Pith. sign in

REVIEW 4 major objections 5 minor 29 references

TRAVEL: Training-Free Retrieval and Alignment for Vision-and-Language Navigation

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A modular, zero-shot pipeline that extracts landmark sequences with an LLM, retrieves goal panoramas with SigLIP, and ranks BFS path hypotheses with GPT-4o grounding achieves an average nDTW of 88.92% on R2R-Habitat, surpassing the VLMaps…

desk verdict Component-level landmark retrieval is a clean, likely reproducible result; the full-pipeline navigation claim is unverified because Table 2 lacks a VLMaps baseline. read the letter →

arxiv 2502.07306 v2 pith:PFK46X77 submitted 2025-02-11 cs.CV cs.AIcs.CLcs.LGcs.RO

Navid Rajabi , Jana Kosecka This is my paper · ORCID
classification cs.CVcs.AIcs.CLcs.LGcs.RO
keywords vision-and-languagenavigationzero-shotlandmarkretrievalSigLIPGPT-4odynamicprogrammingtopologicalmapR2R-Habitat
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

This paper tries to establish that a modular, zero-shot system can outperform joint semantic-map approaches on complex vision-and-language navigation. It decomposes the task into landmark extraction by an LLM, goal retrieval by SigLIP, path hypothesis generation by BFS on a topological map, and path ranking by VLM-based alignment with dynamic programming. On the R2R-Habitat benchmark the system reports an average nDTW of 88.92% and landmark retrieval Precision@10 of 70.0%, versus 34.4% for VLMaps. If correct, the result suggests that sophisticated end-to-end training is not necessary when modern LLMs and VLMs are combined with an explicit map and a simple alignment procedure.

What carries the argument

The load-bearing object is the Pano2Land alignment matrix: a binary matrix $M$ whose rows are landmark phrases, columns are panoramas along a path hypothesis, and entries are GPT-4o groundings (1 if the landmark is visible in that panorama). The Pano2Land algorithm runs a dynamic program over this matrix, analogous to the longest common subsequence recurrence, computing the maximum number of landmarks grounded in the correct order, normalized by path length. This alignment score ranks the BFS-generated path hypotheses, and the top-scoring path is the agent's answer. The mechanism's key property is that it couples the retrieval bottleneck (top-k goals) to a cheap, interpretable sequence-alignment step rather than to a trained policy.

What would settle it

Choose a path uniformly at random from the top-k BFS hypotheses and measure nDTW against the ground truth; if random selection already yields roughly the same nDTW as the full pipeline, then the GPT-4o alignment and dynamic programming stages are not carrying the improvement that the paper attributes to them.

Watch

Extended reading notes

Core claim

The central claim is that treating vision-and-language navigation as retrieve-then-align rather than learn-to-follow outperforms joint semantic-map baselines without any fine-tuning. Given an instruction, an LLM extracts an ordered list of landmark phrases; SigLIP retrieves the top-k panoramas matching the last landmark; BFS shortest paths from the start to each candidate goal form k path hypotheses; and a dynamic-programming alignment called Pano2Land, which scores a panorama sequence against the landmark phrases using GPT-4o binary grounding scores, picks the best path. The authors report Precision@10 of 70.0% for last-landmark retrieval over 127 landmarks, versus 34.4% for VLMaps, and an average nDTW of 88.92% on five R2R-Habitat environments, with the caveat that the correct path is in the hypothesis set only 65.72% of the time.

Load-bearing premise

The method only works when the correct destination is somewhere in the short list of images the vision model picks out, and when the true route is one of the shortest routes to that destination; the paper's own measurement puts this condition at only 65.72% of episodes.

Editorial extensions

If this is right

  • A retrieval-first pipeline using panoramic landmark retrieval with SigLIP more than doubles goal-retrieval precision over CLIP-based occupancy maps, reaching 70.0% Precision@10 versus 34.4%.
  • Comparing the two ranking schemes shows that per-panorama binary grounding plus dynamic programming (Approach I) yields higher nDTW than whole-sequence rating by GPT-4o (Approach II): 88.92% versus 88.34%.
  • Because the correct path appears in the hypothesis set only 65.72% of the time, the final path-ranking stage inherits this ceiling; the system's upper bound is set by goal retrieval and BFS hypothesis coverage, not by the ranking mechanism.
  • The modular, training-free design means each component can be upgraded or swapped independently, so future improvements to landmark extraction or visual grounding can be slotted in without retraining the rest of the pipeline.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The Pano2Land alignment is a generic sequence-alignment operation; the same binary-grounding-plus-LCS scheme could be applied to other embodied tasks that require grounding ordered instructions in ordered observations, such as video-guided manipulation or procedural task execution.
  • The reported 88.92% nDTW depends on the authors' 87% success threshold; evaluating the same pipeline on the full R2R val split with standard success metrics would clarify whether the retrieval advantage translates into full navigation success, not just path fidelity.
  • A controlled test that varies the distinctiveness of the final landmark (for example, 'bedroom' versus 'door') would isolate when top-k retrieval fails, since the paper's own discussion lists frequent and ambiguous landmarks as a failure mode.
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

