Pith. sign in

REVIEW 3 major objections 6 minor 6 cited by

TopoNav claims that a dynamic topological memory graph, built from semantic point clouds and read by a vision-language model, achieves 0.601 SR on HM3D and 0.455 SR on MP3D, surpassing training-free baselines in success and path efficiency.

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 zero-shot object navigation system that builds a text-based topological memory graph, queried by GPT-4o, reports state-of-the-art success rates of 60.1% on HM3D and 45.5% on MP3D.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Solid zero-shot ObjectNav system with a useful topological-memory idea, but the SOTA claim rests on single-run numbers with no error bars. the 3 major comments →

arxiv 2509.01364 v1 pith:IOVZU2QS submitted 2025-09-01 cs.RO

TopoNav: Topological Graphs as a Key Enabler for Advanced Object Navigation

Peiran Liu , Qiang Zhang , Daojie Peng , Lingfeng Zhang , Yihao Qin , Hang Zhou , Jun Ma , Renjing Xu
show 1 more author
Yiding Ji
This is my paper
classification cs.RO
keywords object navigationtopological memory graphzero-shot navigationvision-language modelssemantic point cloudaffordance-based planningembodied AIHM3D
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

TopoNav claims that an agent navigating to an object in an unseen indoor scene can treat a topological graph as its spatial memory: instead of storing raw pixels, it compresses what it has seen into nodes labeled with room type, detected objects, and frontier counts, plus edges that preserve which spaces connect. A vision-language model reads this growing text graph alongside the current panorama and chooses a node and direction; a scoring function over a semantic point cloud then picks the exact waypoint, balancing the VLM's advice with frontier exploration and history avoidance. On the HM3D benchmark the paper reports 0.601 success rate and 0.346 SPL, the best among training-free baselines, and on MP3D it reports 0.455 success rate with the lowest distance to goal. The claim matters because long-horizon ObjectNav often fails from fragmented memory, and this design offers a simple, parameter-light way to make navigation decisions depend on accumulated structure rather than only the current frame.

Core claim

TopoNav establishes that topological structure can serve as an evolving, actionable memory for zero-shot ObjectNav. The agent builds a semantic point cloud plus a text-based graph whose nodes carry position, room type, nearby object classes, and frontier counts, with merging rules that keep the graph compact. A prompt manager feeds the target category, the graph text, trajectory history, and current panorama to a VLM, which returns the next node, a preferred direction, and whether the target is visible. A composite affordance over navigable points—direction, node, frontier, history in exploration; direction plus semantic proximity after detection—picks the waypoint. Reported results: 0.601 S

What carries the argument

The dynamic topological memory graph—a text-readable graph whose nodes carry position, room type, nearby object classes, and frontier counts, with edges implied by traversal history and merge rules—is the central carrier of the argument. It converts ephemeral RGB-D observations into a stable, semantic abstraction that a VLM can reason over, and it feeds an affordance-based waypoint selector that combines VLM direction, node frontier density, global exploration pressure, and history avoidance into one score over the navigable point cloud.

Load-bearing premise

TopoNav's gains assume that GPT-4o always reads the text graph and panorama correctly and that GLEE never misses the target, with no recovery mechanism when either black-box model fails, and that the hand-set thresholds transfer across HM3D, MP3D, and real scenes.

What would settle it

Run TopoNav on HM3D with the topological text graph removed from the prompt while keeping the semantic point cloud and all affordances identical; if success rate stays near 0.601, the graph is not the key enabler. Alternatively, replace the VLM-selected node with a randomly chosen node and keep everything else; a negligible SR drop would show the graph reasoning is not load-bearing.

