Pith. sign in

REVIEW 3 major objections 7 minor 25 references

SignScene claims that rendering the local scene as a sign-centric abstract top-view map lets a vision-language model ground diverse navigational signs to concrete local paths, achieving 88% grounding accuracy and enabling mapless robot navi

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 sign-centric abstract map lets a vision-language model turn navigational sign instructions into correct paths 88.6% of the time across nine environment types.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Solid sign-grounding system with a genuinely useful AToM representation and a credible 101/114 result, but the open-world claim rests on a self-produced benchmark and a two-category path taxonomy that isn't tested outside it. the 3 major comments →

arxiv 2602.12686 v2 pith:CKP7OMQF submitted 2026-02-13 cs.RO

SignScene: Visual Sign Grounding for Mapless Navigation

Nicky Zimmerman , Joel Loo , Benjamin Koh , Zishuo Wang , David Hsu This is my paper
classification cs.RO
keywords sign groundingmapless navigationvision-language modelsabstract top-view mapspatial reasoningrobot navigationwayfindingtop-down rendering
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

SignScene aims to let robots navigate unfamiliar human-built environments without maps by reading signs, much as humans do. It formalizes sign grounding as the problem of mapping a sign's semantic instruction to a concrete path in the local 3D scene, and argues that the key is the spatial representation, not the model alone. The paper proposes the Abstract Top-view Map (AToM), a simplified 2D diagram centered on the sign and aligned with its viewing direction, which lets a vision-language model reason effectively about spatial relationships. On a benchmark of 114 queries across nine environments, the approach grounds signs correctly 88% of the time and demonstrates real-world navigation on a quadruped robot using only signs.

Core claim

The paper claims that sign grounding can be solved by a vision-language model if the local scene is presented as an Abstract Top-view Map (AToM) centered on the sign and aligned with its viewing direction. AToM stores three kinds of navigation-relevant elements: signs with their parsed content, explicit structures such as stairs and escalators as labeled 3D boxes, and implicit paths as traversable point clouds. Rendered as a simplified 2D top-down diagram with lettered candidate paths and labeled boxes, AToM lets the VLM map an instruction like 'take the escalator' or 'forward-left' to a concrete path. The claim is supported by 101/114 (88.6%) grounding accuracy on a nine-environment benchma

What carries the argument

The Abstract Top-view Map (AToM) is the central object. It converts a metric-semantic 3D map of the local scene into a sign-centric 2D diagram: the map is always centered on the sign, and the upward direction is aligned with the sign's head-on viewing direction. Implicit paths are represented as a polygon with lettered points on convex protrusions (candidate discrete paths), and explicit structures are labeled bounding boxes. This representation carries the argument because it turns abstract instructions into a visual spatial layout that a VLM can reason over; the paper's ablations show that rotating into the sign's frame is critical (accuracy drops from 88% to 65% without it) and that addin

Load-bearing premise

The approach assumes that every path a sign can reference is either a known explicit structure (stairs, escalators, lifts) or an implicit path that appears as a convex protrusion of the traversable ground polygon; signs referencing anything else are outside the representation's vocabulary.

What would settle it

Curate a set of additional navigational signs whose referenced paths are not convex protrusions of the traversable polygon—for example, a sign pointing to a route across an open plaza or through a door not in the fixed structure list—and run SignScene's pipeline on them. If grounding accuracy on these cases drops sharply below the 88% on the original 114 queries, the taxonomy is the load-bearing premise; if accuracy stays high, the representation is more general than the paper claims.

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

If this is right

  • A robot can navigate to a goal in an unseen building, mall, or outdoor area using only signs, RGB images, and odometry—no prebuilt map or expensive sensor suite.
  • The sign-grounding problem separates into sign understanding and sign grounding; each can be improved independently, and both are VLM-queried, so progress in VLMs translates directly into better sign grounding.
  • The sign-centric AToM can store multiple signs and re-center on the relevant one, so a single local map can service different goal queries.
  • The exploration module builds AToM incrementally around a sign, making the approach practical for online deployment despite partial observability.

