Pith. sign in

REVIEW 3 major objections 5 minor 42 references

A zero-shot navigator that represents its options as a Steiner tree of paths—rather than isolated waypoints—consistently ranks among the top methods in success rate and path efficiency across Gibson, HM3D, and HSSD.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 17:53 UTC pith:HBUJVJMD

load-bearing objection The tree-of-paths option space is a real idea worth taking seriously; the paper just doesn't run the experiment that would prove its central point. the 3 major comments →

arxiv 2603.18624 v2 pith:HBUJVJMD submitted 2026-03-19 cs.RO cs.AIcs.CV

REST: Receding Horizon Explorative Steiner Tree for Zero-Shot Object-Goal Navigation

Shuqi Xiao , Maani Ghaffari , Chengzhong Xu , Hui Kong This is my paper
classification cs.RO cs.AIcs.CV
keywords zero-shot object navigationnext-best-pathSteiner treelarge language model reasoningopen-vocabulary semantic mappingreceding horizon planninginformation gainhierarchical embodied agent
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.

This paper claims that the candidate subgoals a zero-shot navigation agent chooses among—its option space—should be organized as a tree of full paths, not a flat set of isolated waypoints. The claim matters because destination-only scoring systematically misses information gained along the way, and a flat list hides which candidates share the same route. The paper builds REST, a training-free system that constructs such a tree via a Euclidean Steiner Tree optimizer over a sampling-based roadmap, turns each branch into a natural-language caption, and lets an LLM pick the next subtree by chain-of-thought reasoning. On the Gibson, HM3D, and HSSD benchmarks, REST reports success rates and path efficiencies at or near the top, suggesting that the option space is an underexplored design axis for embodied AI.

Core claim

In its own terms: the option space for zero-shot object-goal navigation should be a tree of paths—path-grounded to make en-route information gain explicit, and tree-structured to let a language-model policy discard or commit to entire regions before comparing individual routes. REST instantiates this by compressing an RT-RRT* roadmap into an obstacle-avoiding Euclidean Steiner minimum tree, textualizing each root-level branch via a virtual-camera sweep over an open-vocabulary 3D map, and selecting the next-best subtree through chain-of-thought scoring with an explicit 'none of the above' fallback. The authors report that this representation yields a favorable efficiency-success trade-off on

What carries the argument

The navigation decision tree: an agent-centric Euclidean Steiner tree computed on top of an RT-RRT* roadmap, whose terminals are information-gain-gated viewpoints and whose auxiliary Steiner points merge shared path segments. The Steiner optimization (approximated by Weiszfeld-median candidate generation plus Kruskal MST, iterated on a warm start) turns thousands of redundant paths into a trunk-to-branch-to-leaf structure that surfaces decision points such as corridor forks and room entrances. Each branch is then narrated into a concise caption by a summarizing LLM, and a separate decision LLM performs chain-of-thought scoring over captions.

Load-bearing premise

The summarizing 2B-parameter model compresses each subtree's per-edge spatial annotations into a short caption without dropping the distances, directions, or object-layout cues that the decision LLM needs to pick the branch leading toward the target.

What would settle it

Run REST with subtree captions that explicitly omit all numeric distances and directions (keeping only object names and room types) and measure success rate; if the drop is no larger than removing the Steiner tree entirely, then the caption funnel is not the load-bearing component—or, conversely, if a variant with unlimited raw annotations still underperforms the compacted version, the compression itself is what enables good decisions.

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

If this is right

  • If correct, the 'next-best-path' paradigm replaces the 'next-best-waypoint' paradigm in training-free ObjectNav: options carry cumulative en-route utility rather than terminal utility alone.
  • The separation of a goal-agnostic semantic map from the language-reasoning layer means goal changes do not require recomputing the value map, potentially enabling fast multi-goal queries.
  • Steiner-compacted trees reduce the combinatorial path space to a handful of branches, making LLM-based reasoning tractable and scalable to larger environments.
  • The same option-space design principle—tree-structured, path-grounded candidates—could transfer to other hierarchical embodied agents beyond ObjectNav.

