Pith. sign in

REVIEW 4 major objections 6 minor 46 references

NavVerse is a physics-enabled benchmark that connects indoor and outdoor scenes in one continuous robot episode, and its zero-shot results show a clear indoor-to-outdoor transition gap, with the best agent's success dropping from 17.65% to

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

2026-08-01 11:58 UTC pith:GAZX4OAG

load-bearing objection A solid, needed benchmark with a plausible but not yet proven transition-gap claim; send it to review, but demand error bars, artifact release, and a path-length-matched comparison. the 4 major comments →

arxiv 2607.19695 v1 pith:GAZX4OAG submitted 2026-07-22 cs.RO cs.CV

NavVerse: Benchmarking Indoor-to-Outdoor Embodied Navigation in Continuous Robot Simulation

Junzhe Wu , Yue Hu , Zeyu Han , Po-Hsun Chang , Yinan Dong , Behrad Rabiei , Maani Ghaffari This is my paper
classification cs.RO cs.CV
keywords embodied navigationindoor-to-outdoor transitionbenchmarkcontinuous robot simulationplace navigationzero-shot evaluationnavigation safetyphysics-enabled simulation
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.

NavVerse is a physics-enabled simulator benchmark that connects indoor, outdoor, and hybrid indoor-to-outdoor scenes in a single continuous robot episode, with 10,000 episodes spanning object, vision-and-language, and a new place-navigation task. The paper's central claim is that indoor-to-outdoor transition is a distinct and currently unmeasured failure mode: in zero-shot evaluation, the strongest end-to-end agent's success falls from 17.65% to 3.64% on place navigation when the episode goes from pure outdoor to indoor-to-outdoor, with similar drops in object and language navigation. Stage-wise analysis shows that 25–48% of hybrid failures happen before the robot reaches outdoor space, and that coverage efficiency drops after exit for every method. The paper argues this transition gap means adaptation, not navigation alone, is the bottleneck, and that physics-enabled metrics expose safety and kinodynamic failures hidden by discrete-action benchmarks. A sympathetic reader would care because real-world deployment—delivery, campus, emergency response—requires exactly this kind of continuous indoor-to-outdoor behavior.

Core claim

On its own terms, the discovery is that a benchmark can isolate the indoor-to-outdoor transition and that this transition is hard for current agents. NavVerse builds 50 hybrid scenes by embedding a single-floor indoor layout behind a road-facing facade opening and aligning the entrance with the terrain, so the robot traverses door, corridor, and street without teleportation or viewpoint switching. Evaluating an end-to-end vision-language-action model, two reinforcement-learned policies, and a modular planner under the same physics-enabled waypoint interface, the paper finds a clear outdoor-to-hybrid success drop—most severely in PlaceNav, from 17.65% to 3.64%—and a pre-exit failure rate of 2

What carries the argument

The central object is NavVerse itself, and within it the door-to-facade scene-assembly mechanism: a single-floor indoor mesh is inserted behind a road-facing facade, the facade is opened at the entrance, blocking geometry is removed, and entrance height is adjusted to match the local terrain, so indoor and outdoor become one continuous physics scene. That assembly is what creates the pre-exit, boundary, and post-exit stages the benchmark then measures with a stage-wise transition analysis (indoor-only versus reached-outdoor episodes, pre/post-exit coverage efficiency) and a new long-horizon task, PlaceNav, which sets place-level goals such as restaurants or banks. The stage-wise analysis is

Load-bearing premise

The load-bearing premise is that the procedurally assembled hybrid scenes—a single-floor indoor mesh connected through a facade opening to an outdoor city, with goals confined to a 5–50 m neighborhood and roads depressed by 0.2 m—reproduce the essential difficulty of real indoor-to-outdoor navigation, so that failure on NavVerse predicts failure on a real robot.

