Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

DRIVE-Nav: Directional Reasoning, Inspection, and Verification for Efficient Open-Vocabulary Navigation

T0 review · 3 major / 6 minor · reviewed 2026-07-13 · grok-4.5

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

Pith's one-line read Open-vocabulary navigation is more efficient when robots reason over persistent directions, not raw frontiers.

desk verdict Solid systems paper: persistent FMM directions + selective 240° inspection deliver real SPL gains; the efficiency story is the main contribution, not a new scientific paradigm. read the letter →

arxiv 2603.28691 v2 pith:VKUYKQ6H submitted 2026-03-30 cs.RO

Maoguo Gao , Zejun Zhu , Zhiming Sun , Zhengwei Ma , Longze Yuan , Zhongjing Ma , Zhigang Gao , Jinhui Zhang
show 1 more author
Suli Zou
This is my paper
classification cs.RO
keywords Open-VocabularyObjectNavigationembodieddirectionalreasoningfrontierexplorationvision-languagemodelspathefficiencyzero-shotmobilerobotics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

DRIVE-Nav claims that zero-shot open-vocabulary object navigation wastes steps when agents choose among dense frontier points under incomplete views, or when they re-scan and re-open directions already inspected. The fix is to convert frontiers into persistent local exit directions via weighted path analysis, track those directions over time, inspect each with a representative view, and thereafter decide only among still-relevant directions in a forward 240-degree window. Direction-level views also feed a vision-language model that both picks the next branch and enriches open-vocabulary segmentation prompts; a short cross-frame check then rejects false target detections before they enter the map. On standard Habitat benchmarks the method raises success rate and, more sharply, path-efficiency metrics, and the same loop runs on a physical humanoid robot. A sympathetic reader cares because the paper ties semantic reliability and shorter paths to one structural choice: treat directions as the stable units of exploration instead of points.

What carries the argument

Persistent directions: frame-wise bearings obtained by intersecting obstacle-aware, centerline-biased FMM paths with a local circle around the robot, clustered by angular gap, associated over time by frontier-set overlap, and each given a representative RGB view for vision-language selection and grounding.

What would settle it

On long HM3D-OVON episodes, force full 360-degree re-inspection at every junction or disable direction tracking so the same physical branch is re-treated as new; if average successful path length and SPL then match or beat the 240-degree tracked version, the efficiency claim fails.

Watch

Extended reading notes

Core claim

Organizing zero-shot open-vocabulary object navigation around persistent directions—extracted from weighted Fast Marching paths, tracked across map growth, and inspected with representative views—reduces redundant revisits and raises path efficiency, while the same directional evidence supports prompt enrichment and multi-frame verification for more reliable target grounding.

Load-bearing premise

Once a direction has been inspected at a junction, later decisions can safely ignore the rear view and only re-examine still-open directions inside a forward 240-degree window without losing good routes.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. DRIVE-Nav reformulates zero-shot open-vocabulary object navigation as directional reasoning over persistent local exits rather than dense frontier points. From RGB-D and odometry it builds obstacle/traversable maps, extracts frontiers, and converts them into path-induced directional candidates via weighted FMM (obstacle penalty plus Voronoi-centerline bias), then clusters bearings, tracks directions over time, and attaches representative RGB views. At decision points the agent inspects still-relevant directions within a forward 240° range (not full panoramas), uses Qwen3-VL for direction selection and target-aware prompt enrichment into SAM3, and applies a three-frame confirm-or-discard Qwen3-VL verifier before writing targets to the object map. On HM3D-OVON it reports 50.2% SR / 32.6% SPL (vs MSGNav 48.3/27.0); it also reports best SPL on HM3Dv2 (41.3%) and MP3D (22.6%), fewer steps than panoramic and frontier baselines on long successful episodes, component ablations, and five successful real-world trials on a Unitree G1 humanoid.