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

If this is right

  • Training-free ObjectNav can exceed a 0.6 success rate on HM3D, approaching the trained SGM baseline and surpassing all other zero-shot systems in the comparison.
  • Adding semantic memory to the map raises path efficiency: SPL improves 10.9% over VLFM, meaning fewer redundant revisits even when success rates are similar.
  • Each node attribute is load-bearing: dropping frontier count, room type, or object type each lowers SR on HM3D by roughly 5 to 10 points.
  • The system is deployable on a real quadruped robot with an on-board GPU, indicating the graph memory adds little runtime overhead beyond the perception call.
  • The combination of an open-vocabulary detector with a VLM is necessary: using either alone yields SR below 0.48, while together they reach 0.601.

Where Pith is reading between the lines

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

  • If the same graph were paired with a smaller open-weights language model and the HM3D score stayed near 0.6, it would indicate the topological memory, not the specific proprietary model, is what drives the gain; the paper does not run this test, but it is directly implied by the framework's design.
  • The text graph format is task-agnostic, so the same nodes could answer 'find the kitchen' or 'go to the bathroom' without retraining, turning ObjectNav into a general instruction-following memory system; the paper only evaluates object goals.
  • Because GLEE is open-vocabulary, the system should generalize beyond the six HM3D target classes used in the benchmark; a quick check with rare object categories would test whether the memory graph's semantics scale beyond the closed label set.
  • The hand-set thresholds (merge distance, frontier radius, obstacle safety margin) are evaluated only at their default values; a sensitivity sweep across scenes would show whether the method needs per-scene tuning or transfers robustly.
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

3 major / 6 minor

Summary. The paper proposes TopoNav, a training-free ObjectNav system that couples a semantic point-cloud map with a dynamically maintained topological memory graph. Waypoint selection is driven by GPT-4o (VLM) and GLEE (open-vocabulary detector) through a weighted affordance combination that switches between exploration and target-acquisition phases. The system is evaluated in Habitat on HM3D and MP3D, reporting SR, SPL, and DTG, plus a single qualitative real-robot demonstration. The central claim is state-of-the-art training-free performance: on HM3D, SR 0.601 vs. 0.581 for WMNav (+0.020); on MP3D, SR 0.455 vs. 0.454 (+0.001), with a claimed new DTG benchmark of 4.21. Ablations examine the contributions of VLM/GLEE and of the topological node information types.

Significance. If the reported numbers are reproducible, TopoNav is a meaningful contribution: it demonstrates that a topological memory graph can be effectively integrated with a VLM for zero-shot ObjectNav, and the benchmark results are competitive with or better than prior training-free methods. The ablations of node information are informative and support the design choices. The main limitation is that the empirical claim is not yet established statistically: the headline gains are small and reported without variance or significance testing. The method's reliance on closed black-box models and missing implementation details further weaken independent verification. With repeated-run evaluation and full prompt/parameter disclosure, the contribution could be solid.