What would settle it

A concrete check: run the same zero-shot agents on real or high-fidelity scanned building-to-street episodes under the same task definitions. If their hybrid success is close to their outdoor success (no large drop), or if the failure distribution does not show the pre-exit and post-exit pattern, the transition gap is an artifact of NavVerse's scene assembly. A cheaper in-silico ablation: vary the facade-opening geometry, lighting discontinuity, road depression depth, and goal radius; if the 17.65% to 3.64% drop disappears or reverses under any of these variations, the benchmark's headline fin

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

If this is right

  • If the transition gap is real, indoor and outdoor navigation should be trained and evaluated as one continuous capability, not as separate benchmarks.
  • Exit-finding is a first-order bottleneck: since a quarter to half of hybrid failures occur before reaching outdoor space, agents need explicit mechanisms to discover and traverse exits.
  • Reaching outdoor space is insufficient: the universal post-exit coverage-efficiency drop implies agents must re-anchor their exploration to larger scale and different topology after crossing.
  • Kinodynamic and safety constraints should be part of navigation evaluation: oracle trajectories that succeed with a legged robot under physics fail or slow a wheeled robot, so discrete-action benchmarks can overestimate executability.
  • Success and safety are orthogonal axes: the method with the highest success rate is not the one with the best safety metrics, so evaluation should report both.

Where Pith is reading between the lines

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

  • Inference: the benchmark's transition gap would be more convincing if validated against real-robot runs through actual building exits; without that, the abrupt lighting and scale change at the door-to-facade connection could be inflating the drop.
  • Inference: the gap between providing the goal location and not providing it suggests future work could decompose NavVerse into separate search and locomotion scores; if the search component is the main driver, improving target grounding and re-localization after exit should be prioritized over low-level control.
  • Inference: the finding that language-conditioned navigation degrades less than place search across the transition implies that route instructions provide portable structure; a testable extension is to add lightweight exit-oriented cues to PlaceNav instructions and measure whether the 3.64% hybrid success rises.
  • Inference: the fixed 0.2 m road depression and 5–50 m goal radius are design choices; ablating them (for example depth 0 or radius 100 m) would reveal whether the measured adaptation bottleneck is robust to scene-assembly parameters.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. NavVerse is a physics-enabled simulation benchmark built on Isaac Sim for indoor, outdoor, and indoor-to-outdoor embodied navigation. It contributes 100 indoor, 50 outdoor, and 50 hybrid scenes, with 10,000 episodes spanning Object Navigation, Vision-and-Language Navigation, and a newly introduced Place Navigation task. The paper describes scene generation, episode verification, evaluation metrics (success, efficiency, safety), and zero-shot experiments with four baselines: SGImagineNav, PoliFormer, UniNaVid, and LongNav-R1. The headline result is that current agents are far from solving cross-context navigation, and that PlaceNav shows a sharp drop from outdoor (17.65% SR for UniNaVid) to indoor-to-outdoor (3.64% SR), which the authors attribute to an 'adaptation bottleneck.' The paper includes oracle sanity checks (GT Trajectory achieves 100% SR) and a detailed failure-mode analysis.

Significance. If the benchmark is accepted as a valid proxy for indoor-to-outdoor robot navigation, it fills a genuine gap: most existing benchmarks evaluate indoor and outdoor settings separately and abstract away physics. NavVerse's strengths are its physics-backed oracle verification, GT Goal/Trajectory sanity checks, the richness of the episode inventory, and the diagnostic metrics for safety and coverage efficiency. The inclusion of a legged embodiment and the controlled oracle-trajectory experiment (Table 7) are also valuable. However, the main empirical claim—that the observed performance drop is caused by cross-context adaptation rather than by route-difficulty or sampling artifacts—is not yet supported by the evidence presented. The benchmark itself is a substantial community resource, but the headline result needs firmer statistical and experimental grounding.