Significance. If the efficiency and reliability claims hold under stronger isolation of the 240° no-reinspect policy, this is a solid systems contribution to zero-shot OVON: it cleanly reframes frontier redundancy as direction-level entities, couples inspection to both exploration and grounding (prompt enrichment), and demonstrates consistent SPL gains plus physical humanoid transfer without task-specific training. The multi-benchmark tables, stepwise ablations (Tables III–IV), long-episode step study (Table II), and real-robot deployment are concrete strengths that go beyond typical simulation-only VLM navigation papers. The work is incremental relative to topology-aware and VLM-frontier lines (VLFM, ApexNav, SG-Nav, MSGNav, VoroNav) but the persistent-direction + restricted-reinspection + verification loop is a useful engineering synthesis with practical path-efficiency impact.

major comments (3)
  1. [Sec. IV-A; Tables II–III] Sec. IV-A (visual summarization) and the central SPL claim: the load-bearing efficiency assumption is that previously inspected directions remain adequately known, so later decisions may restrict new views to a forward 240° range and only re-open a branch when the farthest frontier along the current clustered direction disappears. Table III’s “w/o 240° Inspection View” ablation (SPL 26.8 vs 27.5) is run under an incomplete pipeline without prompt enrichment/verification, so it does not isolate the policy’s contribution to the full 32.6 SPL. Table II’s panoramic comparator is a full 360° re-scan at every decision point, not a controlled re-inspection of previously tracked rear directions under map growth, occlusion change, or junction revisit. Please add (i) a full-pipeline ablation of 240° vs selective re-inspection of tracked rear directions, and (ii) quantitative rates for how often a
  2. [Abstract; Sec. V; Table I] Abstract vs body inconsistency on benchmarks: the abstract claims evaluation on HM3D-OVON, HM3Dv1, HM3Dv2, and MP3D and “best SPL on HM3Dv1, HM3Dv2, and MP3D,” but Sec. V and Table I report only HM3D-OVON, HM3Dv2, and MP3D (no HM3Dv1 numbers). Either add HM3Dv1 results or correct the abstract/claims so the reported scope matches the experiments.
  3. [Sec. IV-A; Sec. IV-C] Sec. IV-A free parameters and sensitivity: direction clustering uses a 45° angle-gap threshold; representative views use ±15° yaw tolerance; FMM uses λ, robs, β, rvor; verification uses a fixed three-frame window with a confirm-if-any-accept / discard-if-any-reject rule and default-accept fallback. These are central to candidate stability and false-positive suppression, yet no sensitivity or failure analysis is provided. At minimum, report sensitivity of SR/SPL (or step count) to the clustering gap and the 240° range, and clarify how often the three-frame verifier falls back to default-accept vs reject, since the confirm-or-discard rule can be brittle under intermittent VLM noise.
minor comments (6)
  1. [Fig. 1; Sec. IV-A] Fig. 1 caption and body text: “Friontier Pionts” and similar typos; also “V oronoi” / “V oronoi-like” spacing artifacts appear repeatedly (e.g., Fvor definition and related work).
  2. [Table I] Table I: blank cells for methods not evaluated on a benchmark are fine, but please state explicitly whether missing entries mean “not reported by authors” vs “not applicable,” and ensure training-free flags match cited papers.
  3. [Sec. V-A; Sec. V-E] Sec. V-A Implementation: SAM3 and Qwen3-VL-8B (sim) vs Qwen3-VL-4B (real) are named, but prompt templates for direction selection, prompt enrichment, and verification are not provided. Including them (appendix) would aid reproducibility.
  4. [Abstract; Keywords] Keywords and abstract: “Open-V ocabulary” spacing; abstract says “HM3Dv1” while body does not—align terminology with Habitat-Matterport naming used in citations.
  5. [Fig. 3; Fig. 4] Fig. 3 / Fig. 4: useful qualitative sequences, but axis/time labels and which colored boxes correspond to which tracked direction IDs are hard to parse in text form; a short legend tying colors to direction IDs would help.
  6. [Sec. II] Related work: VoroNav and topology-aware methods are cited; a sharper one-paragraph contrast on what “persistent directions with representative views + 240° reconsideration” adds beyond Voronoi/topology waypoints would strengthen positioning.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: external Habitat SR/SPL evaluations and ablations; design choices are not predictions forced by construction.