Where Pith is reading between the lines

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

  • If the taxonomy of paths is incomplete—say, a sign points across an open plaza that produces no convex protrusion—the approach would fail on such cases, so the measured 88% is bounded by that taxonomy.
  • Since the paper uses the same VLM for parsing and grounding, part of the accuracy may be model-specific; re-running with a smaller or weaker VLM would reveal how much of the benefit comes from the representation alone.
  • The sign-centric alignment principle could transfer to other VLM spatial-reasoning tasks, such as interpreting arrows in human-robot interaction or instructions in augmented-reality overlays, where aligning the reference frame to the instruction's origin may be similarly important.
  • The paper's own failure analysis lists object detection, compound-sign parsing, crowded environments, and reasoning failures as remaining bottlenecks, so the headline accuracy should be read as conditional on these components working.
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 / 7 minor

Summary. SignScene proposes a VLM-based pipeline for sign grounding: given an RGB observation of a navigational sign and local odometry, it parses the sign into (location, instruction) pairs, builds a sign-centric Abstract Top-view Map (AToM) of the surrounding scene, and queries a VLM to select the path or structure that matches the parsed instruction. The system is evaluated on a self-collected 114-query benchmark spanning nine environment types, reporting 101/114 (88.6%) grounding accuracy, with a no-rotation ablation at 74/114 and a single baseline (ReasonNav) at 30/114. A parsing ablation improves from 38/114 to 93/114. A Spot robot demonstration shows mapless navigation to a goal using only signs. The paper claims that this constitutes open-world sign grounding for mapless navigation.

Significance. If the result holds, the paper makes a useful contribution: AToM is a compact, VLM-friendly representation that lets a robot interpret diverse real-world signs without a prebuilt map. The paper's strengths include the no-rotation ablation (74 to 101 correct), the parsing improvement from 38 to 93 correct, the breadth of environments, and the real Spot demonstration. The authors also give a candid failure analysis and limitations section. However, the open-world claim is contingent on an untested completeness assumption about the two-class path taxonomy, and the headline number comes from a private benchmark with a single mismatched baseline and no error bars. The contribution is defensible if scoped to the tested taxonomy and evaluated more carefully; as stated, the evidence does not fully support 'open world' generalization.

major comments (3)
  1. [§IV-B, §IV-D] The two-class path model is load-bearing. AToM represents implicit paths only as convex protrusions of the simplified ground polygon and explicit paths only as a fixed-vocabulary set of structures. The paper justifies this taxonomy solely by 'our collected dataset' (Sec. IV-B), and the benchmark is the same 114-query set. No example in Table I or Appendix A tests a sign referencing a straight corridor continuation, an open plaza, floor-marked paths, a bridge, or a tunnel; under the current rendering such a path would have no lettered frontier and could not be selected. The Limitations section (Sec. VI) does not acknowledge this taxonomy risk. Please either narrow the claim to the tested taxonomy or provide external evidence of completeness, e.g., a taxonomy survey of navigational signage or held-out scenes containing non-protruding path types.
  2. [§V-B, Table I; §IV-C] The headline 101/114 (88.6%) is a point estimate on the authors' own benchmark. There are no confidence intervals, no per-environment variance summaries, and no significance test. The only baseline, ReasonNav, is mismatched by design: it assumes four cardinal directions and planar-geometry path detection, and it does not perform exploration; the paper itself lists these as reasons for its low score. The comparison therefore supports 'better than ReasonNav on this benchmark', not the general claim of 'significantly outperforms baselines'. Moreover, the thresholds τ_dist and τ_angle (Sec. IV-C) were empirically tuned using the same VLMs under test, so the reported number is a selected-best estimate. I recommend reporting confidence intervals or per-environment variance, adding a second baseline (e.g., a semantic top-down map without the abstract rendering), and/or releasing the benchmark.
  3. [§IV-C, §V-D, Appendix B] Explicit structures are detected only through a fixed vocabulary drawn from the 20-symbol sign dictionary (Listing 1). The paper states that the approach 'naturally extends to larger or open sets', but this is not an evaluation. The failure analysis (Sec. V-D) reports high false-positive and false-negative rates even for a class like 'escalator'. The evidence therefore supports open-world generalization only for the closed set of structures and symbol meanings present in the authors' dictionary and benchmark. Please either demonstrate generalization to held-out structure types or symbols, or soften the 'open-world' framing in the abstract and conclusion.
minor comments (7)
  1. [Abstract] Typo: 'we formalize this assign grounding' should be 'as sign grounding'.
  2. [II-B] Typos in Related Work: 'ad VLMS' should be 'and VLMs'; 'topoligical' should be 'topological'; 'bsed' should be 'based'.
  3. [References] Reference [21] has a malformed author list ('R MapNav Xu'); please correct the citation.
  4. [§IV-B] The lettered frontier points are introduced in the AToM rendering description but their role in grounding is only explained later in Sec. IV-E; please define the notation earlier for readability.
  5. [§V-A] The text says 36 sequences were recorded but does not explicitly state how the 114 queries are distributed across them. Please clarify the relationship between sequences and queries.
  6. [Appendix B] VLM Prompt 2's output dictionary omits 'u-turn' and 'travelator' categories that appear in the symbol dictionary (Listing 1); align the prompt schema with the label list.
  7. [Table I] Consider adding environment-level totals and success rates so the 'stable across environments' claim is easy to verify from the table alone.