major comments (4)
  1. [§4.2 / Table 2 vs. Table 6; Appendix C.1] The headline PlaceNav drop for UniNaVid (17.65% outdoor vs 3.64% indoor-to-outdoor) is confounded by path-length difficulty. Episode sampling (C.1) enforces only 5 m ≤ L ≤ 50 m and does not balance path-length distributions across scene types. Table 6 shows UniNaVid PlaceNav SR is 14.63% for both Easy and Medium tertiles but 4.88% for the Hard tertile. If indoor-to-outdoor episodes are over-represented in the Hard tertile, the observed drop may be partly or entirely a route-length effect. Please provide a path-length-matched comparison (e.g., stratify by difficulty tertile, or regress success on scene type with path length as a covariate) before claiming an 'adaptation bottleneck.'
  2. [Table 2 / §4.2] The evaluation set is small: PlaceNav has only 68 outdoor and 55 indoor-to-outdoor episodes. A 17.65% vs 3.64% difference corresponds to about 12 vs 2 successes. With no confidence intervals or significance tests, this 'clear drop' may be within sampling noise; the 95% CI for 2/55 is roughly 1–12%. Report binomial confidence intervals, and ideally per-scene clustered errors. The same concern applies to the small cell counts in Table 6 (e.g., Hard tertile cells) and to the VLN rows.
  3. [Table 4 / §4.3; Eq. (11)] The post-exit CE drop is cited as evidence of a 'systematic efficiency drop' after reaching outdoor space. Since CE (Eq. 11) is the number of unique 1-m cells visited per meter traveled, it is naturally affected by scene density and structure: open outdoor spaces with longer straight-line segments may yield lower CE simply because the agent spends more distance per newly covered cell. No pure-outdoor control or policy-independent baseline is reported for CE. To support the adaptation interpretation, compare pre- and post-exit CE against outdoor-only episodes with matched path length and goal distribution, or report a waypoint-following baseline on the same segments.
  4. [§3.1 / B.6 / C.1] The representativeness of the constructed hybrid scenes is assumed rather than demonstrated. The assembly places a single-floor indoor mesh behind a facade opening, restricts goals to a 5–50 m neighborhood, and applies a uniform 0.2 m road depression. No sensitivity analysis over these parameters is given, and no external validation (e.g., comparison with a real-world indoor-to-outdoor navigation dataset, or with a known strong outdoor navigator) is provided. This leaves open the possibility that the measured 'transition gap' is an artifact of the scene-assembly process, such as the abrupt lighting/scale change at the door-to-facade connection or the bounded goal placement. At minimum, an ablation of the goal-distance bound and a GT-exit-control setting would help isolate the transition effect from goal-distance effects.
minor comments (6)
  1. [Abstract / §4.2] The phrase 'adaptation remains major bottleneck' should be 'adaptation remains a major bottleneck.'
  2. [Table 1] The row for NavVerse contains 'DiscreteContinuousWaypoints202' with missing separators; the action-space column is garbled. Similarly, Table 4's header arrows (Ind-Only↓, Reach-Out↑) are not explained in the caption.
  3. [§3.1] There is a typo: 'GRScenesmeshes' should be 'GRScenes meshes.'
  4. [Table 7] The entries like 'µ= 0.5100.00' are missing whitespace; also clarify the table caption: the columns report SR and velocity under different friction coefficients, not only friction.
  5. [Appendix A.1] '10FPS' should be '10 FPS.' Also, the depth range of [0.1, 1000] meters is stated but no noise model is applied; this should be reiterated in the main text if not already clear.
  6. [Appendix D.1] The eval split contains 900 episodes, of which VLN has 150 per scene type (450 total) but ObjNav/PlaceNav have fewer; the imbalance should be noted in the main text when interpreting the comparison across tasks.

Circularity Check

0 steps flagged

No significant circularity: NavVerse's episode generation and baseline evaluations are independent measurements, not reductions of the paper's conclusions to its inputs.

full rationale

NavVerse is a benchmark-construction paper. Its central quantitative claims—that current methods succeed at low rates and that success drops from outdoor to indoor-to-outdoor episodes—are empirical measurements on episodes whose construction is independent of the policies being evaluated. Episode generation uses NavMesh-based start/goal sampling, physics rollouts, oracle-feasibility verification, and human/VLM filtering (Appendix C.1–C.3, Appendix B.7); no baseline result is used to define episodes, and no fitted parameter is relabeled as a prediction. The failure-mode analyses (Tables 4–5) and oracle-trajectory diagnostics (Table 7, Appendix E) are controlled experiments rather than derivations. Two baselines (SGImagineNav, LongNav-R1) are authored by members of this team, but they are used as evaluated systems, not as load-bearing citations for the benchmark's novelty or validity; the headline transition gap also appears in the external UniNaVid baseline. The paper's own Limitations section acknowledges scope limits (single-floor meshes, limited dynamics, no outdoor-to-indoor). The skeptic's concern that the indoor-to-outdoor drop may be confounded by path length or episode difficulty is a validity/correctness issue, not a circularity issue: the paper reports path-length tertile results (Table 6) but does not use them to define or predict the drop. No quoted equation in the paper reduces any claimed result to its input by construction. Under the stated rules, this is a non-finding: score 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claims rest on benchmark protocol choices (success thresholds, timers, NavMesh settings, episode filtering) and on domain assumptions about the realism of the scenes, the physics simulator, and the interface. No new physical entities are introduced. The listed parameters are hand-chosen design choices, not fitted to produce the measured failures; no evidence suggests they were tuned to the baselines.