Where Pith is reading between the lines

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

  • Editorial extension: the branch-and-leaf selection schema mirrors human coarse-to-fine navigation and may make policies more interpretable, since each decision is a linguistic choice among named regions rather than a numeric score over coordinates.
  • The paper's own failure mode on HM3D reflective surfaces suggests a testable extension: fusing the unknown-voxel information gain with a geometric traversability prior, or discounting gains that arise from mirror/glass holes, could close the SR gap to the top method.
  • A stress test of the claim: if a small set of hand-crafted region captions (without any Steiner optimization) matched REST's performance, then the tree structure itself, rather than the specific Steiner compaction, would be the active ingredient—a distinction the current ablation does not fully separate.
  • The receding-horizon, event-triggered LLM invocation means the approach's compute cost is concentrated at topological junctions; a practical implication is that the same architecture could run on edge robots with sparse LLM calls.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes REST, a training-free zero-shot object-goal navigation system whose central claim is that the option space presented to a high-level policy should be a tree of full paths, rather than a set of independently scored waypoints. REST builds an open-vocabulary 3D semantic map from RGB-D streams, maintains an agent-centric RT-RRT* roadmap, extracts informative viewpoints by information-gain gating, compacts the resulting path set into a Steiner tree, textualizes each subtree into a spatial narrative, and selects the next branch by chain-of-thought LLM reasoning in a receding-horizon loop. The method is evaluated on Gibson, HM3D, and HSSD against recent zero-shot baselines, and an ablation on HSSD removes LLM reasoning and the Steiner optimizer separately. The paper reports competitive success rates and the best or second-best path efficiency on the three benchmarks, and attributes this to the tree-of-paths option representation.

Significance. If the central claim holds, the paper identifies a genuinely under-explored design axis: the representation of the subgoal option space, as opposed to belief or policy alone. This would be a transferable contribution to hierarchical embodied agents beyond ObjectNav. The strengths are that the framework is fully training-free, uses explicit and open-vocabulary 3D semantics, and is evaluated on three standard benchmarks; the ablation in Table II supports the direction of the two main design choices, and the best/competitive SPL results are enabled by the path-level planning. However, the paper's headline comparative claim — that destination-only waypoint scoring is systematically inferior — is not directly tested, and the published algorithm for the central Steiner-tree construction is logically broken as written. These issues are fixable but require additional evidence and correction before the central claims are supported.

major comments (3)
  1. [§III-B2, Algorithm 1] The repeat-until loop in Algorithm 1 is logically broken as printed. At line 10, T* is updated only if C(T') < C*, but line 11 unconditionally sets C* to C(T'), and line 13 tests `until C(T') >= C*`. Since C* equals C(T') immediately after line 11, the loop always terminates after the first iteration. Thus the text's claim that the solver "converges to a local minimum" and "refines the tree in the background and serves the best-so-far result on demand" is contradicted by the pseudocode. The pseudocode needs a separate current-cost variable, e.g., compare C(T') against the pre-update C*, and only update C* when an improvement is accepted. This is not merely cosmetic: the iterative improvement argument is part of the mechanism that supposedly distinguishes the Steiner compaction from a one-shot MST over candidate Steiner points.
  2. [§IV-C, Table I and §IV-D, Table II] The central claim that "destination-only waypoint scoring is systematically inferior" to the tree-of-paths option space is not directly supported by the experiments. Table I compares REST to systems that differ in perception backbones, mapping, and decision rules, so the SR/SPL gaps could arise from stronger perception, better captioning, or LLM reasoning rather than from the option-space representation. The internal ablations do not fill this gap: the 'w/o LLM' variant removes semantic reasoning entirely, and the 'w/o Steiner' variant retains path-based options. There is no ablation that replaces the Steiner tree with isolated waypoint scoring while keeping the same semantic map, narrator, and LLM policy. Without such a matched control, the paper should either weaken the 'systematically inferior' claim to a comparative systems-level observation or add a direct point-based scoring varian
  3. [§IV-D, Table II] The ablation 'w/o Steiner tree' is confounded by the narration pipeline. The authors explain that without compaction, the raw RRT* subtree creates longer captions, more ray-casting latency, and repetitive tokens that degrade the summarizer. This shows that the Steiner tree helps the current text summarizer, but it does not isolate whether the decision-tree structure itself improves decision quality, as opposed to merely keeping captions within the summarizer's capacity. I would like to see caption length/token statistics and a variant that controls for caption length or uses a stronger summarizer, so the improvement can be attributed to the option representation rather than to the string length presented to the downstream LLM.
minor comments (5)
  1. [§III-B1] The information-gain threshold tau_IG = 10 is reported in units of unknown voxels, but the view frustum and voxel size (5 cm) are only defined later. It would help to state the sensor geometry and the frustum volume used so the threshold is interpretable. The Poisson-disc radius r = 1.25 m also deserves a sensitivity comment, even a one-line statement that results are stable within a reasonable range.
  2. [§IV-A] The evaluation uses the standard 500-step budget, but the number of episodes per dataset and the standard deviation across episode splits are not reported. Given that several results are within 0.1–1.0 SR of each other (e.g., Gibson 85.1 vs 85.2), error bars or at least episode counts would help assess whether the claimed ranking is significant.
  3. [§III-C1] The description of the virtual panoramic camera sweep is clear in concept, but the exact 'designated informative viewpoints' and how the panoramic field of view is constructed (full 360°, or stitched from several orientations) are not specified. Please include the concrete procedure so the narration step is reproducible.
  4. [§IV-D] The sentence 'w/o LLM reasoning defaults to the nearest informative viewpoint' still follows the path to that viewpoint. To keep the ablation clean, it would help to state explicitly whether the option is represented as the viewpoint only or as the full path, and what geometric criterion (distance? gain?) is used in the default.
  5. [Algorithm 1, line 3] The variable VT is reused: it is an input set of viewpoints and is then overwritten inside the loop as 'root or goal'. This is confusing; use a different name for the terminal set in the MST step.

