Pith. sign in

REVIEW 3 major objections 6 minor 51 references

This paper claims that jointly optimizing reward-driven exploration and imitation learning on corrected rollouts — rather than running them in sequence — sets a new state of the art in vision-and-language navigation, reaching 69.9% success

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 05:07 UTC pith:QPHCFXIT

load-bearing objection A strong R2R SOTA with a clear staged recipe, but the headline gain rests partly on an unaablated trajectory-selection heuristic that borrows a token-level idea without proof it transfers. the 3 major comments →

arxiv 2607.13461 v1 pith:QPHCFXIT submitted 2026-07-15 cs.RO

Joint On-and-Off Policy Learning for Vision-and-Language Navigation

Qingrong He , Lin Zhao , Kevin Zheng , Liang Lin This is my paper
classification cs.RO
keywords vision-and-language navigationimitation learningreinforcement learningdataset aggregationgroup-relative policy optimizationtrajectory selectionVLN-CER2R
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 tries to show that the two dominant training paradigms for vision-and-language navigation — imitation learning from expert or corrected demonstrations, and reinforcement learning driven by reward — should be run together, not sequentially, and that doing so materially improves how well an embodied agent generalizes to environments it has never seen. The authors build a three-stage pipeline: first, multi-task imitation on expert data; second, imitation on off-policy rollouts whose mistakes have been corrected by an oracle; third, a joint objective that combines a group-relative policy-optimization term on high-entropy rollouts with an imitation term on error-priority-sorted demonstrations. If the paper is right, this joint scheme pushes the success rate on the R2R Val-Unseen split to 69.9%, a new state of the art for agents using only single-view RGB input, and the same recipe yields 68.0% SR on the longer-horizon RxR split. A sympathetic reader would care because the core question — whether exploration can be added on top of error recovery without losing either benefit — is the bottleneck in making instruction-following agents robust outside their training distribution.

Core claim

On its own terms, JOP-VLN's central claim is that the on-policy reinforcement objective and the off-policy imitation objective reinforce each other when optimized jointly with a decaying weight schedule: the imitation term keeps early RL updates stable, and the RL term keeps the agent from overfitting to off-policy expert data. Concretely, the paper reports that adding this joint stage to an already strong imitation-trained model raises SR from 64.4% to 69.9% on R2R Val-Unseen, with each component — high-entropy trajectory sampling, error-correction-prioritized sorting, and the joint weighting — contributing to the final number. The authors additionally claim that a single-RGB agent trained

What carries the argument

The central mechanism is a combined training objective that dynamically weights a group-relative policy-optimization loss and an imitation-learning loss, with the weight µ decaying from 0.5 to 0.05 over 200 training steps. The RL term is computed on only the top-20%-highest-entropy rollouts collected during dataset aggregation, while the IL term is cross-entropy on oracle-corrected demonstrations, reordered so that trajectories with many oracle corrections are seen first. The two schedules work together so the model starts with balanced supervision and gradually shifts toward exploration.

Load-bearing premise

The final-stage gain rests on the untested transfer of a token-level 'top 20% high-entropy' sampling rule to whole navigation trajectories; if that filter instead discards useful rollouts, the reported RL and state-of-the-art improvements could be a sampling artifact rather than evidence for the joint-learning mechanism.

What would settle it

Run Stage 3 with group-relative policy optimization on all collected trajectories and on a randomly chosen 20% subset, comparing R2R Val-Unseen SR against the top-20%-entropy version; if either control matches or beats the filtered version, the high-entropy sampling claim is not supported.

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

If this is right

  • If the reported results are correct, JOP-VLN sets the new state of the art on R2R Val-Unseen among single-RGB observation methods, 4.8 SR points above the previous best.
  • The ablation trajectory implies that DAgger-style imitation and reward-driven exploration are complementary: the largest single jump comes from Stage-2 imitation (49.9 to 64.4 SR), and the joint Stage-3 adds a further gain to reach 69.9 SR.
  • The method's performance on RxR (68.0 SR, 59.3 SPL) suggests the joint training scheme scales to longer, more ambiguous instructions, not just the shorter R2R routes.
  • The real-world 10-shot adaptation results, if they hold up, indicate the trained policy is a practical starting point for physical deployment rather than only a simulator artifact.