Circularity Check

0 steps flagged

No circular-by-construction derivation; the 88% result is an empirical benchmark result, with self-produced data and data-driven design choices raising external-validity concerns rather than circularity.

full rationale

I walked the derivation chain from the problem formulation (Sec. III) through AToM (Sec. IV-B), construction (Sec. IV-C), reasoning (Sec. IV-E), and evaluation (Sec. V-B). The reported 101/114 grounding accuracy is obtained by running a VLM over an AToM render, converting the selected letter/object to a 3D subgoal, and picking the multiple-choice option closest to that subgoal (Sec. IV-E). This is an empirical measurement, not a quantity defined as its own input. The AToM two-class path taxonomy and the convex-protrusion heuristic (Sec. IV-B, IV-D) are design assumptions motivated by 'our collected dataset' (Sec. V-A); they limit the class of representable paths and thus the scope of the open-world claim, but they do not define the accuracy number. The viewpoint thresholds in Sec. IV-C are tuned 'through tests with VLMs including GPT-5 and Gemini-2.5-Pro', and the same VLM families are used in evaluation; this is a possible source of optimistic evaluation, but the grounding score is not a fitted parameter renamed as a prediction. The sign-understanding module builds on the authors' prior work [1], but it is re-implemented, ablated, and compared against [1] on the same benchmark (Table II), so the self-citation is not an unverified load-bearing black box. No uniqueness theorem, forbidden-alternative argument, or ansatz is imported solely from self-citations. The paper's own failure analysis and Limitations section acknowledge failures in detection, parsing, mapping, and reasoning, which further indicates the result is contingent on external models rather than forced by construction. The main weaknesses—a self-collected benchmark, no external dataset, and the completeness assumption on path types—are external-validity and overfitting concerns, not circular derivation. Under the rule that circularity requires a specific exhibited reduction (same equation on both sides, or a fitted parameter renamed as a prediction), no such step exists.

Axiom & Free-Parameter Ledger

4 free parameters · 7 axioms · 1 invented entities

The central system rests on hand-tuned thresholds, a hand-authored instruction taxonomy, and a completeness assumption about path types; these are inputs rather than derived facts. No new physical entities are introduced; AToM is a software representation without independent evidence.