4 major / 5 minor

Summary. The paper proposes TRAVEL, a modular, training-free pipeline for vision-and-language navigation on the R2R-Habitat dataset. The method extracts ordered landmark phrases from an instruction with an LLM, retrieves top-k candidate goal panoramas with SigLIP, generates k BFS path hypotheses on a topological map built from training episodes, and ranks these hypotheses either by a GPT-4o-based binary grounding matrix aligned with a dynamic-programming/LCS score (Approach I) or by direct GPT-4o sequence ratings (Approach II). The paper reports a landmark-retrieval Precision@10 of 70.0% for SigLIP versus 34.4% for VLMaps/CLIP, and full-pipeline average nDTW of 88.92% over 105 episodes from 5 environments, with a success threshold of 87% nDTW.

Significance. If fully substantiated, the modular zero-shot design is a useful contribution: it avoids end-to-end training, decomposes error sources across interpretable stages, and the component-level comparison of SigLIP versus CLIP-based semantic maps is clean and compelling. The use of dynamic programming for landmark-to-panorama alignment is a reasonable and transparent way to score path hypotheses. However, the paper's central navigation-level claim of superiority over VLMaps is not supported by the evidence presented, because Table 2 contains no VLMaps navigation baseline and the evaluation is confined to 105 episodes with a post hoc success threshold. The strengths are real but the headline claim needs either additional experimental support or explicit qualification.

major comments (4)
  1. [Abstract; Section 3, Table 2] The abstract and Section 3 claim "superior performance compared to other approaches that use joint semantic maps like VLMaps" for navigation, but Table 2 reports nDTW and accuracy only for TRAVEL approaches; the sole VLMaps comparison is the component-level landmark-retrieval Precision@10 in Table 1. Landmark-retrieval superiority does not imply navigation superiority, since the full pipeline also depends on BFS hypothesis generation, GPT-4o grounding, and DP ranking. The paper should either run VLMaps (or an equivalent semantic-map navigation baseline) on the same 105 episodes and report the navigation metrics, or explicitly restrict the superiority claim to the landmark-retrieval component.
  2. [Section 4; Table 2] The reported HYPOPATHGEN accuracy of 65.72% means that for roughly one third of the 105 episodes the ground-truth path or a highly similar path is not among the generated hypotheses. Section 4 correctly states that "the quality of the path hypotheses eventually determines the upper bound," so the average nDTW of 88.92% is conditional on successful goal retrieval and hypothesis generation and cannot be interpreted as a general navigation success rate. The paper should report metrics separately for episodes where the correct path is and is not in the hypothesis set, and should quantify the fraction of episodes for which the final ranking can succeed at all.
  3. [Section 3, Table 2] The 87% nDTW success threshold is described as "based on our empirical analysis," meaning it was chosen post hoc on the same episodes used to report results. This makes the accuracy numbers difficult to interpret. The authors should justify the threshold in advance from prior work or an independent validation set, or report continuous nDTW distributions and standard VLN metrics such as SR and nDTW at commonly used thresholds.
  4. [Section 3, Table 2; Conclusion] The full-pipeline evaluation is limited to 105 episodes across 5 environments (21 per environment). This is a small sample for a claim of general navigation superiority, and environment-level variability is evident from the per-environment HYPOPATHGEN values ranging from 57.1% to 76.2%. The paper should report per-episode distributions, environment-level confidence intervals, and results on a larger or independently defined split (for example a held-out set of R2R episodes). In addition, no code or data release is mentioned, which further hampers verification of the reported 88.92% nDTW and of the retrieval comparison.
minor comments (5)
  1. [Section 3, Algorithm 1] The text refers to "Algorithm 14" when describing the PANO2LAND procedure, but the algorithm is numbered Algorithm 1; the cross-reference should be corrected.
  2. [References] Reference [2] is an exact duplicate of reference [1]; one should be removed or replaced with the intended citation.
  3. [Section 3, Table 2] The definition of HYPOPATHGEN accuracy should specify what counts as a "highly similar" path and how this similarity is measured, since the reported 65.72% is a key quantity for interpreting the pipeline's upper bound.
  4. [Figure 2; Section 3] The topological map construction is described only briefly; the figure caption "All Unique Waypoints & Trajectories" is vague. Please clarify whether only train-split waypoints are used, how edges are defined from the R2R trajectories, and whether the evaluation episodes overlap with the map-building episodes.
  5. [Section 3, Table 2] The table caption and text state that random tie-breaking was repeated up to 3 times and that mean and standard deviation are reported, but it is unclear whether the reported variance is across the tie-breaking repeats or across episodes; this should be stated explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pipeline composes externally pretrained LLMs/VLMs and is evaluated against ground-truth paths and an external VLMaps benchmark.