Where Pith is reading between the lines

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

  • The paper transfers a token-level sampling rule (train on only the top 20% high-entropy tokens) to whole trajectories without an ablation that compares group-relative policy optimization on all trajectories against the filtered 20%; that comparison is the natural next experiment to pin down where the RL gain really comes from.
  • If high-entropy trajectories correlate with difficult or ambiguous instructions, the entropy filter could be acting as a curriculum rather than as a general efficiency booster; separating the two explanations would require matching on instruction difficulty.
  • The same dynamically weighted IL+RL recipe could be tried on continuous-control navigation or other embodied tasks where off-policy rollouts are expensive; the paper's discrete action-chunk setting leaves that transfer untested.

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 / 6 minor

Summary. The paper proposes JOP-VLN, a three-stage training pipeline for VLM-based vision-and-language navigation in continuous environments. Stage 1 combines action-prediction imitation learning with trajectory summarization; Stage 2 fine-tunes on DAgger-collected trajectories with oracle re-labeling; Stage 3 applies CHORD-style joint optimization of GRPO and IL on DAgger trajectories, adding a high-entropy trajectory filter and an error-correction-prioritized sorting strategy. On the R2R Val-Unseen split the authors report 69.9% SR with single RGB observations, surpassing CorrectNav by 4.8 SR; on RxR they report 68.0% SR, below CorrectNav's 69.3% SR. The central claims are the new R2R state of the art and the benefit of jointly combining off-policy DAgger imitation with on-policy GRPO.

Significance. If the reported numbers are robust, this is a meaningful systems-level advance: it gives a concrete recipe for integrating off-policy DAgger imitation and on-policy GRPO in VLN-CE, with a staged ablation showing a large Stage-2 gain (49.9% to 64.4% SR) and additional gains from Stage-3 joint learning. The paper is positioned against strong recent baselines, uses standard benchmarks, and includes qualitative real-world validation. The staged ablation in Table II is a strength, isolating trajectory summarization, DAgger IL, GRPO, CHORD, and error-correction sorting. The main risk is that the high-entropy trajectory sampling component — one of the two named technical contributions — is not directly ablated and relies on an extrapolation from token-level LLM-RL results, so the causal attribution of the SOTA result to the proposed mechanism is not yet established. No circularity concern is present: the CHORD framework is cited as an external method with no author overlap.