major comments (3)
  1. [§IV.C, Table I] The central SOTA claim rests on very small differences: HM3D SR +0.020 over WMNav and MP3D SR +0.001 over WMNav. No error bars, confidence intervals, or significance tests are reported anywhere. Since the policy uses GPT-4o API calls (which are stochastic) and the evaluation can vary with episode ordering and RNG, a single benchmark pass cannot rule out that these differences are within run-to-run noise. The authors should run multiple independent trials (e.g., at least 5 seeds), report mean±std or bootstrap confidence intervals, and apply a paired significance test over episodes. This is load-bearing for the paper's core claim.
  2. [§III.C–E] Several parameters are introduced but never assigned values or ranges: r_pcd, r_topo, d_merge, δ, Δ_step, r_g, and σ. The composite affordance A(p) is described as a sum but the weights for the exploration components (A_dir, A_node, A_front, A_hist) are not given, and ε in the normalization is not defined. The full VLM prompt is also omitted. These omissions prevent independent reproduction and make it impossible to assess sensitivity to hand-set thresholds. The paper should provide a complete parameter table, the exact prompt text, and ideally release code and prompts.
  3. [§III.E, §IV.C] The framework delegates all high-level decisions to GPT-4o and GLEE, which are closed black-box models. No per-decision accuracy (e.g., how often the VLM selects the correct node or direction), no prompt-phrasing robustness study, and no failure analysis are reported. Because the benchmark numbers are generated through this pipeline, they may be contingent on the exact API version and sampling temperature. The authors should specify model versions and inference parameters, and provide at least a basic analysis of decision-level reliability or variance across prompt formulations.
minor comments (6)
  1. [Throughout] Typos and formatting issues: 'specifed' in the Introduction, 'Camparison' in Fig. 5, 'Point Clould' in Fig. 2, and 'lacks' should be 'lack' in Section II.A. Please proofread.
  2. [§III.E] The equation for the composite affordance is written in an incomplete notation: 'A(p) = (P A{dir,node,front,hist} exploration' is not a valid equation. Define the explicit summation for exploration and target-acquisition phases, and clarify whether the four components are equally weighted.
  3. [Table I] Many baseline rows have '-' for DTG, so the claim of a 'new DTG benchmark' is only among methods that report DTG. Please state this explicitly and, if possible, obtain DTG values for the missing baselines to make the comparison fair.
  4. [Fig. 5] The per-object-category comparison against InstructNav shows no error bars or statistical testing. Given the small overall gains, these per-category differences may not be significant; either add confidence intervals or describe the plot as qualitative.
  5. [References] Reference [16] lists both a CVPR venue and an arXiv identifier; [20] lacks full publication details. Also, several cited works are preprints; this should be noted consistently.
  6. [§IV.E] The real-robot experiment is a single qualitative demonstration (Fig. 2). The contribution statement says 'extensive real-world experiments,' which is not supported. Either provide quantitative real-world trials (number of episodes, success rate, path efficiency) or temper the claim.

Circularity Check

0 steps flagged

No significant circularity: TopoNav's central benchmark claims are externally evaluated and do not reduce to its own definitions, fitted parameters, or self-citations.

full rationale

The paper's central claim is the Table I comparison on HM3D and MP3D using standard metrics SR, SPL, and DTG, obtained by running the full system in Habitat. These numbers are external benchmark outcomes; none of the equations in Section III (point-cloud construction, topological node merging, or affordance scoring) encode or assume the reported success rates. The affordance function is a waypoint-ranking heuristic, not a fitted predictor of the benchmark result. The VLM does consume the topological memory map that the system itself builds, but this is a closed-loop design choice rather than a definitional equivalence: the graph does not define the benchmark outcome, and the VLM output is subsequently tested against the environment. The ablations in Tables II and III provide controlled comparisons showing that removing components changes performance, which further indicates the main result is not forced by construction. The self-citations [12] and [13] (TriHelper and Multi-floor ObjectNav) appear only as related-work examples of zero-shot LLM-based navigation; they are not used to justify TopoNav's design, not invoked as a uniqueness theorem, and not load-bearing for any derivation. No fitted parameter is renamed as a prediction, and no known result is merely relabeled as an organization. Concerns about missing error bars, stochastic GPT-4o/GLEE behavior, hand-set thresholds, and lack of code/prompts are correctness and reproducibility risks, not circularity. The paper is therefore self-contained with respect to its headline empirical comparison, and no circular step is present.

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 0 invented entities

The paper's central claim rests on many hand-set thresholds that are not reported, on the reliability of closed commercial models (GPT-4o, GLEE), and on standard robotics assumptions about floor planarity and frontier extraction. The topological memory graph is a data structure, not a new physical entity, so it is not listed as an invented entity.