free parameters (5)
  • success tolerance = 1.6 m (1.0 + r_robot; r_robot=0.6 m)
    Defines SR/SPL; if enlarged, apparent success rises. Hand-chosen per robot, not fitted to baselines.
  • episode timeout = 300 s
    Determines timeout failure rate; hand-chosen, affects all failure analyses.
  • NavMesh agent radius/height = 0.60 m / 0.60 m
    Defines traversable regions and filters episodes; if the radius mismatches Spot's real footprint, difficulty changes.
  • episode path-length bounds = 5–50 m
    Retained episodes only in this range; constrains task difficulty and removes very short/long tasks.
  • road depression depth = 0.2 m
    Modifies terrain traversability for the wheeled vs legged comparison; hand-chosen.
axioms (5)
  • domain assumption Isaac Sim's rigid-body physics and the trained locomotion policy produce physically faithful kinodynamic outcomes for the Spot robot.
    The entire safety/kinodynamic claim rests on the simulator's physical fidelity; no real-robot validation is provided.
  • domain assumption GRScenes (GRUtopia) and Virtual Community meshes are representative of generic indoor and outdoor navigation environments.
    The 100 indoor + 50 outdoor source meshes stand in for 'indoor' and 'outdoor' reality; no diversity analysis beyond counts.
  • domain assumption NavMesh circle-based traversability (agent radius 0.6 m) matches Spot's actual physical traversability.
    Oracle paths and episode filtering assume a simple disc footprint models a legged robot; terrain features like stairs are simplified.
  • domain assumption Storefront assets generated by Gemini 3 Pro Image are semantically plausible and correctly placed per OSM.
    PlaceNav goal validity relies on the storefront-to-facade placement being semantically and geometrically correct; only manual inspection, no objective metric.
  • domain assumption Providing the ground-truth base pose to agents does not invalidate the navigation conclusions.
    All baselines receive perfect localization (A.1), which is unrealistically favorable and may mask failures that would occur in real deployment; relative comparisons may still hold.

pith-pipeline@v1.3.0-alltime-deepseek · 18717 in / 19149 out tokens · 197522 ms · 2026-08-01T11:58:48.589345+00:00 · methodology

0 comments
read the original abstract

Robots deployed in delivery, campus, and emergency-response settings often need to navigate from buildings to streets within a single continuous episode. Existing benchmarks usually evaluate indoor and outdoor navigation separately, and many abstract away robot execution, leaving exit finding, boundary traversal, adaptation, and kinodynamic failures underexplored. We introduce NavVerse, a physics-enabled benchmark for indoor-to-outdoor embodied navigation. NavVerse contains 100 indoor scenes, 50 urban outdoor scenes, and 50 indoor-to-outdoor scenes, and 10,000 episodes spanning Object Navigation, Vision-and-Language Navigation, and Place Navigation tasks, where agents search for semantic points of interest such as restaurants or banks. Agents are evaluated through executable robot interfaces using task-success, path-efficiency, and safety metrics. Zero-shot experiments with RL, VLA, and modular baselines show that current agents remain far from solving cross-context navigation: end-to-end VLAs obtain the highest zero-shot success, while the modular method provides the strongest safety profile. PlaceNav further reveals a clear drop from outdoor to indoor-to-outdoor scenes, indicating that adaptation remains major bottleneck.

Figures

Figures reproduced from arXiv: 2607.19695 by Behrad Rabiei, Junzhe Wu, Maani Ghaffari, Po-Hsun Chang, Yinan Dong, Yue Hu, Zeyu Han.