free parameters (4)
  • tau_dist (sign viewpoint distance threshold) = not reported
    Empirically tuned through VLM tests (Sec. IV-C); controls whether a sign view is parsed and which navigational cues enter AToM.
  • tau_angle (sign viewpoint angular threshold) = not reported
    Same empirical tuning; controls the head-on alignment assumption for sign parsing.
  • Polygon simplification / convex-protrusion parameters = not reported
    AToM rendering and exploration infer discrete paths from convex protrusions of the path polygon (Sec. IV-B/IV-D); no quantitative setting or sensitivity analysis is given.
  • Symbol dictionary and fixed detector vocabulary = 20 labels plus fixed prompt
    Hand-authored output space for sign parsing and object detection; constrains what instructions can be represented (Sec. IV-A, Prompt 2, Fig. 15).
axioms (7)
  • domain assumption Robot has access to RGB observations and odometry
    Problem formulation assumes odometry and RGB input (Sec. III); no global map or localization is needed.
  • domain assumption Signs are viewed head-on or the robot can align with them
    Sign understanding assumes camera aligned with the sign-centric frame (Sec. IV-A); real-world module actively aligns the robot, but if alignment fails parsing degrades.
  • ad hoc to paper All navigation-relevant paths fall into two classes: explicit structures and implicit traversable terrain
    Sec. IV-B states this is based on the authors' collected dataset; signs referencing other path types would not be representable.
  • ad hoc to paper Discrete paths appear as convex protrusions of the top-down path polygon
    Used to propose path frontier points and exploration targets (Sec. IV-B, IV-D); no theoretical or independent empirical justification.
  • domain assumption A VLM can perform spatial grounding on the sign-centric abstract render
    Core hypothesis of the method (Sec. IV-E); validated only on the author-collected benchmark.
  • domain assumption Hand-annotated ground-truth queries and paths are correct
    Evaluation assumes annotations accurately capture the intended sign instructions (Sec. V-A).
  • domain assumption Black-box perception models (GeNIE, Metric3Dv2, GroundingDINO) are sufficiently reliable
    Underlying depth, path, and object models are used without modification (Sec. IV-C); their failures propagate to AToM, as acknowledged in Sec. VI.
invented entities (1)
  • AToM (Abstract Top-view Map) no independent evidence
    purpose: Sign-centric spatial-semantic representation rendered as a simplified 2D diagram for VLM reasoning
    Introduced in Sec. IV-B; its utility is shown only on the authors' own benchmark and a single Spot demo, with no external falsifiable handle.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SignScene: Visual Sign Grounding for Mapless Navigation." pith.science (2026). https://pith.science/paper/CKP7OMQF

@misc{pith2026260212686,
  author       = {Pith},
  title        = {Pith review of: SignScene: Visual Sign Grounding for Mapless Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CKP7OMQF}},
  note         = {Machine review of arXiv:2602.12686}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Navigational signs enable humans to navigate unfamiliar environments without maps. This work studies how robots can similarly exploit signs for mapless navigation in the open world. A central challenge lies in interpreting signs: real-world signs are diverse and complex, and their abstract semantic contents need to be grounded in the local 3D scene. We formalize this as sign grounding, the problem of mapping semantic instructions on signs to corresponding scene elements and navigational actions. Recent Vision-Language Models (VLMs) offer the semantic common-sense and reasoning capabilities required for this task, but are sensitive to how spatial information is represented. We propose SignScene, a sign-centric spatial-semantic representation that captures navigation-relevant scene elements and sign information, and presents them to VLMs in a form conducive to effective reasoning. We evaluate our grounding approach on a dataset of 114 queries collected across nine diverse environment types, achieving 88% grounding accuracy and significantly outperforming baselines. Finally, we demonstrate that it enables real-world mapless navigation on a Spot robot using only signs.

Figures

Figures reproduced from arXiv: 2602.12686 by Benjamin Koh, David Hsu, Joel Loo, Nicky Zimmerman, Zishuo Wang.