major comments (3)
  1. [§IV.C.1 / Table II] The top-20% high-entropy trajectory filter is not ablated. Rows 5–7 of Table II all apply GRPO/CHORD to the high-entropy subset; there is no row using all DAgger trajectories, nor a random 20% subset. This is load-bearing because the paper's motivating citations [33,34] establish token-level filtering in LLM reasoning, whereas Eq. (5) computes normalized advantages per response. Low entropy does not by itself imply zero advantage unless all G rewards coincide. Please add ablations: CHORD on all trajectories, CHORD on a random 20% subset, and CHORD on the high-entropy subset; also report how token entropies are aggregated to a trajectory-level score and whether the subset is selected once before Stage 3 even though the policy changes during GRPO. Without these, the 69.9% SR cannot be causally attributed to the high-entropy mechanism.
  2. [§V-C, Tables I–II] All benchmark numbers appear to come from a single run. The SOTA claim rests on a 4.8 SR margin over CorrectNav, and the final ablation increment (Table II rows 6 vs 7) is 1.2 SR. In VLN-CE these gaps can be within seed noise. Please report means and standard deviations over at least three seeds for the final model and the key ablations, or provide another statistical comparison. If multi-seed evaluation is infeasible, state the compute budget and explicitly flag the single-run caveat.
  3. [§IV.C.1, Eq. (5)] The 'zero-gradient' motivation for entropy filtering is imprecise. Eq. (5) yields zero advantage only when all G responses have equal reward, not merely when the model is confident/low-entropy. A confidently correct trajectory can still produce nonzero advantages if the reward group is mixed. If the real mechanism is increased reward variance, the paper should report reward statistics or gradient norms for filtered versus unfiltered subsets. Otherwise the filtering rationale rests on an unverified extrapolation from token-level LLM-RL findings.
minor comments (6)
  1. [§II-A, §III-A, §IV-C] Typos: 'incresing', 'histroy', and 'Fiture 4' should be fixed.
  2. [§IV.C.1] Clarify the definition of trajectory-level entropy. The text records 'entropy of each model-predicted action token' but then filters 'top 20% of samples': specify whether the trajectory score is the mean, sum, or maximum token entropy, and what the 20% threshold is relative to.
  3. [Eq. (5)] Add a small epsilon to the standard deviation in the advantage normalization to avoid division by zero when all G responses receive the same reward.
  4. [§V-C, Table I] The sentence claiming JOP-VLN 'even outperforms methods employing panoramic views' is only true on R2R SR; on RxR, OmniNav reports higher SR. Please qualify the claim by benchmark and metric.
  5. [References] Reference entries are duplicated: [14] and [16] are the same paper, and [15] and [18] are the same paper. The author list of [31] also contains a stray formatting artifact.
  6. [Figure 3] The caption uses the same arrow symbol for model-predicted and oracle-corrected actions; distinguish the two with different symbols or explicit labels.

Circularity Check

0 steps flagged

No significant circularity: the pipeline is an empirical recipe and all reported gains are measured on held-out splits; the CHORD loss is explicitly attributed and the high-entropy filter is an unaablated heuristic, not a fitted prediction.

full rationale

JOP-VLN does not claim a mathematical derivation that reduces to its inputs. The central result is a training recipe (IL -> DAgger -> CHORD-style joint optimization) evaluated on R2R/RxR Val-Unseen, which is held out, so the 69.9% SR number is not produced by fitting a parameter to that split. The joint loss in Eq. (8) is the CHORD loss, explicitly cited to [13], and the paper's own contribution is the two VLN-specific data-selection heuristics, not a claimed first-principles derivation. The high-entropy trajectory filter (Section IV.C.1) is a borrowing of a token-level empirical result [33,34] to the trajectory level; it is unaablated (Table II rows 5-7 never compare against 'CHORD without the entropy filter' or a random 20% subset), and the transfer is logically unsupported, but this is a missing-ablation/correctness concern, not circularity: the top-20% selection is not defined in terms of the R2R success rate it is supposed to explain. The error-correction-prioritized sorting strategy (Section IV.C.3) sorts trajectories by proportion of corrected tokens, so the statement that it exposes the model to past mistakes is tautological, but the ablation claim that this improves SR is empirical and not forced by the sorting definition. No uniqueness theorem is imported from the authors, and the only apparent author-overlap citation ([1], which shares co-author L. Zhao) appears in related-work context and is not load-bearing. The strongest critique is that the high-entropy filter is never ablated, which weakens the causal attribution, but that is a validation gap, not a circular derivation.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The method is empirical and relies on standard RL/IL assumptions plus several hand-picked hyperparameters that are not independently validated. The most distinctive premises are the transfer of high-entropy sampling from token-level to trajectory-level and the use of oracle-corrected token proportions as a learning-priority signal. No new physical or conceptual entities are introduced.