full rationale

DRIVE-Nav is an empirical systems paper whose central claims are measured success rate and path efficiency (SR/SPL/steps) on external Habitat benchmarks (HM3D-OVON, HM3Dv2, MP3D) against prior methods. The directional pipeline (weighted FMM path-induced exits, angular clustering, temporal tracking, 240° representative-view selection, VLM selection, SAM3 prompt enrichment, cross-frame verification) consists of engineering design choices and modules whose contributions are isolated by ablation (Tables III–IV) and analytical step counts (Table II), not quantities defined to equal a fitted objective. There is no self-definitional loop (metrics are not defined in terms of the method), no fitted parameter renamed as a prediction, no load-bearing uniqueness theorem imported from overlapping authors, and no ansatz smuggled in via self-citation that forces the reported SPL. Ordinary self-reference to standard tools (FMM, SAM3, Qwen3-VL) and prior ObjectNav baselines is independent external support. Concerns about whether the 240° no-reinspect policy is fully stress-tested are validation/assumption risks, not circular reductions of the derivation chain.

Assumptions & free parameters 6 free parameters · 5 assumptions · 3 invented entities

The central claim is empirical systems performance, so the ledger is dominated by design thresholds and foundation-model/domain assumptions rather than fitted physical constants. Free parameters are hand-chosen angular and FMM weights that shape candidate directions and inspection cost. Axioms are standard path-planning math plus the domain bet that VLMs/SAM3 give reliable open-vocabulary judgments from the constructed views. Invented entities are the paper’s directional abstractions—useful engineering constructs, not new physical particles—with independent evidence only via the reported navigation metrics and robot trials.

free parameters (6)
  • direction clustering angle-gap threshold = 45 degrees
    Adjacent bearings split into new directions only when angular gaps exceed 45° (Sec. IV-A); this hand threshold controls candidate count and redundancy.
  • representative-view yaw tolerance = ±15 degrees
    A frame is valid for a direction only if camera yaw is within ±15° of the clustered bearing; controls which image is stored for VLM selection.
  • forward inspection angular range = 240 degrees
    Decisions use a 240° forward view range instead of 360°, a core efficiency design choice not derived from first principles.
  • FMM obstacle penalty parameters λ, robs = unspecified
    F_obs = 1 − λ exp(−d_obs/robs) shapes safer paths; values are design parameters not reported numerically.
  • FMM Voronoi attraction parameters β, rvor = unspecified
    F_vor = 1 + β exp(−d_vor/rvor) biases centerline paths; again hand-chosen and not numerically specified.
  • cross-frame verification window length = 3 frames
    Candidates are judged over three consecutive approach frames with confirm-or-discard rules (Sec. IV-C).
assumptions (5)
  • standard math Eikonal/FMM travel-time fields on traversable free space yield feasible paths whose first local-circle intersections are meaningful direction carriers.
    Sec. IV-A invokes the Eikonal equation and FMM backtracking as established geometric planning tools.
  • domain assumption Qwen3-VL can select among representative directional images and produce useful target descriptions for open-vocabulary navigation under partial indoor observations.
    Central reasoning engine in Sec. IV-B and experiments; performance inherits VLM reliability limits.
  • domain assumption SAM3 open-vocabulary phrase segmentation plus short multi-view VLM checks are adequate for writing targets into the object map.
    Grounding and verification pipeline in Sec. IV-B/C; no task-specific detector training.
  • domain assumption Habitat RGB-D episodes with 500-step budgets and standard SR/SPL success criteria are a valid proxy for the claimed navigation efficiency.
    Experimental setup Sec. V-A; real-world section only partially stress-tests this proxy.
  • ad hoc to paper Previously inspected rear directions need not be re-observed at later junctions except when the current direction’s farthest frontier vanishes.
    Explicit policy choice justifying 240° inspection (Sec. IV-A); load-bearing for the efficiency claim.
invented entities (3)
  • persistent directional entities (tracked local exits with representative views)
    purpose: Replace dense frontier-point candidates with stable direction-level options for VLM selection and inspection.
    Core abstraction of Sec. IV-A; defined operationally via FMM bearings, clustering, and temporal association rather than independent physical measurement.
  • inspection-guided prompt enrichment loop into SAM3
    purpose: Turn VLM scene descriptions from selected directions into refined segmentation prompts for open-vocabulary grounding.
    Sec. IV-B mechanism; evidence is ablation gains, not an external predicted observable outside the navigation task.
  • cross-frame confirm-or-discard verifier with failed-position memory
    purpose: Suppress false-positive target writes during approach without extra navigation steps.
    Sec. IV-C; evaluated only via end-task metrics inside this system.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DRIVE-Nav: Directional Reasoning, Inspection, and Verification for Efficient Open-Vocabulary Navigation." pith.science (2026). https://pith.science/paper/VKUYKQ6H