Figure 1
Figure 1. Figure 1: Navigational sign grounding aims to enable robot mapless navigation. We highlight the challenges of sign grounding: [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Parsing. We use in-context learning to improve the sign understanding performance. We include in the prompt a sign dictionary, with examples of commonly encountered signs and their labels. The output is a list of pairs, where each pair includes a location and its associated navigational instruction. through text and icons. To improve robustness in parsing symbolic content, we prompt the VLM with a small se… view at source ↗
Figure 3
Figure 3. Figure 3: AToM. The 3D map represents signs and explicit navigational structures as abstract objects, and implicit paths as dense point clouds, and can be rendered into top-down, sign-centric views for VLM reasoning. for estimating 3D geometry, we apply a monocular depth estimation model to RGB observations. We use these inputs to estimate the three key types of scene elements in AToM: Navigational signs. We detect,… view at source ↗
Figure 4
Figure 4. Figure 4: SignScene. In a given a scene, SignScene constructs AToM from RGB observations and robot poses. When deployed on a real robot, it offers modules to align the robot head-on to signs, parse their content, and select signs containing goal-relevant information. The resulting AToM supports queries about directions and actions to take in the local scene. The final cue set C i merged is produced by selecting the … view at source ↗
Figure 5
Figure 5. Figure 5: Failure Analysis. Figures (a) and (b) address reasoning failure. Despite accurate map construction and correct parsing, the VLM reasoning step can still fail. The red dot indicates the location of the sign. In (a) the VLM fails to ground correctly the compound instruction ”left-then-forward”. In (b) the VLM’s choice of boundary point C is influenced by ”exit c” in the parsed location. Figures (c) addresses… view at source ↗
Figure 6
Figure 6. Figure 6: Real-world mapless navigation with signs. The robot is given “TERRACE” as a goal: (a) it explores multiple signs until it finds one with information relevant to the goal, then (b) explores the local environment to build AToM, enabling successful grounding and guiding the robot to take the stairs to reach “TERRACE” in (c). queries Gemini-2.5-Pro (3 seconds), while sign grounding queries GPT-5 (20 seconds). … view at source ↗
Figure 7
Figure 7. Figure 7: Lobby of large hospital. SignScene selects the sign in the AToM (highlighted in orange) as being relevant to the query of main building pharmacy. It grounds the queried location to a navigation subgoal in the AToM render. This subgoal is closest to point C in the query, resulting in the answer “C”. APPENDIX A: VQA EXAMPLES We present several VQA examples to showcase the diversity of our benchmark and to hi… view at source ↗
Figure 8
Figure 8. Figure 8: Near fare gates in subway station. We make multiple queries on a single AToM. For each query, SignScene selects the sign containing information relevant to the query from the AToM. It grounds the queried location to a navigation subgoal in the AToM render, then identifies the point in the image query closest to the computed subgoal, returning that point as the answer to the query. In (a), SignScene selects… view at source ↗
Figure 9
Figure 9. Figure 9: Near atrium of large mall. SignScene selects the sign in the AToM (highlighted in orange) as being relevant to the query of sentosa express. It grounds the queried location to a navigation subgoal in the AToM render. This subgoal is closest to point D in the query, resulting in the answer “D” [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Main road on university campus. SignScene selects the sign in the AToM (highlighted in orange) as being relevant to the query of Temasek Life Sciences Laboratory. It grounds the queried location to a navigation subgoal in the AToM render. This subgoal is closest to point A in the query, resulting in the answer “A” [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Main corridor of large hospital. SignScene selects the sign in the AToM (highlighted in orange) as being relevant to the query of discharge pharmacy. It grounds the queried location to a navigation subgoal in the AToM render. This subgoal is closest to point C in the query, resulting in the answer “C” [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Walking path connecting different campus buildings. SignScene selects the sign in the AToM (highlighted in orange) as being relevant to the query of DSO. It grounds the queried location to a navigation subgoal in the AToM render. This subgoal is closest to point C in the query, resulting in the answer “C” [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Airport terminal. SignScene selects the sign in the AToM (highlighted in orange) as being relevant to the query of lounges. It grounds the queried location to a navigation subgoal in the AToM render. This subgoal is closest to point C in the query, resulting in the answer “C” [PITH_FULL_IMAGE:figures/full_fig_p015_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Shopping mall. SignScene selects the sign in the AToM (highlighted in orange) as being relevant to the query of baby changing. It grounds the queried location to a navigation subgoal in the AToM render. This subgoal is closest to point D in the query, resulting in the answer “D” [PITH_FULL_IMAGE:figures/full_fig_p015_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Symbol Dictionary. ’up-stairs’ : [content], ’down-stairs’ : [content], ’down-escalator’: [content], ’up-escalator’: [content], ’locational’: [content] }. Return only the dict (no comments or formatting). VLM Prompt 3: Sign grounding prompt with AToM Select a letter in a circle or a object name in a bounding box in this image that is potentially closest to {location}, given the following list that consist … 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

25 extracted references · 6 linked inside Pith

  1. [1]

    Sign language: Towards sign understanding for robot autonomy.arXiv preprint arXiv:2506.02556, 2025

    Ayush Agrawal, Joel Loo, Nicky Zimmerman, and David Hsu. Sign language: Towards sign understanding for robot autonomy.arXiv preprint arXiv:2506.02556, 2025

    arXiv 2025

  2. [2]

    RayFronts: Open- Set Semantic Ray Frontiers for Online Scene Under- standing and Exploration, 2025

    Omar Alama, Avigyan Bhattacharya, Haoyang He, Se- ungchan Kim, Yuheng Qiu, Wenshan Wang, Cherie Ho, Nikhil Keetha, and Sebastian Scherer. RayFronts: Open- Set Semantic Ray Frontiers for Online Scene Under- standing and Exploration, 2025. URL https://arxiv.org/ abs/2504.06994

