selection/tuning protocol, and reports no sensitivity analysis. Because SkillNav is advertised as training-free and zero-shot, the reader cannot tell whether these constants were chosen a priori or tuned on the evaluation episodes. Please report all values and include a sensitivity sweep or a calibration on a held-out validation split.
  4. [Tables 1-3] All reported scores are single point estimates with no error bars or repeated runs. VLM decoding is stochastic, and several differences are small (e.g., +1.1 SR on HM3D v0.2 and +1.7 SR on HM3D v0.1 relative to MerNav). Without variance estimates, these differences are not statistically meaningful. Please provide means and standard deviations over at least three independent runs, or explicitly justify that the evaluation pipeline is deterministic and insensitive to VLM sampling.
minor comments (4)
  1. [Section 3.3] The 'uniqueness' claim of the composition order is stated without proof. As written, the three properties P1-P3 are more a restatement of the intended tier semantics than a nontrivial theorem. Either provide a formal proof or reword the claim to avoid implying a deep stability result.
  2. [Section 4.3 / Table 2] The failure-mode abbreviations FP and MS are used but never precisely defined. Please state the exact definitions of false-positive rate and max-steps rate, and how they are computed from the episode outcomes.
  3. [Section 3.4] The experience-prompt library is central to the Memory ablation, but the paper does not provide the prompt templates, the list of category hints, or the value of L_max. Full reproducibility requires this material in the supplementary material.
  4. [Throughout] The text uses 'training-free' and 'zero-shot' somewhat interchangeably. Table 1 lists both TF and ZS columns but does not define the distinction; please clarify the taxonomy (e.g., training-free vs. zero-shot generalization).

Circularity Check

0 steps flagged

No circular derivation: SkillNav's SPL/SR claims are external benchmark measurements; the only self-citation is non-load-bearing, and backbone/filtering issues are empirical confounds, not circular reductions.

full rationale

The paper's central claims are empirical results on MP3D and HM3D, not quantities derived from fitted parameters. The Refinement Layer operators (Eqs. 4-7) are defined independently of the reported SPL/SR numbers; thresholds such as τ_exh, τ_cov, τ_move, and W are hyperparameters, not fitted to reproduce the headline metrics or renamed as predictions. The Sec. 3.3 claim that the composition R = Hard Override ∘ Lower-Bound Boost ∘ Soft Scaling is 'the unique one' is a genuine algebraic consequence of properties (P1)-(P3), which are stated as constraints independent of the order; it is not an imported uniqueness theorem and does not presuppose the paper's empirical conclusions. The only self-citation is WMNav (co-authored by Guo and Duan), used as a base pipeline and baseline and compared against in Table 1 rather than assumed; the curiosity-map substrate is also attributed to VLFM and L3MVN, so the premise is externally supported. Two empirical confounds are noted but are not circularity: Sec. 4.2 reports HM3D v0.1 results 'after filtering these problematic instances,' which is a benchmark-comparison risk; and Table 2's SOTA row uses Gemini-3-Flash-Preview without a Gemini-no-skill control, so the VLM-backbone contribution is not decomposed. These affect interpretation of the benchmark comparisons, but no equation, fitted value, or cited theorem is shown to reduce to its own input. Hence no load-bearing circular step; the score of 2 reflects only minor self-citation and unverified threshold/backbone assumptions, not circularity.

Axiom & Free-Parameter Ledger

9 free parameters · 3 axioms · 0 invented entities

The central claim depends on a set of hand-chosen thresholds and scaling values; the paper gives values for α_max, α_soft, and W, but leaves several others unspecified. No tuning or sensitivity analysis is provided. Additionally, the framework assumes the underlying VLM navigator's value map is accessible and that modifying it will not degrade the VLM's other functions. No new physical or conceptual entities are introduced.