free parameters (8)
  • r_pcd (voxel downsample resolution)
    Used in point cloud construction (III-C); value not reported.
  • r_topo (node attribute radius)
    Defines the neighborhood for object lists and frontier counts per node (III-D); value not reported.
  • d_merge (node merge distance)
    Controls when nodes are merged (III-D); value not reported.
  • delta (floor height tolerance)
    Defines navigable height band and obstacle selection (III-C); value not reported.
  • Delta_step (interpolation step)
    Used for bridging gaps in navigable point clouds (III-C); value not reported.
  • r_g (frontier grid resolution)
    Resolution of the grid used for frontier detection (III-C); value not reported.
  • sigma (safety threshold)
    Threshold for zeroing affordances near obstacles (III-E); value not reported.
  • Affordance weights = equal weights
    The four exploration affordances are summed with implicit equal weights (III-E); no ablation or sensitivity analysis is given.
axioms (4)
  • domain assumption The floor is roughly planar and navigable space lies within a height band around z_floor.
    Used to construct M_nav and M_obs in Section III-C; not validated per scene.
  • domain assumption Frontier detection via Yamauchi's algorithm on a discretized 2D grid produces reliable exploration targets.
    Invoked in Section III-C for M_fro; standard robotics practice but still an assumption.
  • domain assumption GPT-4o and GLEE provide reliable semantic labels, room type classification, and directional reasoning at every decision step.
    The entire waypoint selection in Section III-E depends on these closed black-box models; no failure analysis or fallback is provided.
  • standard math Camera projection and rigid body transforms are standard pinhole model operations.
    Equations 1 and 2 in Section III-C use standard projective geometry.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of TopoNav: Topological Graphs as a Key Enabler for Advanced Object Navigation." pith.science (2026). https://pith.science/paper/IOVZU2QS

@misc{pith2026250901364,
  author       = {Pith},
  title        = {Pith review of: TopoNav: Topological Graphs as a Key Enabler for Advanced Object Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IOVZU2QS}},
  note         = {Machine review of arXiv:2509.01364}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Object Navigation (ObjectNav) has made great progress with large language models (LLMs), but still faces challenges in memory management, especially in long-horizon tasks and dynamic scenes. To address this, we propose TopoNav, a new framework that leverages topological structures as spatial memory. By building and updating a topological graph that captures scene connections, adjacency, and semantic meaning, TopoNav helps agents accumulate spatial knowledge over time, retrieve key information, and reason effectively toward distant goals. Our experiments show that TopoNav achieves state-of-the-art performance on benchmark ObjectNav datasets, with higher success rates and more efficient paths. It particularly excels in diverse and complex environments, as it connects temporary visual inputs with lasting spatial understanding.

Figures

Figures reproduced from arXiv: 2509.01364 by Daojie Peng, Hang Zhou, Jun Ma, Lingfeng Zhang, Peiran Liu, Qiang Zhang, Renjing Xu, Yiding Ji, Yihao Qin.

Figure 1
Figure 1. Figure 1: Our memorization technique enables backtracking to key nodes when exploration directions prove suboptimal, enhancing navigation efficiency through historical trajectory recall. Our key insight is that topological structures inherently function as a form of compact, enduring spatial memory. Unlike pixel-level visual details that are transient and data￾heavy, topological information captures the essence of e… view at source ↗
Figure 2
Figure 2. Figure 2: Real Robot Experiment Demonstration: When prolonged hallway exploration is detected, the robot backtracks to a node associated with a room and then starts a procedure inside the room to search for chairs. spatial knowledge dynamically—thus supporting coherent reasoning about long-range navigation goals and correcting deviations from optimal paths. Our contributions are threefold: 1) We establish the theore… view at source ↗
Figure 3
Figure 3. Figure 3: Framework Overview. Our approach constructs both a semantic point cloud map and a topological memory map during navigation. A prompt manager integrates the current observation with these map representations. Finally, our affordance-driven waypoint selection strategy dynamically adapts between exploration and target acquisition phases. where x (c,j) w denotes the j-th point belonging to object class c. Thes… view at source ↗
Figure 4
Figure 4. Figure 4: The text-based topological map enables VLMs to perform high-level reasoning and memory-based decisions. 1. Node ID: a unique ID idk to distinguish nodes. 2. Position: nk = (xk, yk) in world coordinates (z￾coordinate omitted as topological reasoning operates primarily in 2D space). 3. Surrounding object classes: Ok ⊂ T identified within a rtopo radius of nk. This semantic context informs object search strat… view at source ↗
Figure 5
Figure 5. Figure 5: Camparison with baseline (InstructNav[28]) on various objects. D. Ablation Study VLM and GLEE. Both VLM and GLEE are quite important for better performance. The combination of VLM + GLEE demonstrates a higher SR and SPL compared to the individual methods only with VLM and only with GLEE (Table. II). This suggests that the integration of these two components enhances the overall success and efficiency of go… view at source ↗