full rationale

The paper does not derive a result from an input that is equivalent to the output by construction. The proposed system decomposes VLN into LLM landmark extraction, SigLIP goal retrieval, BFS hypothesis generation, GPT-4o grounding, and DP-based alignment; each module is an independent pretrained component or a standard algorithm. The final nDTW values are computed against ground-truth trajectories from R2R-Habitat, which are not used to fit any parameter that is then reported as a prediction. The comparison with VLMaps is an external benchmark on landmark retrieval, and although the abstract's broader navigation claim lacks a full-pipeline VLMaps baseline, that is an evidence/comparison gap, not circular reasoning. The empirical 87% nDTW success threshold is a reporting choice, not a fitted parameter disguised as a result. The paper's own stated limitation that HYPOPATHGEN accuracy bounds later ranking performance is a genuine dependency, but it is an acknowledged bottleneck rather than a self-referential step. There are no load-bearing self-citations, no uniqueness arguments imported from the authors' prior work, and no ansatz smuggled in via citation. Accordingly, the derivation chain is self-contained with respect to circularity, and the honest score is 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central pipeline rests on the reliability of LLM parsing and VLM grounding, plus the completeness of the topological map. There are no fitted scientific constants or invented entities; the main free choices are the unspecified top-k and the empirically chosen success threshold.

free parameters (2)
  • top-k goal candidates (k) = not stated in paper
    Number of SigLIP-retrieved goal panoramas used to generate path hypotheses; affects HYPOPATHGEN coverage, but the paper never states the value of k.
  • nDTW success threshold = 0.87
    nDTW above 87% is treated as success, chosen 'based on our empirical analysis'; this post hoc threshold directly sets the reported accuracy values.
assumptions (5)
  • domain assumption LLM (Llama-3.1-8B-Instruct) extracts the correct ordered sequence of landmarks from the instruction.
    Section 3: 'we extract the sequence of landmarks from the natural language instruction using a pre-trained LLM'; errors here propagate to all later stages per Section 4.
  • domain assumption The topological map built from training-split waypoints and trajectories contains all nodes and edges needed for evaluation episodes.
    Section 3: graph G is built from unique waypoints and trajectories of the training episodes; Section 4 states the method only works in previously explored environments.
  • domain assumption SigLIP cosine similarity between a panorama and the landmark text reliably identifies the goal location in the top-k retrieved nodes.
    Section 3: goal node retrieval is 'carried out by computing cosine similarity between panorama images and the textual description of the landmark.'
  • domain assumption GPT-4o's binary grounding scores accurately indicate whether each landmark is visible in each panorama.
    Figure 4 shows the binary prompting protocol; the entire alignment matrix A depends on these scores.
  • domain assumption Navigation instructions can be reduced to an ordered sequence of landmarks, ignoring action, spatial, and metric phrases.
    Section 4: the method 'only works in cases where the natural language instruction is landmarks-based and is not heavily based on spatial and temporal phrases, action phrases, and absolute metric distances.'

how reviews work

0 comments
Cite this review

Pith. "Pith review of TRAVEL: Training-Free Retrieval and Alignment for Vision-and-Language Navigation." pith.science (2026). https://pith.science/paper/PFK46X77

@misc{pith2026250207306,
  author       = {Pith},
  title        = {Pith review of: TRAVEL: Training-Free Retrieval and Alignment for Vision-and-Language Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PFK46X77}},
  note         = {Machine review of arXiv:2502.07306}
}
abstract

In this work, we propose a modular approach for the Vision-Language Navigation (VLN) task by decomposing the problem into four sub-modules that use state-of-the-art Large Language Models (LLMs) and Vision-Language Models (VLMs) in a zero-shot setting. Given navigation instruction in natural language, we first prompt LLM to extract the landmarks and the order in which they are visited. Assuming the known model of the environment, we retrieve the top-k locations of the last landmark and generate $k$ path hypotheses from the starting location to the last landmark using the shortest path algorithm on the topological map of the environment. Each path hypothesis is represented by a sequence of panoramas. We then use dynamic programming to compute the alignment score between the sequence of panoramas and the sequence of landmark names, which match scores obtained from VLM. Finally, we compute the nDTW metric between the hypothesis that yields the highest alignment score to evaluate the path fidelity. We demonstrate superior performance compared to other approaches that use joint semantic maps like VLMaps on the complex R2R-Habitat instruction dataset and quantify in detail the effect of visual grounding on navigation performance.