Circularity Check

0 steps flagged

No significant circularity; the central design claim is empirically tested on held-out benchmarks and by falsifiable ablations, with no fitted-input-as-prediction or self-citation chain.

full rationale

REST is an empirical systems paper rather than a first-principles derivation, so circularity must be judged on whether any reported 'prediction' reduces to its own inputs. The central thesis—that a tree-of-paths option space improves zero-shot ObjectNav—is evaluated on held-out benchmarks (Gibson, HM3Dv1, HSSD) and through ablations in Table II. No parameter is fitted to the reported SR/SPL values; the thresholds (tau_IG=10, r=1.25 m, tau_IoS=0.3, tau_sim=0.8) are fixed engineering choices, not calibrated to test outcomes. The components (Qwen3-VL, YOLO-World, EdgeTAM, RT-RRT*, UFOMap, OAESMT, SemanticFusion-style fusion) are external or standard algorithms, and the reference list contains no self-citations by the authors. The ablations 'w/o LLM reasoning' and 'w/o Steiner tree' could have refuted the component contributions and in fact change performance, so the design claims are falsifiable in principle. The absence of an internal matched control for isolated-waypoint scoring is a genuine experimental-confound issue in the cross-system comparison, but it is not circularity: the benchmark numbers are not definitionally equivalent to the design choices, and nothing in the paper is a fit renamed as a prediction. Therefore no circular step is established.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 2 invented entities

The paper rests on hand-set thresholds (tau_IG, r, tau_IoU, tau_IoS, tau_sim), two domain assumptions about LLM priors and perception reliability, an information-gain model the paper itself shows to be adversarially wrong on mirrors, and the probabilistic completeness of its roadmap. None of these are derived; they are engineering choices, which is typical for systems papers but limits the strength of any claim to generalization.