discussion (0)

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

Forward citations

Cited by 6 Pith papers

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

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

    cs.RO 2026-07 conditional novelty 7.0

    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.

  2. MCNav: Memory-Aware Dynamic Cognitive Map for Zero-shot Goal-oriented Navigation

    cs.RO 2026-05 unverdicted novelty 7.0

    MCNav builds a dynamic cognitive map with goal re-validation and missed-goal re-exploration to reach state-of-the-art results on instance-level zero-shot navigation in HM3D environments.

  3. SSTG-Nav: Metric-Grounded Spatial-Semantic Topological Graphs for Reusable Object Navigation

    cs.RO 2026-08 conditional novelty 6.0

    A pre-explored metric-semantic topology with depth-grounded standoffs, multi-view fusion, and sequential verification achieves high success in repeated object navigation.

  4. OneVLA: A Unified Framework for Embodied Tasks

    cs.RO 2026-05 unverdicted novelty 6.0

    OneVLA is a unified VLA model using a shared action head and multi-stage progressive training with CoT fine-tuning that reports state-of-the-art results on both navigation and manipulation in simulation and real-world...

  5. Weather-Conditioned Branch Routing for Robust LiDAR-Radar 3D Object Detection

    cs.CV 2026-04 unverdicted novelty 6.0

    A routing framework maintains three parallel 3D feature streams for LiDAR, 4D radar, and fusion, with a lightweight router using weather prompts to dynamically weight them and auxiliary supervision to keep branches di...

  6. MerNav: A Highly Generalizable Memory-Execute-Review Framework for Zero-Shot Object Goal Navigation

    cs.CV 2026-02 unverdicted novelty 6.0

    MerNav's Memory-Execute-Review framework improves success rates in zero-shot object goal navigation by 5-8% over baselines on four datasets while outperforming both training-free and supervised methods on key benchmarks.

Reference graph

Works this paper leans on

37 extracted references · 25 canonical work pages · cited by 6 Pith papers

  1. [1]

    Efficient robotic object search via hiem: Hierar- chical policy learning with intrinsic-extrinsic modeling,

    X. Ye and Y . Yang, “Efficient robotic object search via hiem: Hierar- chical policy learning with intrinsic-extrinsic modeling,”IEEE robotics and automation letters , vol. 6, no. 3, pp. 4425–4432, 2021

    work page 2021

  2. [2]

    Procthor: Large- scale embodied ai using procedural generation,

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

    work page 2022

  3. [3]

    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,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 5228–5234

    work page 2024

  4. [4]

    Stubborn: A strong baseline for indoor object navigation. in 2022 ieee,

    H. Luo, A. Yue, Z.-W. Hong, and P. Agrawal, “Stubborn: A strong baseline for indoor object navigation. in 2022 ieee,” in RSJ Inter- national Conference on Intelligent Robots and Systems (IROS) , pp. 3287–3293

    work page 2022

  5. [5]

    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

    work page 2020

  6. [6]

    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,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 18 890–18 900

    work page 2022

  7. [7]

    Peanut: Predicting and navigating to unseen targets,

    A. J. Zhai and S. Wang, “Peanut: Predicting and navigating to unseen targets,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 10 926–10 935

    work page 2023

  8. [8]