@misc{pith2026260328691,
  author       = {Pith},
  title        = {Pith review of: DRIVE-Nav: Directional Reasoning, Inspection, and Verification for Efficient Open-Vocabulary Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VKUYKQ6H}},
  note         = {Machine review of arXiv:2603.28691}
}
read the original abstract

Open-Vocabulary Object Navigation (OVON) requires an embodied agent to locate a language-specified target in unknown environments. Many zero-shot methods rely on frontier-candidate reasoning under incomplete observations, while topology-aware methods reduce candidate redundancy but may still introduce panoramic inspection overhead and repeated reconsideration. We present DRIVE-Nav, a structured framework that organizes exploration around persistent directions rather than raw frontiers. By inspecting encountered directions more completely and restricting subsequent decisions to still-relevant directions within a forward 240-degree view range, DRIVE-Nav reduces redundant revisits and improves path efficiency. The framework extracts and tracks directional candidates from weighted Fast Marching Method (FMM) paths, maintains representative views for semantic inspection, and combines vision-language-guided prompt enrichment with cross-frame verification to improve grounding reliability. Experiments on HM3D-OVON, HM3Dv1, HM3Dv2, and MP3D demonstrate strong overall performance and consistent efficiency gains. On HM3D-OVON, DRIVE-Nav achieves 50.2% SR and 32.6% SPL, improving the previous best method by 1.9% SR and 5.6% SPL. It also delivers the best SPL on HM3Dv1, HM3Dv2, and MP3D and transfers to a physical humanoid robot. Real-world deployment also demonstrates its effectiveness.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. SpikingNav: Robust Embodied Navigation with Spiking Neural Policies

    cs.RO 2026-08 conditional novelty 6.0 of 10

    A spiking sensing encoder and spiking policy network improve ObjectNav success under visual corruptions (8.45% to 13.71%) while using fewer parameters and fewer FLOPs than a matched ANN baseline.

  2. SpaceVLN: A Zero-Shot Vision-and-Language Navigation Agent with Online Spatial Cognitive Memory and Reasoning

    cs.RO 2026-06 unverdicted novelty 6.0 of 10

    SpaceVLN proposes a stagewise closed-loop framework using Spatial Cognitive Memory and Spatial-CoT for zero-shot vision-and-language navigation and object-goal navigation, reporting SOTA results on R2R-CE, RxR-CE, GN-...

Reference graph

Works this paper leans on

30 extracted references · 9 linked inside Pith · cited by 2 Pith papers

  1. [1]

    ObjectNav revisited: On evaluation of embodied agents navigating to objects,

    D. Batra, A. Gokaslan, A. Kembhavi, O. Maksymets, R. Mottaghi, M. Savva, A. Toshev, and E. Wijmans, “ObjectNav revisited: On evaluation of embodied agents navigating to objects,”arXiv preprint arXiv:2006.13171, 2020

    arXiv 2006

  2. [2]

    HM3D- OVON: A dataset and benchmark for open-vocabulary object goal navigation,

    N. Yokoyama, R. Ramrakhya, A. Das, D. Batra, and S. Ha, “HM3D- OVON: A dataset and benchmark for open-vocabulary object goal navigation,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 5543–5550