free parameters (6)
  • Information-gain threshold tau_IG = 10
    Retains only RT-RRT* nodes whose view frustum contains at least 10 unknown visible voxels (Sec. III-B1). Hand-chosen; directly controls which options enter the tree.
  • Poisson-disc radius r = 1.25 m
    Spatial thinning radius for informative viewpoints (Sec. III-B1a). Hand-chosen; sets option density in the roadmap.
  • Entity tracking IoU threshold tau_IoU = 0.5
    3D bounding-box IoU threshold for matching clusters between snapshots (Sec. III-A2b).
  • Entity merge thresholds tau_IoS and tau_sim = 0.3 / 0.8
    Spatial overlap (IoS) and embedding cosine-similarity thresholds for merging fragmented labels into entities (Sec. III-A2b).
  • Steiner snap radius delta
    Algorithm 1 line 6 snaps each Weiszfeld median to the nearest existing RT-RRT* node within delta; delta is unspecified.
  • LLM prompt/temperature and narration template
    Chain-of-thought scoring prompt, captioning instructions, and decoding settings for Qwen3-VL are not specified; Sec. IV-D shows caption quality changes SR by 3.6 points, so outputs are sensitive to these choices.
axioms (6)
  • domain assumption The cascaded SVLM->OVOD->SAM perception pipeline and the vote-fusion/DBSCAN mapping produce entity maps accurate enough for navigation decisions.
    Load-bearing for tree narration quality (Sec. III-A2); if detections are noisy, captions mislead the LLM. No quantitative perception accuracy evaluation is provided.
  • domain assumption LLM commonsense priors on object co-occurrence and spatial layout transfer to the benchmark homes (Gibson, HM3D, HSSD).
    The CoT scoring in Sec. III-C2c relies on priors like 'beds are not usually in living rooms'; benchmark homes are real-world-like scans, so the LLM's web-scale priors plausibly match the eval distribution, but the leakage risk is unmeasured.
  • ad hoc to paper Information gain as count of unknown visible voxels (Eq. 1) is an adequate epistemic utility for ObjectNav.
    Used in viewpoint gating and geometric fallback; the paper itself notes HM3D mirror/glass holes inflate this gain and degrade SR (Sec. IV-C), i.e., the metric is adversarially wrong in some scenes.
  • domain assumption An RT-RRT* roadmap re-rooted to the agent, grown at sensor rate, covers all navigational relevant states (probabilistic completeness).
    The option space is a subset of roadmap nodes; decisions not represented by sampled viewpoints are never considered (Sec. III-A3).
  • domain assumption Steiner-tree compaction preserves the best decision structure: merging paths via Steiner points does not discard the optimal route, only its independent representation.
    Algorithm 1 preserves connectivity of all terminals but sacrifices per-path optimality (Sec. III-B2); if a terminal's best path is detoured severely, the narrated branch no longer matches the executable path.
  • standard math MST over collision-free line-of-sight edges plus Weiszfeld-median candidates is a valid approximation of the obstacle-avoiding Euclidean Steiner tree.
    Standard heuristic approach, but no optimality gap is reported, and the claimed convergence to a local minimum is undermined by the loop-termination bug in Algorithm 1.
invented entities (2)
  • Navigation decision tree no independent evidence
    purpose: Hierarchical option space: a Steiner-compacted tree of safe paths rooted at the agent, whose subtrees are the candidates presented to the LLM policy.
    The central new representation. Evidence for its value is the internal HSSD ablation (w/o Steiner tree: SR 53.1 vs 56.7, SPL 25.3 vs 29.1, Table II); no external falsifiable handle (e.g., a third-party reproduction) is provided.
  • Spatial narrative (tree narration) no independent evidence
    purpose: Converts each branch's geometry and semantics into natural-language captions for LLM branch selection.
    A textualization construct; its fidelity to the real map is unmeasured, and the paper's own ablation shows narrator quality materially changes performance.

pith-pipeline@v1.3.0-alltime-deepseek · 11572 in / 21810 out tokens · 216567 ms · 2026-08-02T17:53:33.294353+00:00 · methodology

0 comments
read the original abstract

Zero-shot object-goal navigation (ZSON) requires navigating unknown environments to find a target object without task-specific training. Prior hierarchical solutions mainly focus on either scene understanding and representations (belief) or high-level decision-making and planning (policy), yet treat the option, i.e., the subgoal candidate that belief proposes and policy selects, as an interface inherited from adjacent modules rather than a design axis in its own right. In practice, options are predominantly single waypoints scored by destination utility: a lone destination hides the value gathered en route, and a flat list obscures the relationships among candidates. Our insight is that the option space should be a tree of paths. Full paths expose en-route information gain that destination-only scoring systematically neglects; a tree of shared segments enables coarse-to-fine LLM reasoning that dismisses or pursues entire branches before examining individual leaves, compressing the combinatorial path space into an efficient hierarchy. We instantiate this insight in REST (Receding Horizon Explorative Steiner Tree), a training-free framework that (1) builds an explicit open-vocabulary 3D map from online RGB-D streams; (2) grows an agent-centric tree of safe and informative paths as the option space via sampling-based planning; and (3) textualizes each branch into a spatial narrative and selects the next-best path through chain-of-thought LLM reasoning. Across the Gibson, HM3D, and HSSD benchmarks, REST consistently ranks among the top methods in success rate and path efficiency.