Figure 1
Figure 1. Figure 1: NAVVERSE connects indoor, outdoor, and hybrid indoor-to-outdoor scenes for ObjNav, VLN, and PlaceNav tasks under continuous physics-enabled robot execution, and evaluates task success, efficiency, and safety. Abstract: Robots deployed in delivery, campus, and emergency-response settings often need to navigate from buildings to streets within a single continuous episode. Existing benchmarks usually evaluate… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the NAVVERSE benchmark construction pipeline. Isaac Sim provides the execution engine, while NAVVERSE contributes connected indoor, outdoor, and indoor-to-outdoor scene construction, task generation, oracle validation, and diagnostic evaluation. We defer implementation details for asset generation, mesh processing, object placement, storefront construction, terrain modeling, and indoor-to-outdo… view at source ↗
Figure 3
Figure 3. Figure 3: Generated BEV maps for representative indoor, outdoor, and indoor-to-outdoor scenes. [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: ObjNav target category distribution [PITH_FULL_IMAGE:figures/full_fig_p020_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: PlaceNav storefront category distribution. [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Treemaps of the most frequent tokens in VLN instructions. Top: action verbs. Bottom: [PITH_FULL_IMAGE:figures/full_fig_p021_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: GT Trajectory qualitative examples. Each row shows one episode: the left panel is a BEV overview with the oracle reference path, and the right panels are egocentric RGB frames. The start is marked by a yellow dot and the goal by a red cross in the BEV, and the red overlay in RGB frames indicates the commanded oracle reference path. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

46 extracted references · 11 linked inside Pith

  1. [1]

    H. Yin, X. Xu, L. Zhao, Z. Wang, J. Zhou, and J. Lu. Unigoal: Towards universal zero-shot goal-oriented navigation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 19057– 19066, 2025

    2025

  2. [2]

    Zhang, K

    J. Zhang, K. Wang, S. Wang, M. Li, H. Liu, S. Wei, Z. Wang, Z. Zhang, and H. Wang. Uni-navid: A video-based vision-language-action model for unifying embodied navigation tasks.arXiv preprint arXiv:2412.06224, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    J. Li, A. Padmakumar, G. Sukhatme, and M. Bansal. Vln-video: Utilizing driving videos for outdoor vision-and-language navigation. InProceedings of the AAAI Conference on Artificial Intelligence, vol- ume 38, pages 18517–18526, 2024

    2024

  4. [4]

    Deitke, E

    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, 35:5982–5994, 2022

    2022

  5. [5]

    M. Wei, C. Wan, X. Yu, T. Wang, Y . Yang, X. Mao, C. Zhu, W. Cai, H. Wang, Y . Chen, et al. Streamvln: Streaming vision-and-language navigation via slowfast context modeling.arXiv preprint arXiv:2507.05240, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Z. Wang, S. Lee, and G. H. Lee. Dynam3d: Dynamic layered 3d tokens empower vlm for vision-and- language navigation.arXiv preprint arXiv:2505.11383, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    H. Tian, J. Meng, W.-S. Zheng, Y .-M. Li, J. Yan, and Y . Zhang. Loc4plan: Locating before planning for outdoor vision and language navigation. InProceedings of the 32nd ACM International Conference on Multimedia, pages 4073–4081, 2024

    2024

  8. [8]

    K.-H. Zeng, Z. Zhang, K. Ehsani, R. Hendrix, J. Salvador, A. Herrasti, R. Girshick, A. Kembhavi, and L. Weihs. Poliformer: Scaling on-policy rl with transformers results in masterful navigators. InConfer- ence on Robot Learning, pages 408–432. PMLR, 2025

    2025

  9. [9]

    Y . Long, W. Cai, H. Wang, G. Zhan, and H. Dong. Instructnav: Zero-shot system for generic instruction navigation in unexplored environment. InConference on Robot Learning, pages 2049–2060. PMLR, 2025

    2049

  10. [10]

    Yokoyama, S

    N. Yokoyama, S. Ha, D. Batra, J. Wang, and B. Bucher. Vlfm: Vision-language frontier maps for zero- shot semantic navigation. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 42–48. IEEE, 2024

    2024

  11. [11]

    Y . Hu, J. Wu, R. Xu, H. Liu, A. Xi, H. X. Liu, R. Vasudevan, and M. Ghaffari. Imaginative world modeling with scene graphs for embodied agent navigation.arXiv preprint arXiv:2508.06990, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    Huang, Y

    H. Huang, Y . Hao, C. Wen, A. Tzes, Y . Fang, et al. Gamap: Zero-shot object goal navigation with multi-scale geometric-affordance guidance.Advances in Neural Information Processing Systems, 37: 39386–39408, 2024