free parameters (5)
  • high_entropy_fraction = 0.2 (top 20%)
    GRPO training uses only the top 20% of trajectories by recorded action-token entropy; adopted from token-level LLM-RL studies [33,34] and not ablated in this paper (Sec. IV.C.1).
  • mu_schedule = 0.5 -> 0.05 cosine decay over first 200 steps
    Trade-off weight between GRPO and IL in the CHORD loss; chosen by hand and not varied/ablated (Sec. IV.C.3).
  • dagger_oracle_intervention_threshold = not specified
    The heuristic DAgger strategy triggers oracle intervention when deviation from the ground-truth path exceeds an unspecified threshold; this directly shapes the collected trajectories and is not quantified (Sec. IV.B).
  • reward_combination_weights = not specified
    Four reward functions (format, exact match, position, distance) are defined, but the paper does not state how they are combined into the scalar reward used in GRPO advantages (Sec. IV.C.1).
  • grpo_hyperparameters = G=8, beta=0.1, epsilon=0.2
    Standard GRPO constants, but hand-set and not ablated; central to the RL update in Stage 3 (Sec. V.B).
axioms (4)
  • domain assumption DAgger oracle labels, obtained by replaying visited states, are correct and sufficient as off-policy supervision.
    Stage 2 and Stage 3 assume oracle action chunks provide reliable corrective labels; if oracle replay is noisy, both IL and reward computation degrade (Sec. IV.B).
  • domain assumption Qwen3-VL-8B-Instruct provides adequate visual-language priors for navigation.
    The entire framework is built on this backbone; no sensitivity analysis to the backbone is provided (Sec. III.B).
  • ad hoc to paper Token-level high-entropy sampling transfers to trajectory-level sampling in VLN.
    Sec. IV.C.1 applies the top-20% high-entropy token finding from LLM reasoning [33,34] to navigation trajectories without validating the transfer.
  • ad hoc to paper Trajectory 'learning value' is well proxied by the proportion of oracle-corrected action tokens.
    Sec. IV.C.2 defines sorting priority this way; no evidence is given that this proxy maximizes error-recovery learning.

pith-pipeline@v1.3.0-alltime-deepseek · 11902 in / 17022 out tokens · 159688 ms · 2026-08-02T05:07:11.447700+00:00 · methodology

0 comments
read the original abstract

Vision-and-Language Navigation (VLN) necessitates an embodied agent to navigate in the physical world by adhering to natural language instructions. Recent advancements in Vision-Language Models (VLM) have propelled the development of VLM-based VLN methods with two predominant paradigms: (1) imitation learning (IL) on expert demonstrations, followed by the Dataset Aggregation (DAgger) algorithm to bolster error recovery capabilities; (2) reinforcement learning (RL) driven by verifiable rewards to enhance reasoning and exploration. A notable gap is the absence of integration between these two distinct paradigms. This paper introduces JOP-VLN, a novel VLN framework that synergistically combines off-policy imitation learning and on-policy exploration within a three-stage training pipeline. Initially, IL is employed on expert demonstrations to acquire basic navigation skills. Subsequently, the DAgger algorithm is utilized to generate heuristic exploration trajectories, which are then used for imitation learning to improve error recovery capabilities. Finally, a joint on-and-off policy learning framework is implemented, featuring high-entropy trajectory sampling to enhance RL training efficiency and an error-correction-prioritized trajectory sorting strategy for effective error correction. Extensive experiments demonstrate the efficacy of JOP-VLN, achieving success rates of 69.9% and 68.0% on the VLN-CE R2R and RxR benchmarks, respectively, setting a new state-of-the-art on R2R. Project page: https://qingrongh.github.io/JOP-VLN.

Figures

Figures reproduced from arXiv: 2607.13461 by Kevin Zheng, Liang Lin, Lin Zhao, Qingrong He.

Figure 1
Figure 1. Figure 1: Multi-turn Dialogue Paradigm. Our model is built upon a large vision-language model, i.e., Qwen3-VL [31]. As shown in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Stage 1. Multi-task training. Stage 2. IL on DAgger collected trajectories [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Stage 3. Joint on-and-off policy learning with high-entropy and [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: µ schedule with cosine decay. Error-correction-prioritised trajectory sorting. The dy￾namic scheduling strategy for µ prioritizes imitation learning during the early stage of training. To improve the model’s error recovery capabilities, we devise an error-correction￾prioritised trajectory sorting strategy. Specifically, for a multi-turn dialogue trajectory sample as described in Sec￾tion III-B, we identify… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results of JOP-VLN in representative real-world [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative results of JOP-VLN on the VLN-CE benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p007_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