Figures

Figures reproduced from arXiv: 2603.18624 by Chengzhong Xu, Hui Kong, Maani Ghaffari, Shuqi Xiao.

Figure 1
Figure 1. Figure 1: REST reasons over an agent-centric tree of safe and informative paths rather than evaluating isolated waypoints. Here, REST selects the next-best [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of REST, a training-free ObjectNav framework that replans in a receding-horizon manner. At each decision cycle, the agent updates [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The RT-RRT* subtree connecting current agent (indexed by 0) to [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

42 extracted references · 4 linked inside Pith

  1. [1]

    VLFM: Vision-Language Frontier Maps for Zero- Shot Semantic Navigation,

    N. Yokoyamaet al., “VLFM: Vision-Language Frontier Maps for Zero- Shot Semantic Navigation,” inICRA. IEEE, May 2024, pp. 42–48

    2024

  2. [2]

    ObjectNav Revisited: On Evaluation of Embodied Agents Navigating to Objects,

    D. Batraet al., “ObjectNav Revisited: On Evaluation of Embodied Agents Navigating to Objects,” Aug. 2020. [Online]. Available: http://arxiv.org/abs/2006.13171

    Pith/arXiv arXiv 2020

  3. [3]

    VTNet: Visual transformer network for object goal navigation,

    H. Duet al., “VTNet: Visual transformer network for object goal navigation,” inICLR, 2021

    2021

  4. [4]

    Pirlnav: Pretraining with imitation and rl finetun- ing for objectnav,

    R. Ramrakhyaet al., “Pirlnav: Pretraining with imitation and rl finetun- ing for objectnav,” inCVPR, 2023, pp. 17 896–17 906

    2023

  5. [5]

    Object Goal Navigation with Recursive Implicit Maps,

    S. Chenet al., “Object Goal Navigation with Recursive Implicit Maps,” inIROS. IEEE, Oct. 2023, pp. 7089–7096

    2023

  6. [6]

    ZSON: Zero-shot object-goal navigation using multimodal goal embeddings,

    A. Majumdaret al., “ZSON: Zero-shot object-goal navigation using multimodal goal embeddings,” inNeurIPS), S. Koyejoet al., Eds., vol. 35. Curran Associates, Inc., 2022, pp. 32 340–32 352

    2022

  7. [7]

    Learning Transferable Visual Models From Natural Language Supervision,

    A. Radfordet al., “Learning Transferable Visual Models From Natural Language Supervision,” inICML. PMLR, Jul. 2021, pp. 8748–8763

    2021

  8. [8]

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

    K. Zhouet al., “ESC: Exploration with soft commonsense constraints for zero-shot object navigation,” inICML, ser. ICML ’23. JMLR.org, 2023

    2023

  9. [9]

    CoWs on Pasture: Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation,

    S. Y . Gadreet al., “CoWs on Pasture: Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation,” inCVPR. IEEE, Jun. 2023, pp. 23 171–23 181

    2023

  10. [10]

    SG-Nav: Online 3D scene graph prompting for LLM- based zero-shot object navigation,

    H. Yinet al., “SG-Nav: Online 3D scene graph prompting for LLM- based zero-shot object navigation,” inNeurIPS), ser. NeurIPS ’24. Curran Associates Inc., 2024

    2024

  11. [11]

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

    P. Wuet al., “V oroNav: V oronoi-based zero-shot object navigation with large language model,” inICML, 2024

    2024

  12. [12]

    GAMap: Zero-Shot Object Goal Navigation with Multi- Scale Geometric-Affordance Guidance,

    Y . Fanget al., “GAMap: Zero-Shot Object Goal Navigation with Multi- Scale Geometric-Affordance Guidance,” inNeurIPS), vol. 37. Neural Information Processing Systems Foundation, Inc. (NeurIPS), 2024, pp. 39 386–39 408

    2024

  13. [13]

    L3MVN: Leveraging Large Language Models for Visual Target Navigation,

    B. Yuet al., “L3MVN: Leveraging Large Language Models for Visual Target Navigation,” inIROS. IEEE, Oct. 2023, pp. 3554–3560