    2024

  3. [3]

    Object goal navigation using goal-oriented semantic exploration,

    D. S. Chaplot, D. P. Gandhi, A. Gupta, and R. R. Salakhutdinov, “Object goal navigation using goal-oriented semantic exploration,” Advances in Neural Information Processing Systems, vol. 33, pp. 4247–4258, 2020

    2020

  4. [4]

    PONI: Potential functions for Object- Goal navigation with interaction-free learning,

    S. K. Ramakrishnanet al., “PONI: Potential functions for Object- Goal navigation with interaction-free learning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

    2022

  5. [5]

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

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

    2022

  6. [6]

    Cows on pasture: Baselines and benchmarks for language-driven zero-shot object navigation,

    S. Y . Gadre, M. Wortsman, G. Ilharco, L. Schmidt, and S. Song, “Cows on pasture: Baselines and benchmarks for language-driven zero-shot object navigation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 23 171–23 181

    2023

  7. [7]

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

    K. Zhou, K. Zheng, C. Pryor, Y . Shen, H. Jin, L. Getoor, and X. E. Wang, “ESC: Exploration with soft commonsense constraints for zero-shot object navigation,” inInternational Conference on Machine Learning. PMLR, 2023, pp. 42 829–42 842

    2023

  8. [8]

    L3MVN: Leveraging large language models for visual target navigation,

    B. Yu, H. Kasaei, and M. Cao, “L3MVN: Leveraging large language models for visual target navigation,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 3554–3560

    2023

Show all 30 references
  1. [9]

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

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

    2024

  2. [10]

    A frontier-based approach for autonomous exploration,

    B. Yamauchi, “A frontier-based approach for autonomous exploration,” inProceedings of the 1997 IEEE International Symposium on Com- putational Intelligence in Robotics and Automation. IEEE, 1997, pp. 146–151

    1997

  3. [11]

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

    P. Wu, Y . Mu, B. Wu, Y . Hou, J. Ma, S. Zhang, and C. Liu, “V oroNav: V oronoi-based zero-shot object navigation with large language model,” arXiv preprint arXiv:2401.02695, 2024

    2024 arXiv

  4. [12]

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

    B. Linet al., “SG-Nav: Online 3D scene graph prompting for LLM- based zero-shot object navigation,” inAdvances in Neural Information Processing Systems, 2024

    2024

  5. [13]

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

    M. Zhang, Y . Du, C. Wu, J. Zhou, Z. Qi, J. Ma, and B. Zhou, “ApexNav: An adaptive exploration strategy for zero-shot object navigation with target-centric semantic fusion,”IEEE Robotics and Automation Letters, 2025

    2025

  6. [14]

    A fast marching level set method for monotonically advancing fronts,

    J. A. Sethian, “A fast marching level set method for monotonically advancing fronts,”Proceedings of the National Academy of Sciences, vol. 93, no. 4, pp. 1591–1595, 1996

    1996

  7. [15]

    SAM 3: Segment anything with concepts,

    N. Carion, L. Gustafson, Y .-T. Hu, S. Debnath, R. Hu, D. Suris, C. Ryali, K. V . Alwala, H. Khedr, A. Huang,et al., “SAM 3: Segment anything with concepts,”arXiv preprint arXiv:2511.16719, 2025

    2025 arXiv

  8. [16]

    Qwen3-VL technical report,

    S. Bai, Y . Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge,et al., “Qwen3-VL technical report,”arXiv preprint arXiv:2511.21631, 2025

    2025 arXiv

  9. [17]

    Habitat-Matterport 3D semantics dataset,

    K. Yadav, R. Ramrakhya, S. K. Ramakrishnan, T. Gervet, J. Turner, A. Gokaslan, N. Maestre, A. X. Chang, D. Batra, M. Savva,et al., “Habitat-Matterport 3D semantics dataset,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 4927–4936

    2023

  10. [18]

    Matterport3D: Learning from RGB- D data in indoor environments,

    A. Chang, A. Dai, T. Funkhouser, M. Halber, M. Niessner, M. Savva, S. Song, A. Zeng, and Y . Zhang, “Matterport3D: Learning from RGB- D data in indoor environments,”arXiv preprint arXiv:1709.06158, 2017

    2017 arXiv

  11. [19]

    Learning transferable visual models from natural language supervi- sion,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervi- sion,” inProceedings of the 38th International Conference on Machine ...