Figures

Figures reproduced from arXiv: 2502.07306 by the authors.

Figure 1
Figure 1. Bird’s Eye View visualization of a sample VLN [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Topological Map Construction natural language instructions into linear temporal logic (LTL). Lang2LTL is advantageous because it disam￾biguates the goal specification and facilitates incorporat￾ing temporal constraints. The limitations of Lang2LTL are the need for a parallel dataset of natural language in￾structions and their corresponding fixed set of LTL for￾mulas for fine-tuning the LLMs for the translation stage… view at source ↗
Figure 3
Figure 3. SigLIP vs. VLMaps Query Result for Last Landmark Indexing [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: GPT-4o Landmark Grounding Score Extraction [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Sequence Alignment for Path Ranking (Pano2Land) [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: GPT-4o Full Prompting for Entire Sequence Scoring [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 22 canonical work pages

  1. [1]

    Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments

    Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko S ¨underhauf, Ian Reid, Stephen Gould, and Anton 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 recog- nition, pages 3674–3683, 2018. 1, 2, 3, 7

    work page 2018

  2. [2]

    Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments

    Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko S ¨underhauf, Ian Reid, Stephen Gould, and Anton 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 recog- nition, pages 3674–3683, 2018. 2

    work page 2018

  3. [3]

    Language models are few-shot learners.Ad- vances in neural information processing systems, 33: 1877–1901, 2020

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared D Kaplan, Prafulla Dhariwal, Arvind Nee- lakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.Ad- vances in neural information processing systems, 33: 1877–1901, 2020. 2

    work page 1901

  4. [4]

    Matterport3d: Learning from rgb-d data in indoor environments.arXiv preprint arXiv:1709.06158, 2017

    Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments.arXiv preprint arXiv:1709.06158, 2017. 1, 7

    arXiv 2017

  5. [5]

    Evaluating large language models trained on code

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021. 2

    arXiv 2021

  6. [6]

    History aware multimodal transformer for vision-and-language navigation.Advances in neural in- formation processing systems, 34:5834–5847, 2021

    Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation.Advances in neural in- formation processing systems, 34:5834–5847, 2021. 2

    work page 2021

  7. [7]

    Clip-nav: Using clip for zero-shot vision-and-language navigation.arXiv preprint arXiv:2211.16649, 2022

    Vishnu Sashank Dorbala, Gunnar Sigurdsson, Robinson Piramuthu, Jesse Thomason, and Gaurav S Sukhatme. Clip-nav: Using clip for zero-shot vision-and-language navigation.arXiv preprint arXiv:2211.16649, 2022. 2

    arXiv 2022

  8. [8]

    Speaker-follower models for vision-and- language navigation.Advances in neural information processing systems, 31, 2018

    Daniel Fried, Ronghang Hu, V olkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Tay- lor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and- language navigation.Advances in neural information processing systems, 31, 2018. 2

    work page 2018

Show all 29 references
  1. [9]

    Cross-modal map learning for vi- sion and language navigation

    Georgios Georgakis, Karl Schmeckpeper, Karan Wan- choo, Soham Dan, Eleni Miltsakaki, Dan Roth, and Kostas Daniilidis. Cross-modal map learning for vi- sion and language navigation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15460...

    2022

  2. [10]

    Vln bert: A recurrent vision- and-language bert for navigation

    Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez- Opazo, and Stephen Gould. Vln bert: A recurrent vision- and-language bert for navigation. InProceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 1643–1653, 2021. 2

    2021

  3. [11]

    Visual language maps for robot navigation

    Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 10608–10615. IEEE, 2023. 1, 2, 3, 4

    2023

  4. [12]

    Language models as zero-shot planners: Ex- tracting actionable knowledge for embodied agents

    Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Ex- tracting actionable knowledge for embodied agents. In International Conference on Machine Learning, pages 9118–9147. PMLR, 2022. 2

    2022

  5. [13]

    Stay on the path: Instruction fidelity in vision-and-language naviga- tion.arXiv preprint arXiv:1905.12255, 2019

    Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language naviga- tion.arXiv preprint arXiv:1905.12255, 2019. 4

    1905 arXiv

  6. [14]

    A new path: Scal- ing vision-and-language navigation with synthetic in- structions and imitation learning

    Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Ja- son Baldridge, and Zarana Parekh. A new path: Scal- ing vision-and-language navigation with synthetic in- structions and imitation learning. InProceedings of the IEEE/CVF Conf...