    2023

  14. [14]

    Bridging Zero-shot Object Navigation and Foundation Models through Pixel-Guided Navigation Skill,

    W. Caiet al., “Bridging Zero-shot Object Navigation and Foundation Models through Pixel-Guided Navigation Skill,” inICRA. IEEE, May 2024, pp. 5228–5234

    2024

  15. [15]

    PanoNav: Mapless Zero-Shot Object Navigation with Panoramic Scene Parsing and Dynamic Memory,

    Q. Jinet al., “PanoNav: Mapless Zero-Shot Object Navigation with Panoramic Scene Parsing and Dynamic Memory,” Nov. 2025. [Online]. Available: http://arxiv.org/abs/2511.06840

    arXiv 2025

  16. [16]

    Object goal navigation using goal-oriented semantic exploration,

    D. S. Chaplotet al., “Object goal navigation using goal-oriented semantic exploration,”NeurIPS), vol. 33, pp. 4247–4258, 2020

    2020

  17. [17]

    Learning to map for active semantic goal naviga- tion,

    G. Georgakiset al., “Learning to map for active semantic goal naviga- tion,” inICLR, 2022

    2022

  18. [18]

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

    M. Zhanget al., “ApexNA V: An adaptive exploration strategy for zero- shot object navigation with target-centric semantic fusion,”IEEE Robot. Autom. Lett., vol. 10, no. 11, pp. 11 530–11 537, 2025

    2025

  19. [19]

    A frontier-based approach for autonomous exploration,

    B. Yamauchi, “A frontier-based approach for autonomous exploration,” inProc. IEEE Int. Symp. Comput. Intell. Robot. Autom.IEEE Comput. Soc. Press, 1997, pp. 146–151

    1997

  20. [20]

    Sensor-based exploration for convex bodies: A new roadmap for a convex-shaped robot,

    Ji Yeong Leeet al., “Sensor-based exploration for convex bodies: A new roadmap for a convex-shaped robot,”IEEE Trans. Robot., vol. 21, no. 2, pp. 240–247, Apr. 2005

    2005

  21. [21]

    A Tree-Based Next-Best-Trajectory Method for 3-D UA V Exploration,

    B. Lindqvistet al., “A Tree-Based Next-Best-Trajectory Method for 3-D UA V Exploration,”IEEE Trans. Robot., vol. 40, pp. 3496–3513, 2024

    2024

  22. [22]

    A Formal Basis for the Heuristic Determination of Minimum Cost Paths,

    P. Hartet al., “A Formal Basis for the Heuristic Determination of Minimum Cost Paths,”IEEE Trans. Syst. Sci. Cybern., vol. 4, no. 2, pp. 100–107, 1968

    1968

  23. [23]

    An Efficient Sampling-Based Method for Online Informative Path Planning in Unknown Environments,

    L. Schmidet al., “An Efficient Sampling-Based Method for Online Informative Path Planning in Unknown Environments,”IEEE Robot. Autom. Lett., vol. 5, no. 2, pp. 1500–1507, Apr. 2020

    2020

  24. [24]

    UFOMap: An efficient probabilistic 3D mapping framework that embraces the unknown,

    D. Duberget al., “UFOMap: An efficient probabilistic 3D mapping framework that embraces the unknown,”IEEE Robot. Autom. Lett., vol. 5, no. 4, pp. 6411–6418, 2020

    2020

  25. [25]

    Sam 2: Segment anything in images and videos,

    N. Raviet al., “Sam 2: Segment anything in images and videos,”arXiv preprint arXiv:2408.00714, 2024. [Online]. Available: https://arxiv.org/abs/2408.00714

    Pith/arXiv arXiv 2024

  26. [26]

    SemanticFusion: Dense 3D semantic mapping with convolutional neural networks,

    J. McCormacet al., “SemanticFusion: Dense 3D semantic mapping with convolutional neural networks,” inICRA, May 2017, pp. 4628–4635

    2017

  27. [27]

    A density-based algorithm for discovering clusters in large spatial databases with noise,

    M. Esteret al., “A density-based algorithm for discovering clusters in large spatial databases with noise,” inKDD, vol. 96, no. 34, 1996, pp. 226–231