    Pith/arXiv arXiv 2025

  3. [3]

    Human-like Navigation in a World Built for Humans

    Bhargav Chandaka, Gloria Wang, Haozhe Chen, Henry Che, Albert Zhai, and Shenlong Wang. Human-like Navigation in a World Built for Humans. InConference on Robot Learning, 2025

    2025

  4. [4]

    Signage-Aware Exploration in Open World using Venue Maps.IEEE Robotics and Automation Letters, 2025

    Chang Chen, Liang Lu, Lei Yang, Yinqiang Zhang, Yizhou Chen, Ruixing Jia, and Jia Pan. Signage-Aware Exploration in Open World using Venue Maps.IEEE Robotics and Automation Letters, 2025

    2025

  5. [5]

    Intentionnet: Map-lite visual navigation at the kilometre scale.arXiv preprint arXiv:2407.03122, 2024

    Wei Gao, Bo Ai, Joel Loo, David Hsu, et al. Intentionnet: Map-lite visual navigation at the kilometre scale.arXiv preprint arXiv:2407.03122, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Metric3D v2: A Versatile Monocular Geometric Foundation Model for Zero-Shot Metric Depth and Surface Normal Estimation

    Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3D v2: A Versatile Monocular Geometric Foundation Model for Zero-Shot Metric Depth and Surface Normal Estimation. 46:10579– 10596, 2024

    2024

  7. [7]

    RA VEN: Resilient Aerial Navigation via Open-Set Semantic Memory and Behavior Adapta- tion.arXiv preprint arXiv:2509.23563, 2025

    Seungchan Kim, Omar Alama, Dmytro Kurdydyk, John Keller, Nikhil Keetha, Wenshan Wang, Yonatan Bisk, and Sebastian Scherer. RA VEN: Resilient Aerial Navigation via Open-Set Semantic Memory and Behavior Adapta- tion.arXiv preprint arXiv:2509.23563, 2025

    arXiv 2025

  8. [8]

    V . I. Levenshtein. Binary codes capable of correcting deletions, insertions, and reversals.Soviet Physics Dok- lady, 10(8):707–710, 1966. Translated from Doklady Akademii Nauk SSSR, 1965, V ol. 163, No. 4, pp. 845– 848

    1966

  9. [9]

    Topviewrs: Vision-language models as top-view spatial reasoners.arXiv preprint arXiv:2406.02537, 2024

    Chengzu Li, Caiqi Zhang, Han Zhou, Nigel Collier, Anna Korhonen, and Ivan Vuli ´c. Topviewrs: Vision-language models as top-view spatial reasoners.arXiv preprint arXiv:2406.02537, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    No map, no problem: A local sensing approach for navigation in human-made spaces using signs

    Claire Liang, Ross A Knepper, and Florian T Pokorny. No map, no problem: A local sensing approach for navigation in human-made spaces using signs. In2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6148–6155. IEEE, 2020

    2020

  11. [11]

    Perceiving signs for navigation guidance in spaces designed for humans

    Claire Liang, Cheng Perng Phoo, Laasya Renganathan, Yingying Yu, Bharath Hariharan, and Hadas Kress- Gazit. Perceiving signs for navigation guidance in spaces designed for humans. InWorkshop on Closing the Academia to Real-World Gap in Service Robotics at Robotics Science and Systems (RSS), 2020

    2020

  12. [12]

    Grounding dino: Marrying dino with grounded pre-training for open- set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding dino: Marrying dino with grounded pre-training for open- set object detection. InComputer Vision – ECCV 2024: 18th European Conference, Milan, Italy, Septem- ber 29–October 4, 2024, Proceedings, Part XLVII...

    doi:10.1007/978-3-031-72970-6 2024

  13. [13]

    Scene action maps: Behavioural maps for navigation without metric information

    Joel Loo and David Hsu. Scene action maps: Behavioural maps for navigation without metric information. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6354–6360. IEEE, 2024

    2024

  14. [14]

    Open scene graphs for open-world object-goal navigation.Interna- tional Journal of Robotics Research, 2025