free parameters (9)
  • α_max = 10
    Global maximum value map score; ceiling for lower-bound boost.
  • α_soft = 7
    Ceiling for soft scaling outputs, chosen to reserve top range for higher tiers.
  • γ_k = not specified
    Per-direction multipliers for soft scaling, set by the room-exhaustion skill; values not given.
  • τ_exh = not specified
    Mean curiosity threshold for detecting room exhaustion.
  • τ_cov = not specified
    Voxel coverage threshold for room-exhaustion trigger.
  • W = 2
    Sliding window size for anti-stuck displacement monitoring.
  • τ_move = not specified
    L2 displacement threshold for triggering escape sub-routine.
  • r = not specified
    Radius of circular region used in lower-bound boost.
  • L_max = not specified
    Maximum number of experience hints injected into the prompt channel.
axioms (3)
  • domain assumption The underlying VLM navigator already maintains a curiosity value map that is readable and writable without disrupting the VLM's inference.
    The entire method is built on modifying the value map; if the map is not exposed or its modification degrades VLM behavior, the framework fails.
  • domain assumption All skills and thresholds generalize across the evaluation datasets and are not overfit to the specific test episodes.
    The paper does not provide a tuning protocol or sensitivity analysis, so the validity of the chosen thresholds is assumed.
  • domain assumption The reported baseline numbers are directly comparable to SkillNav under identical episode filtering.
    The filtering of 'problematic' episodes in HM3D v0.1 is only applied to SkillNav's results per the text, so baseline comparability is assumed.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SkillNav: Score-Level Skill Intervention for Zero-Shot Object Goal Navigation." pith.science (2026). https://pith.science/paper/DUWD32TN

@misc{pith2026260715758,
  author       = {Pith},
  title        = {Pith review of: SkillNav: Score-Level Skill Intervention for Zero-Shot Object Goal Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DUWD32TN}},
  note         = {Machine review of arXiv:2607.15758}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Vision-Language Model (VLM) agents have advanced zero-shot object-goal navigation, yet single-frame reasoning leaves them without the cross-step behavioral awareness an embodied navigator requires, producing recurring failures such as dead-end stalls, in-room loops, and circuitous approaches to detected targets. Prompt-based remedies inflate token budgets across multi-submodule episodes and still struggle to encode inherently spatial signals such as angles, map cells, and viewpoint coordinates. In this paper, we propose SkillNav, an extensible behavioral skill framework for VLM-based navigation that treats the curiosity value map already maintained by modern VLM navigators as a writable substrate on which composable skills inscribe behavioral memory at zero token cost. Skills are stratified into three tiers by their level of behavioral authority, namely soft scaling for proportional reweighting, lower-bound boost for region-level guarantees, and hard override for threshold-triggered forced actions, and cooperate across tiers under a fixed composition order that establishes a predictable, declared priority among skills. This design turns capability improvement into skill registration: new behaviors plug in without retraining the VLM or disturbing existing skills, opening a path for continual refinement. A minimal prompt channel complements the score-level skills with category-level semantic hints, yielding a dual-representation design in which spatial memory lives on the map and semantic memory in short prompts. Training-free, SkillNav establishes new state-of-the-art SPL across MP3D (25.5), HM3D v0.1 (39.3), and HM3D v0.2 (43.2), improving SPL by up to 6.0 absolute over the strongest prior method, and achieves the highest Success Rate on HM3D v0.1 (69.7) and v0.2 (75.9).

Figures

Figures reproduced from arXiv: 2607.15758 by Pinhan Fu, Ruijie Sang, Ruilin Wang, Wei Sui, Xianda Guo, Yiqun Duan.

Figure 1
Figure 1. Figure 1: Three recurring failures caused by the ab [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the SkillNav pipeline. The Predict VLM scores each heading direction, and the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Two memory paradigms for VLM-based ObjectNav. (a) Prompt-based memory appends the action history to every VLM call, inflating the token cost as the episode grows. (b) SkillNav writes behavioral memory directly onto the curios￾ity value map and refines it through score-level skill operators, achieving direct spatial interven￾tion at zero extra token cost. As discussed in Section 1, serializing the his￾tory … 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

34 extracted references · 8 linked inside Pith

  1. [1]

    arXiv preprint arXiv:1807.06757 , year=

    On evaluation of embodied navigation agents , author=. arXiv preprint arXiv:1807.06757 , year=

    Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:2006.13171 , year=

    Objectnav revisited: On evaluation of embodied agents navigating to objects , author=. arXiv preprint arXiv:2006.13171 , year=

    Pith/arXiv arXiv 2006

  3. [3]

    Advances in Neural Information Processing Systems , volume=

    Object goal navigation using goal-oriented semantic exploration , author=. Advances in Neural Information Processing Systems , volume=

  4. [4]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Habitat-web: Learning embodied object-search strategies from human demonstrations at scale , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  5. [5]

    arXiv preprint arXiv:2109.08238 , year=

    Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai , author=. arXiv preprint arXiv:2109.08238 , year=

    Pith/arXiv arXiv

  6. [6]

    arXiv preprint arXiv:1709.06158 , year=

    Matterport3d: Learning from rgb-d data in indoor environments , author=. arXiv preprint arXiv:1709.06158 , year=

    Pith/arXiv arXiv

  7. [7]