    2021

  12. [20]

    MSGNav: Unleashing the power of multi- modal 3D scene graph for zero-shot embodied navigation,

    X. Huang, S. Zhao, Y . Wang, X. Lu, W. Zhang, R. Qu, W. Li, Y . Wang, and C. Wen, “MSGNav: Unleashing the power of multi- modal 3D scene graph for zero-shot embodied navigation,”arXiv preprint arXiv:2511.10376, 2025

    2025

  13. [21]

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

    H. Yin, X. Xu, L. Zhao, Z. Wang, J. Zhou, and J. Lu, “UniGoal: Towards universal zero-shot goal-oriented navigation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, 2025, pp. 19 057–19 066

    2025

  14. [22]

    An object-driven navigation strategy based on active perception and semantic association,

    Y . Guo, J. Sun, R. Zhang, Z. Jiang, Z. Mi, C. Yao, X. Ban, and M. S. Obaidat, “An object-driven navigation strategy based on active perception and semantic association,”IEEE Robotics and Automation Letters, vol. 9, no. 8, pp. 7110–7117, 2024

    2024

  15. [23]

    OpenFMNav: Towards open-set zero-shot object navigation via vision-language foundation models,

    Y . Kuang, H. Lin, and M. Jiang, “OpenFMNav: Towards open-set zero-shot object navigation via vision-language foundation models,” arXiv preprint arXiv:2402.10670, 2024

    2024 arXiv

  16. [24]

    InstructNav: Zero-shot system for generic instruction navigation in unexplored environment,

    Y . Long, W. Cai, H. Wang, G. Zhan, and H. Dong, “InstructNav: Zero-shot system for generic instruction navigation in unexplored environment,”arXiv preprint arXiv:2406.04882, 2024

    2024 arXiv

  17. [25]

    PIRLNav: Pretraining with imitation and RL finetuning for ObjectNav,

    R. Ramrakhya, D. Batra, E. Wijmans, and A. Das, “PIRLNav: Pretraining with imitation and RL finetuning for ObjectNav,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 17 896–17 906

    2023

  18. [26]

    DD-PPO: Learning near-perfect pointgoal navigators from 2.5 billion frames,

    E. Wijmans, A. Kadian, A. Morcos, S. Lee, I. Essa, D. Parikh, M. Savva, and D. Batra, “DD-PPO: Learning near-perfect pointgoal navigators from 2.5 billion frames,”arXiv preprint arXiv:1911.00357, 2019

    1911 arXiv

  19. [27]

    Uni-NaVid: A video-based vision-language- action model for unifying embodied navigation tasks,

    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

    2024 arXiv

  20. [28]

    Move to understand a 3D scene: Bridging visual grounding and exploration for efficient and versatile embodied navigation,

    Z. Zhu, X. Wang, Y . Li, Z. Zhang, X. Ma, Y . Chen, B. Jia, W. Liang, Q. Yu, Z. Deng,et al., “Move to understand a 3D scene: Bridging visual grounding and exploration for efficient and versatile embodied navigation,” inProceedings of the IEEE/CVF International Conference on Co...

    2025

  21. [29]

    TANGO: Training- free embodied AI agents for open-world tasks,

    F. Ziliotto, T. Campari, L. Serafini, and L. Ballan, “TANGO: Training- free embodied AI agents for open-world tasks,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 24 603–24 613

    2025

  22. [30]

    FAST-LIO2: Fast direct LiDAR-inertial odometry,

    W. Xu, Y . Cai, D. He, J. Lin, and F. Zhang, “FAST-LIO2: Fast direct LiDAR-inertial odometry,”IEEE Transactions on Robotics, vol. 38, no. 4, pp. 2053–2073, 2022

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.