    Zson: Zero-shot object-goal navigation using multimodal goal embed- dings,

    A. Majumdar, G. Aggarwal, B. Devnani, J. Hoffman, and D. Batra, “Zson: Zero-shot object-goal navigation using multimodal goal embed- dings,” Advances in Neural Information Processing Systems , vol. 35, pp. 32 340–32 352, 2022

    work page 2022

  9. [9]

    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,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 23 171–23 181

    work page 2023

  10. [10]

    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,” in 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2023, pp. 3554–3560

    work page 2023

  11. [11]

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

    N. Yokoyama, S. Ha, D. Batra, J. Wang, and B. Bucher, “Vlfm: Vision- language frontier maps for zero-shot semantic navigation,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 42–48

    work page 2024

  12. [12]

    Trihelper: Zero-shot object navigation with dynamic assistance,

    L. Zhang, Q. Zhang, H. Wang, E. Xiao, Z. Jiang, H. Chen, and R. Xu, “Trihelper: Zero-shot object navigation with dynamic assistance,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 10 035–10 042

    2024

  13. [13]

    Multi-floor zero-shot object navigation policy,

    L. Zhang, H. Wang, E. Xiao, X. Zhang, Q. Zhang, Z. Jiang, and R. Xu, “Multi-floor zero-shot object navigation policy,” arXiv preprint arXiv:2409.10906, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Apex- nav: An adaptive exploration strategy for zero-shot object navigation with target-centric semantic fusion,

    M. Zhang, Y . Du, C. Wu, J. Zhou, Z. Qi, J. Ma, and B. Zhou, “Apex- nav: An adaptive exploration strategy for zero-shot object navigation with target-centric semantic fusion,” arXiv preprint arXiv:2504.14478, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    Dual memory networks: A versatile adaptation approach for vision-language models,

    Y . Zhang, W. Zhu, H. Tang, Z. Ma, K. Zhou, and L. Zhang, “Dual memory networks: A versatile adaptation approach for vision-language models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2024, pp. 28 718–28 728

    work page 2024

  16. [16]

    Continual vision-and- language navigation,

    M. Jeong, Z. Li, H. Wang, Z. Li, and Y . Lv, “Continual vision-and- language navigation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, arXiv preprint arXiv:2403.15049

    work page arXiv 2024

  17. [17]

    Mem4nav: Boosting vision-and-language navigation with a hierarchical spatial-cognition long-short memory system,

    J. Liu, T. Liu, J. Yu, R. Fu, Y . Pan, and Y . Zhang, “Mem4nav: Boosting vision-and-language navigation with a hierarchical spatial-cognition long-short memory system,” arXiv preprint arXiv:2506.19433 , 2025

    arXiv 2025

  18. [18]

    Exploring spatial representation to enhance llm reasoning in aerial vision-and-language navigation,

    J. Wei, B. Liu, Y . Wu, and Z. Liu, “Exploring spatial representation to enhance llm reasoning in aerial vision-and-language navigation,” arXiv preprint arXiv:2410.08500 , 2024

    Pith/arXiv arXiv 2024

  19. [19]

    Etpnav: Evolving topological planning for vision-language navigation in continuous environments,

    D. An, H. Wang, W. Wang, Z. Wang, Y . Huang, K. He, and L. Wang, “Etpnav: Evolving topological planning for vision-language navigation in continuous environments,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2024

    work page 2024

  20. [20]

    Offline reinforcement learning for visual navigation,

    D. Shah, A. Bhorkar, H. Leen, I. Kostrikov, N. Rhinehart, and S. Levine, “Offline reinforcement learning for visual navigation,” in Conference on Robot Learning (CoRL) , 2022

    work page 2022

  21. [21]

    On the topology awareness and generalization performance of graph neural networks,

    J. Su and C. Wu, “On the topology awareness and generalization performance of graph neural networks,” in European Conference on Computer Vision. Springer, 2024, pp. 73–89

    work page 2024

  22. [22]