    International Conference on Learning Representations , year=

    DD-PPO: Learning Near-Perfect PointGoal Navigators from 2.5 Billion Frames , author=. International Conference on Learning Representations , year=

  8. [8]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Thda: Treasure hunt data augmentation for semantic navigation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  9. [9]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Pirlnav: Pretraining with imitation and rl finetuning for objectnav , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  10. [10]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Poni: Potential functions for objectgoal navigation with interaction-free learning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  11. [11]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Simple but effective: Clip embeddings for embodied ai , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  12. [12]

    The Eleventh International Conference on Learning Representations , year=

    ReAct: Synergizing Reasoning and Acting in Language Models , author=. The Eleventh International Conference on Learning Representations , year=

  13. [13]

    Advances in neural information processing systems , volume=

    Reflexion: Language agents with verbal reinforcement learning , author=. Advances in neural information processing systems , volume=

  14. [14]

    arXiv preprint arXiv:2305.16291 , year=

    Voyager: An open-ended embodied agent with large language models , author=. arXiv preprint arXiv:2305.16291 , year=

    Pith/arXiv arXiv

  15. [15]

    Advances in Neural Information Processing Systems , volume=

    Zson: Zero-shot object-goal navigation using multimodal goal embeddings , author=. Advances in Neural Information Processing Systems , volume=

  16. [16]

    2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Bridging zero-shot object navigation and foundation models through pixel-guided navigation skill , author=. 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2024 , organization=

    2024

  17. [17]

    European Conference on Computer Vision , pages=

    Prioritized semantic learning for zero-shot instance navigation , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  18. [18]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Imagine before go: Self-supervised generative map for object goal navigation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  19. [19]

    2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Vlfm: Vision-language frontier maps for zero-shot semantic navigation , author=. 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2024 , organization=

    2024

  20. [20]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Cows on pasture: Baselines and benchmarks for language-driven zero-shot object navigation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  21. [21]

    International Conference on Machine Learning , pages=

    Esc: Exploration with soft commonsense constraints for zero-shot object navigation , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  22. [22]

    2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    L3mvn: Leveraging large language models for visual target navigation , author=. 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2023 , organization=

    2023

  23. [23]

    International Conference on Machine Learning , pages=

    VoroNav: Voronoi-based Zero-shot Object Navigation with Large Language Model , author=. International Conference on Machine Learning , pages=. 2024 , organization=

    2024

  24. [24]

    arXiv preprint arXiv:2411.16425 , year=

    Topv-nav: Unlocking the top-view spatial reasoning potential of mllm for zero-shot object navigation , author=. arXiv preprint arXiv:2411.16425 , year=

    Pith/arXiv arXiv

  25. [25]

    Findings of the Association for Computational Linguistics: NAACL 2024 , pages=

    Openfmnav: Towards open-set zero-shot object navigation via vision-language foundation models , author=. Findings of the Association for Computational Linguistics: NAACL 2024 , pages=

    2024

  26. [26]

    Advances in neural information processing systems , volume=

    Sg-nav: Online 3d scene graph prompting for llm-based zero-shot object navigation , author=. Advances in neural information processing systems , volume=

  27. [27]

    International Conference on Neuro-symbolic Systems , pages=

    End-to-End Navigation with Vision-Language Models: Transforming Spatial Reasoning into Question-Answering , author=. International Conference on Neuro-symbolic Systems , pages=. 2025 , organization=

    2025

  28. [28]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Unigoal: Towards universal zero-shot goal-oriented navigation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  29. [29]

    Conference on Robot Learning , pages=

    InstructNav: Zero-shot System for Generic Instruction Navigation in Unexplored Environment , author=. Conference on Robot Learning , pages=. 2025 , organization=

    2025

  30. [30]

    2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Multi-floor zero-shot object navigation policy , author=. 2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2025 , organization=

    2025

  31. [31]

    arXiv preprint arXiv:2509.01364 , year=

    Toponav: Topological graphs as a key enabler for advanced object navigation , author=. arXiv preprint arXiv:2509.01364 , year=

    Pith/arXiv arXiv

  32. [32]

    2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    Wmnav: Integrating vision-language models into world models for object goal navigation , author=. 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2025 , organization=

    2025

  33. [33]

    arXiv preprint arXiv:2602.05467 , year=

    MerNav: A Highly Generalizable Memory-Execute-Review Framework for Zero-Shot Object Goal Navigation , author=. arXiv preprint arXiv:2602.05467 , year=

    Pith/arXiv arXiv

  34. [34]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Habitat-matterport 3d semantics dataset , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.