    Joel Loo, Zhanxin Wu, and David Hsu. Open scene graphs for open-world object-goal navigation.Interna- tional Journal of Robotics Research, 2025

    2025

  15. [15]

    Macenski, T

    S. Macenski, T. Moore, DV Lu, A. Merzlyakov, and M. Ferguson. From the desks of ROS maintainers: A sur- vey of modern & capable mobile robotics algorithms in the robot operating system 2.Robotics and Autonomous Systems, 2023

    2023

  16. [16]

    Slam toolbox: Slam for the dynamic world.Journal of Open Source Software, 6(61):2783, 2021

    Steve Macenski and Ivona Jambrecic. Slam toolbox: Slam for the dynamic world.Journal of Open Source Software, 6(61):2783, 2021. doi: 10.21105/joss.02783. URL https://doi.org/10.21105/joss.02783

    doi:10.21105/joss.02783 2021

  17. [17]

    Clio: Real-time task- driven open-set 3d scene graphs.IEEE Robotics and Automation Letters, 2024

    Dominic Maggio, Yun Chang, Nathan Hughes, Matthew Trang, Dan Griffith, Carlyn Dougherty, Eric Cristofalo, Lukas Schmid, and Luca Carlone. Clio: Real-time task- driven open-set 3d scene graphs.IEEE Robotics and Automation Letters, 2024

    2024

  18. [18]

    SAM 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Rong- hang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R ¨adle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. InThe Thirteenth Interna...

    2025

  19. [19]

    Ge- nie: A generalizable navigation system for in-the-wild environments.IEEE Robotics and Automation Letters, 10(12):12628–12635, 2025

    Jiaming Wang, Diwen Liu, Jizhuo Chen, Jiaxuan Da, Nuowen Qian, Minh Man Tram, and Harold Soh. Ge- nie: A generalizable navigation system for in-the-wild environments.IEEE Robotics and Automation Letters, 10(12):12628–12635, 2025. doi: 10.1109/LRA.2025. 3623038

    doi:10.1109/lra.2025 2025

  20. [20]

    Vision-language navigation: a survey and taxonomy.Neural Computing and Applications, 36(7): 3291–3316, 2024

    Wansen Wu, Tao Chang, Xinmeng Li, Quanjun Yin, and Yue Hu. Vision-language navigation: a survey and taxonomy.Neural Computing and Applications, 36(7): 3291–3316, 2024

    2024

  21. [21]

    A novel memory representation via annotated semantic maps for vlm-based vision-and-language navigation.arXiv preprint arXiv:2502.13451, 2025

    L Zhang, X Hao, Q Xu, Q Zhang, X Zhang, P Wang, J Zhang, Z Wang, S Zhang, and R MapNav Xu. A novel memory representation via annotated semantic maps for vlm-based vision-and-language navigation.arXiv preprint arXiv:2502.13451, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Topv- nav: Unlocking the top-view spatial reasoning potential of mllm for zero-shot object navigation.arXiv preprint arXiv:2411.16425, 2024

    Linqing Zhong, Chen Gao, Zihan Ding, Yue Liao, Huimin Ma, Shifeng Zhang, Xu Zhou, and Si Liu. Topv- nav: Unlocking the top-view spatial reasoning potential of mllm for zero-shot object navigation.arXiv preprint arXiv:2411.16425, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Open3D: A modern library for 3D data processing.arXiv preprint arXiv:1801.09847, 2018

    Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing.arXiv preprint arXiv:1801.09847, 2018

    Pith/arXiv arXiv 2018

  24. [24]

    Constructing Metric- Semantic Maps Using Floor Plan Priors for Long-Term Indoor Localization

    Nicky Zimmerman, Matteo Sodano, Elias Marks, Jens Behley, and Cyrill Stachniss. Constructing Metric- Semantic Maps Using Floor Plan Priors for Long-Term Indoor Localization. 2023

    2023

  25. [25]

    1" : "forward

    Nicky Zimmerman, Joel Loo, Ayush Agrawal, and David Hsu. Signloc: Robust localization using navigation signs and public maps.IEEE Robotics and Automation Letters, 2026. Fig. 7:Lobby of large hospital.SignSceneselects the sign in theAToM(highlighted in orange) as being relevant to the query ofmain building pharmacy. It grounds the queried location to a nav...

    2026

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