51 extracted references · 16 linked inside Pith

  1. [1]

    Towards learning a generalist model for embodied navigation,

    D. Zheng, S. Huang, L. Zhao, Y . Zhong, and L. Wang, “Towards learning a generalist model for embodied navigation,” inCVPR, 2024

    2024

  2. [2]

    Navila: Legged robot vision-language- action model for navigation,

    A.-C. Cheng, Y . Ji, Z. Yang, Z. Gongye, X. Zou, J. Kautz, E. Bıyık, H. Yin, S. Liu, and X. Wang, “Navila: Legged robot vision-language- action model for navigation,”Robotics: Science and Systems, 2025

    2025

  3. [3]

    Navid: Video-based vlm plans the next step for vision-and-language navigation,

    J. Zhang, K. Wang, R. Xu, G. Zhou, Y . Hong, X. Fang, Q. Wu, Z. Zhang, and H. Wang, “Navid: Video-based vlm plans the next step for vision-and-language navigation,”Robotics: Science and Systems, 2024

    2024

  4. [4]

    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,”Robotics: Science and Systems, 2025

    2025

  5. [5]

    Streamvln: Streaming vision-and- language navigation via slowfast context modeling,

    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]

    A reduction of imitation learning and structured prediction to no-regret online learning,

    S. Ross, G. Gordon, and D. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” inProceedings of the fourteenth international conference on artificial intelligence and statistics, pp. 627–635, JMLR Workshop and Conference Proceedings, 2011

    2011

  7. [7]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models,

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, M. Zhang, Y . Li, Y . Wu, and D. Guo, “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,”arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Dapo: An open-source llm reinforcement learning system at scale,

    Q. Yu, Z. Zhang, R. Zhu, Y . Yuan, X. Zuo, Y . Yue, W. Dai, T. Fan, G. Liu, L. Liu,et al., “Dapo: An open-source llm reinforcement learning system at scale,”arXiv preprint arXiv:2503.14476, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Group sequence policy optimization,

    C. Zheng, S. Liu, M. Li, X.-H. Chen, B. Yu, C. Gao, K. Dang, Y . Liu, R. Men, A. Yang, J. Zhou, and J. Lin, “Group sequence policy optimization,”arXiv preprint arXiv:2507.18071, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Soft adaptive policy optimization,

    C. Gao, C. Zheng, X.-H. Chen, K. Dang, S. Liu, B. Yu, A. Yang, S. Bai, J. Zhou, and J. Lin, “Soft adaptive policy optimization,”arXiv preprint arXiv:2511.20347, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    Vln-r1: Vision- language navigation via reinforcement fine-tuning,

    Z. Qi, Z. Zhang, Y . Yu, J. Wang, and H. Zhao, “Vln-r1: Vision- language navigation via reinforcement fine-tuning,”arXiv preprint arXiv:2506.17221, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    Octonav: Towards generalist embodied navigation,

    C. Gao, L. Jin, X. Peng, J. Zhang, Y . Deng, A. Li, H. Wang, and S. Liu, “Octonav: Towards generalist embodied navigation,”arXiv preprint arXiv:2506.09839, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    On-policy rl meets off-policy experts: Harmonizing super- vised fine-tuning and reinforcement learning via dynamic weighting,

    W. Zhang, Y . Xie, Y . Sun, Y . Chen, G. Wang, Y . Li, B. Ding, and J. Zhou, “On-policy rl meets off-policy experts: Harmonizing super- vised fine-tuning and reinforcement learning via dynamic weighting,” inICLR, 2026

    2026

  14. [15]

    Room- across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding,

    A. Ku, P. Anderson, R. Patel, E. Ie, and J. Baldridge, “Room- across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding,”arXiv preprint arXiv:2010.07954, 2020

    Pith/arXiv arXiv 2010

  15. [16]