    2023

  7. [15]

    Beyond the nav-graph: Vision- and-language navigation in continuous environments

    Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision- and-language navigation in continuous environments. In Computer Vision–ECCV 2020: 16th European Confer- ence, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVIII 16, pa...

    2020

  8. [16]

    Room-Across-Room: Multilingual vision-and-language navigation with dense spatiotempo- ral grounding

    Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-Across-Room: Multilingual vision-and-language navigation with dense spatiotempo- ral grounding. InConference on Empirical Methods for Natural Language Processing (EMNLP), 2020. 2, 3

    2020

  9. [17]

    Language-driven semantic segmentation

    Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. InInternational Conference on Learning Representations, 2022. 2

    2022

  10. [18]

    Grounding complex natural language commands for tem- poral tasks in unseen environments

    Jason Xinyu Liu, Ziyi Yang, Ifrah Idrees, Sam Liang, Benjamin Schornstein, Stefanie Tellex, and Ankit Shah. Grounding complex natural language commands for tem- poral tasks in unseen environments. InConference on Robot Learning, 2023. 2

    2023

  11. [19]

    Soat: A scene-and object- aware transformer for vision-and-language navigation

    Abhinav Moudgil, Arjun Majumdar, Harsh Agrawal, Stefan Lee, and Dhruv Batra. Soat: A scene-and object- aware transformer for vision-and-language navigation. Advances in Neural Information Processing Systems, 34: 7357–7367, 2021. 2

    2021

  12. [20]

    Learning transferable visual models from natural lan- guage supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sas- try, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural lan- guage supervision. InInternational conference on ma- chine learnin...

    2021

  13. [21]

    Habitat: A Platform for Embodied AI Research

    Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI Research. InProceedings of the IEEE/CVF International Conferenc...

    2019

  14. [22]

    Ving: Learn- ing open-world navigation with visual goals

    Dhruv Shah, Benjamin Eysenbach, Gregory Kahn, Nicholas Rhinehart, and Sergey Levine. Ving: Learn- ing open-world navigation with visual goals. In2021 IEEE International Conference on Robotics and Automa- tion (ICRA), pages 13215–13222. IEEE, 2021. 3

    2021

  15. [23]

    Lm- nav: Robotic navigation with large pre-trained models of language, vision, and action

    Dhruv Shah, Bła ˙zej Osi ´nski, Sergey Levine, et al. Lm- nav: Robotic navigation with large pre-trained models of language, vision, and action. InConference on robot learning, pages 492–504. PMLR, 2023. 3

    2023

  16. [24]

    Learning to navigate unseen environments: Back translation with en- vironmental dropout.arXiv preprint arXiv:1904.04195,

    Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with en- vironmental dropout.arXiv preprint arXiv:1904.04195,

    1904 arXiv

  17. [25]

    Less is more: Generating grounded naviga- tion instructions from landmarks

    Su Wang, Ceslee Montgomery, Jordi Orbay, Vigh- nesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, and Peter An- derson. Less is more: Generating grounded naviga- tion instructions from landmarks. InProceedings of the IEEE/CVF Confere...

    2022

  18. [26]

    Reinforced cross-modal match- ing and self-supervised imitation learning for vision- language navigation

    Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal match- ing and self-supervised imitation learning for vision- language navigation. InProceedings of the IEEE/CVF conference on comput...

    2019

  19. [27]

    Habitat-matterport 3d semantics dataset.ArXiv, abs/2210.05633, 2022

    Karmesh Yadav, Ram Ramrakhya, Santhosh Kumar Ra- makrishnan, Theo Gervet, John Turner, Aaron Gokaslan, Noah Maestre, Angel Xuan Chang, Dhruv Batra, Mano- lis Savva, Alexander William Clegg, and Devendra Singh Chaplot. Habitat-matterport 3d semantics dataset.ArXiv, abs/2210.056...

    arXiv

  20. [28]

    Sigmoid loss for language image pre- training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre- training. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975–11986,

  21. [29]

    Diagnosing vision-and- language navigation: What really matters.arXiv preprint arXiv:2103.16561, 2021

    Wanrong Zhu, Yuankai Qi, Pradyumna Narayana, Kazoo Sone, Sugato Basu, Xin Eric Wang, Qi Wu, Miguel Eck- stein, and William Yang Wang. Diagnosing vision-and- language navigation: What really matters.arXiv preprint arXiv:2103.16561, 2021. 2

    2021 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.