    2024

  13. [13]

    Y . Sun, Y . Qiu, Y . Aoki, and H. Kataoka. Outdoor vision-and-language navigation needs object-level alignment.Sensors, 23(13):6028, 2023

    2023

  14. [14]

    Chang, A

    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

    2017

  15. [15]

    S. K. Ramakrishnan, A. Gokaslan, E. Wijmans, O. Maksymets, A. Clegg, J. M. Turner, E. Undersander, W. Galuba, A. Westbury, A. X. Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. InThirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021

    2021

  16. [16]

    Deitke, W

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

    2020

  17. [17]

    Geiger, P

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun. Vision meets robotics: The kitti dataset.The international journal of robotics research, 32(11):1231–1237, 2013

    2013

  18. [18]

    Caesar, V

    H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom. nuscenes: A multimodal dataset for autonomous driving. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631, 2020. 23

    2020

  19. [19]

    H. Chen, A. Suhr, D. Misra, N. Snavely, and Y . Artzi. Touchdown: Natural language navigation and spatial reasoning in visual street environments. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12538–12547, 2019

    2019

  20. [20]

    De Vries, K

    H. De Vries, K. Shuster, D. Batra, D. Parikh, J. Weston, and D. Kiela. Talk the walk: Navigating new york city through grounded dialogue.arXiv preprint arXiv:1807.03367, 2018

    Pith/arXiv arXiv 2018

  21. [21]

    Anderson, Q

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

    2018

  22. [22]

    A. Ku, P. Anderson, R. Patel, E. Ie, and J. Baldridge. Room-across-room: Multilingual vision-and- language navigation with dense spatiotemporal grounding. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4392–4412, 2020

    2020

  23. [23]

    Kolve, R

    E. Kolve, R. Mottaghi, W. Han, E. VanderBilt, L. Weihs, A. Herrasti, M. Deitke, K. Ehsani, D. Gordon, Y . Zhu, et al. Ai2-thor: An interactive 3d environment for visual ai.arXiv preprint arXiv:1712.05474, 2017

    Pith/arXiv arXiv 2017

  24. [24]

    Savva, A

    M. Savva, A. Kadian, O. Maksymets, Y . Zhao, E. Wijmans, B. Jain, J. Straub, J. Liu, V . Koltun, J. Malik, et al. Habitat: A platform for embodied ai research. InProceedings of the IEEE/CVF international conference on computer vision, pages 9339–9347, 2019

    2019

  25. [25]

    Yadav, S

    K. Yadav, S. K. Ramakrishnan, J. Turner, A. Gokaslan, O. Maksymets, R. Jain, R. Ramrakhya, A. X. Chang, A. Clegg, M. Savva, E. Undersander, D. S. Chaplot, and D. Batra. Habitat challenge 2022. https://aihabitat.org/challenge/2022/, 2022

    2022

  26. [26]

    Krantz, E

    J. Krantz, E. Wijmans, A. Majumdar, D. Batra, and S. Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. InEuropean Conference on Computer Vision, pages 104–120. Springer, 2020

    2020

  27. [27]

    L. Wang, X. Xia, H. Zhao, H. Wang, T. Wang, Y . Chen, C. Liu, Q. Chen, and J. Pang. Rethinking the embodied gap in vision-and-language navigation: A holistic study of physical and visual disparities. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9455–9465, 2025

    2025

  28. [28]

    Isaac Sim, 2025

    NVIDIA. Isaac Sim, 2025. URLhttps://github.com/isaac-sim/IsaacSim

    2025

  29. [29]

    A. Szot, A. Clegg, E. Undersander, E. Wijmans, Y . Zhao, J. Turner, N. Maestre, M. Mukadam, D. S. Chaplot, O. Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat.Advances in neural information processing systems, 34:251–266, 2021

    2021

  30. [30]

    X. Puig, E. Undersander, A. Szot, M. D. Cote, T.-Y . Yang, R. Partsey, R. Desai, A. W. Clegg, M. Hlavac, S. Y . Min, et al. Habitat 3.0: A co-habitat for humans, avatars and robots.arXiv preprint arXiv:2310.13724, 2023

    Pith/arXiv arXiv 2023

  31. [31]