    Objectnav 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, “Objectnav revisited: On evaluation of embodied agents navigating to objects,” arXiv preprint arXiv:2006.13171, 2020

    Pith/arXiv arXiv 2006

  23. [23]

    A frontier-based approach for autonomous exploration,

    B. Yamauchi, “A frontier-based approach for autonomous exploration,” in Proceedings 1997 IEEE International Symposium on Computational Intelligence in Robotics and Automation CIRA’97. ’Towards New Com- putational Principles for Robotics and Automation’ . IEEE, 1997, pp. 146–151

    work page 1997

  24. [24]

    Semantic policy network for zero-shot object goal visual navigation,

    Q. Zhao, L. Zhang, B. He, and Z. Liu, “Semantic policy network for zero-shot object goal visual navigation,” IEEE Robotics and Automation Letters, vol. 8, no. 11, pp. 7655–7662, 2023

    2023

  25. [25]

    Imagine before go: Self-supervised generative map for object goal navigation,

    S. Zhang, X. Yu, X. Song, X. Wang, and S. Jiang, “Imagine before go: Self-supervised generative map for object goal navigation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16 414–16 425

    work page 2024

  26. [26]

    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,” in International Conference on Machine Learning. PMLR, 2023, pp. 42 829–42 842

    work page 2023

  27. [27]

    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

  28. [28]

    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,” in Conference on Robot Learning . PMLR, 2025, pp. 2049–2060

    work page 2025

  29. [29]

    Gamap: Zero-shot object goal navigation with multi-scale geometric-affordance guid- ance,

    H. Huang, Y . Hao, C. Wen, A. Tzes, Y . Fanget al., “Gamap: Zero-shot object goal navigation with multi-scale geometric-affordance guid- ance,” Advances in Neural Information Processing Systems , vol. 37, pp. 39 386–39 408, 2024

    work page 2024

  30. [30]

    Unigoal: Towards universal zero-shot goal-oriented navigation,

    H. Yin, X. Xu, L. Zhao, Z. Wang, J. Zhou, and J. Lu, “Unigoal: Towards universal zero-shot goal-oriented navigation,” in Proceedings of the Computer Vision and Pattern Recognition Conference , 2025, pp. 19 057–19 066

    work page 2025

  31. [31]

    Wmnav: Integrating vision-language models into world models for object goal navigation,

    D. Nie, X. Guo, Y . Duan, R. Zhang, and L. Chen, “Wmnav: Integrating vision-language models into world models for object goal navigation,” arXiv preprint arXiv:2503.02247 , 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Habitat 2.0: Training home assistants to rearrange their habitat,

    A. Szot, A. Clegg, E. Undersander, E. Wijmans, Y . Zhao, J. Turner, N. Maestre, M. Mukadam, D. Chaplot, O. Maksymets, A. Gokaslan, V . V ondrus, S. Dharur, F. Meier, W. Galuba, A. Chang, Z. Kira, V . Koltun, J. Malik, M. Savva, and D. Batra, “Habitat 2.0: Training home assistants to rearrange their habitat,” in Advances in Neural Information Processing Sy...

    work page 2021

  33. [33]

    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...

    Pith/arXiv arXiv 2021

  34. [34]

    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,” International Conference on 3D Vision (3DV), 2017

    work page 2017

  35. [35]

    Gpt-4o system card,

    A. Hurst, A. Lerer, A. P. Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Welihinda, A. Hayes, A. Radford et al., “Gpt-4o system card,” arXiv preprint arXiv:2410.21276 , 2024

    Pith/arXiv arXiv 2024

  36. [36]

    General object foundation model for images and videos at scale,

    J. Wu, Y . Jiang, Q. Liu, Z. Yuan, X. Bai, and S. Bai, “General object foundation model for images and videos at scale,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2024, pp. 3783–3795

    work page 2024

  37. [37]

    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

    Pith/arXiv arXiv 2018

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.