    1996

  28. [28]

    MobileCLIP2: Improving multi-modal reinforced training,

    F. Faghriet al., “MobileCLIP2: Improving multi-modal reinforced training,”TMLR, 2025

    2025

  29. [29]

    RT-RRT*: A real-time path planning algorithm based on RRT*,

    K. Naderiet al., “RT-RRT*: A real-time path planning algorithm based on RRT*,” inProc. ACM SIGGRAPH Conf. Motion in Games. ACM, Nov. 2015, pp. 113–118

    2015

  30. [30]

    Zachariasenet al.,Obstacle-Avoiding Euclidean Steiner Trees in the Plane: An Exact Algorithm, 1999, pp

    M. Zachariasenet al.,Obstacle-Avoiding Euclidean Steiner Trees in the Plane: An Exact Algorithm, 1999, pp. 286–299

    1999

  31. [31]

    Sur le point pour lequel la somme des distances de n points donnés est minimum,

    E. Weiszfeld, “Sur le point pour lequel la somme des distances de n points donnés est minimum,”Tohoku Math. J., First Ser., vol. 43, pp. 355–386, 1937

    1937

  32. [32]

    On the shortest spanning subtree of a graph and the traveling salesman problem,

    J. B. Kruskal, “On the shortest spanning subtree of a graph and the traveling salesman problem,”Proc. Amer. Math. Soc., vol. 7, no. 1, pp. 48–50, 1956

    1956

  33. [33]

    Gibson Env: Real-World Perception for Embodied Agents,

    F. Xiaet al., “Gibson Env: Real-World Perception for Embodied Agents,” inCVPR. IEEE, Jun. 2018, pp. 9068–9079

    2018

  34. [34]

    Habitat-matterport 3D dataset (HM3D): 1000 large-scale 3D environments for embodied AI,

    S. K. Ramakrishnanet al., “Habitat-matterport 3D dataset (HM3D): 1000 large-scale 3D environments for embodied AI,” 2021. [Online]. Available: https://arxiv.org/abs/2109.08238

    Pith/arXiv arXiv 2021

  35. [35]

    Habitat Synthetic Scenes Dataset (HSSD-200): An Analysis of 3D Scene Scale and Realism Tradeoffs for ObjectGoal Navigation,

    M. Khannaet al., “Habitat Synthetic Scenes Dataset (HSSD-200): An Analysis of 3D Scene Scale and Realism Tradeoffs for ObjectGoal Navigation,” inCVPR. IEEE, Jun. 2024, pp. 16 384–16 393

    2024

  36. [36]

    Habitat challenge 2023,

    K. Yadavet al., “Habitat challenge 2023,” https://aihabitat.org/challenge/ 2023/, 2023

    2023

  37. [37]

    Qwen3-VL Technical Report,

    S. Baiet al., “Qwen3-VL Technical Report,” Nov. 2025. [Online]. Available: http://arxiv.org/abs/2511.21631

    Pith/arXiv arXiv 2025

  38. [38]

    YOLO-World: Real-Time Open-V ocabulary Object Detection,

    T. Chenget al., “YOLO-World: Real-Time Open-V ocabulary Object Detection,” inCVPR. IEEE, Jun. 2024, pp. 16 901–16 911

    2024

  39. [39]

    EdgeTAM: On-device track anything model,

    C. Zhouet al., “EdgeTAM: On-device track anything model,” inCVPR, 2025, pp. 13 832–13 842

    2025

  40. [40]

    TriHelper: Zero-Shot Object Navigation with Dynamic Assistance,

    L. Zhanget al., “TriHelper: Zero-Shot Object Navigation with Dynamic Assistance,” inIROS. IEEE, Oct. 2024, pp. 10 035–10 042

    2024

  41. [41]

    ImagineNav: Prompting vision-language models as embodied navigator through scene imagination,

    X. Zhaoet al., “ImagineNav: Prompting vision-language models as embodied navigator through scene imagination,” inICLR, 2025

    2025

  42. [42]

    UniGoal: Towards Universal Zero-shot Goal-oriented Navigation,

    H. Yinet al., “UniGoal: Towards Universal Zero-shot Goal-oriented Navigation,” inCVPR. IEEE, Jun. 2025, pp. 19 057–19 066

    2025