    Todorov, T

    E. Todorov, T. Erez, and Y . Tassa. Mujoco: A physics engine for model-based control. In2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026–5033, 2012. doi:10.1109/IROS. 2012.6386109

    arXiv 2012

  32. [32]

    Koenig and A

    N. Koenig and A. Howard. Design and use paradigms for gazebo, an open-source multi-robot simula- tor. In2004 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (IEEE Cat. No.04CH37566), volume 3, pages 2149–2154 vol.3, 2004. doi:10.1109/IROS.2004.1389727

    Pith/arXiv arXiv 2004

  33. [33]

    F. Xia, A. R. Zamir, Z. He, A. Sax, J. Malik, and S. Savarese. Gibson env: Real-world perception for embodied agents. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 9068–9079, 2018

    2018

  34. [34]

    Thomason, M

    J. Thomason, M. Murray, M. Cakmak, and L. Zettlemoyer. Vision-and-dialog navigation. InConference on Robot Learning, pages 394–406. PMLR, 2020

    2020

  35. [35]

    Y . Qiao, H. Hong, W. Lyu, D. An, S. Zhang, Y . Xie, X. Wang, and Q. Wu. Navbench: Probing multimodal large language models for embodied navigation.arXiv preprint arXiv:2506.01031, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    Shridhar, J

    M. Shridhar, J. Thomason, D. Gordon, Y . Bisk, W. Han, R. Mottaghi, L. Zettlemoyer, and D. Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740–10749, 2020. 24

    2020

  37. [37]

    F. Zhu, X. Liang, Y . Zhu, Q. Yu, X. Chang, and X. Liang. Soon: Scenario oriented object navigation with graph-based exploration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12689–12699, 2021

    2021

  38. [38]

    Y . Qi, Q. Wu, P. Anderson, X. Wang, W. Y . Wang, C. Shen, and A. v. d. Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982–9991, 2020

    2020

  39. [39]

    Z. Wang, Y . Zhu, G. H. Lee, and Y . Fan. Navrag: Generating user demand instructions for embodied navigation through retrieval-augmented llm.arXiv preprint arXiv:2502.11142, 2025

    Pith/arXiv arXiv 2025

  40. [40]

    Schumann and S

    R. Schumann and S. Riezler. Generating landmark navigation instructions from maps as a graph-to-text problem. InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 489–502, 2021

    2021

  41. [41]

    Liang, D

    J. Liang, D. Das, D. Song, M. N. H. Shuvo, M. Durrani, K. Taranath, I. Penskiy, D. Manocha, and X. Xiao. Gnd: Global navigation dataset with multi-modal perception and multi-category traversability in outdoor campus environments. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 2383–2390. IEEE, 2025

    2025

  42. [42]

    Q. Zhou, H. Zhang, X. Lin, Z. Zhang, Y . Chen, W. Liu, Z. Zhang, S. Chen, L. Fang, Q. Lyu, et al. Virtual community: An open world for humans, robots, and society.arXiv preprint arXiv:2508.14893, 2025

    arXiv 2025

  43. [43]

    H. Wang, J. Chen, W. Huang, Q. Ben, T. Wang, B. Mi, T. Huang, S. Zhao, Y . Chen, S. Yang, et al. Grutopia: Dream general robots in a city at scale.arXiv preprint arXiv:2407.10943, 2024

    Pith/arXiv arXiv 2024

  44. [44]

    Deitke, D

    M. Deitke, D. Schwenk, J. Salvador, L. Weihs, O. Michel, E. VanderBilt, L. Schmidt, K. Ehsani, A. Kem- bhavi, and A. Farhadi. Objaverse: A universe of annotated 3d objects. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13142–13153, 2023

    2023

  45. [45]

    Y . Hu, A. Xi, Q. Xiao, S. Isaacson, H. X. Liu, R. Vasudevan, and M. Ghaffari. LongNav-R1: Horizon- Adaptive Multi-Turn RL for Long-Horizon VLA Navigation, Feb. 2026. URLhttp://arxiv.org/ abs/2602.12351. arXiv:2602.12351 [cs.RO]

    arXiv 2026

  46. [46]

    https://docs.cloud.google.com/vertex-ai/generative- ai/docs/models/gemini/3-pro-image

    Gemini 3 Pro Image|Generative AI on Vertex AI. https://docs.cloud.google.com/vertex-ai/generative- ai/docs/models/gemini/3-pro-image. 25