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

    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,” inCVPR, 2018

    2018

  16. [17]

    Beyond the nav-graph: Vision and language navigation in continuous environ- ments,

    J. Krantz, E. Wijmans, A. Majundar, D. Batra, and S. Lee, “Beyond the nav-graph: Vision and language navigation in continuous environ- ments,” inECCV, 2020

    2020

  17. [18]

    Room- across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding,

    A. Ku, P. Anderson, R. Patel, E. Ie, and J. Baldridge, “Room- across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding,” inEMNLP, 2020

    2020

  18. [19]

    Reverie: Remote embodied visual referring expression in real indoor environments,

    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,” inCVPR, 2020

    2020

  19. [20]

    Vision- and-dialog navigation,

    J. Thomason, M. Murray, M. Cakmak, and L. Zettlemoyer, “Vision- and-dialog navigation,” inCoRL, PMLR, 2020

    2020

  20. [21]

    Babywalk: Going farther in vision-and-language navigation by taking baby steps,

    W. Zhu, H. Hu, J. Chen, Z. Deng, V . Jain, E. Ie, and F. Sha, “Babywalk: Going farther in vision-and-language navigation by taking baby steps,” inACL, 2020

    2020

  21. [22]

    Curriculum learning for vision-and- language navigation,

    J. Zhang, J. Fan, J. Peng,et al., “Curriculum learning for vision-and- language navigation,” inNeurIPS, 2021

    2021

  22. [23]

    Landmark-rxr: Solving vision-and-language navigation with fine- grained alignment supervision,

    K. He, Y . Huang, Q. Wu, J. Yang, D. An, S. Sima, and L. Wang, “Landmark-rxr: Solving vision-and-language navigation with fine- grained alignment supervision,” inNeurIPS, 2021

    2021

  23. [24]

    Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation,

    X. Wang, Q. Huang, A. Celikyilmaz, J. Gao, D. Shen, Y .-F. Wang, W. Y . Wang, and L. Zhang, “Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation,” in CVPR, 2019

    2019

  24. [25]

    Activevln: Towards active exploration via multi-turn rl in vision-and- language navigation,

    Z. Zhang, W. Zhu, H. Pan, X. Wang, R. Xu, X. Sun, and F. Zheng, “Activevln: Towards active exploration via multi-turn rl in vision-and- language navigation,”arXiv preprint arXiv:2509.12618, 2025

    arXiv 2025

  25. [26]

    Mobilevla-r1: Reinforcing vision-language-action for mobile robots,

    T. Huang, D. Li, R. Yang, Z. Zhang, Z. Yang, and H. Tang, “Mobilevla-r1: Reinforcing vision-language-action for mobile robots,” arXiv preprint arXiv:2511.17889, 2025

    arXiv 2025

  26. [27]

    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

    Pith/arXiv arXiv 2024

  27. [28]

    Dynam3d: Dynamic layered 3d tokens empower vlm for vision-and-language navigation,

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

    2025

  28. [29]

    Navgpt: Explicit reasoning in vision- and-language navigation with large language models,

    G. Zhou, Y . Hong, and Q. Wu, “Navgpt: Explicit reasoning in vision- and-language navigation with large language models,” inAAAI, 2024

    2024

  29. [30]

    Correctnav: Self-correction flywheel empowers vision- language-action navigation model,

    Z. Yu, Y . Long, Z. Yang, C. Zeng, H. Fan, J. Zhang, and H. Dong, “Correctnav: Self-correction flywheel empowers vision- language-action navigation model,” inAAAI, 2026

    2026

  30. [31]

    Qwen3-vl technical report,

    S. Bai, Y . Cai, R. Chen, K. Chen, X. Chen, othersZesen Cheng, L. Deng, W. Ding, C. Gao, C. Ge, W. Ge, Z. Guo, Q. Huang, J. Huang, F. Huang, B. Hui, S. Jiang, Z. Li, M. Li, M. Li, K. Li, Z. Lin, J. Lin, X. Liu, J. Liu, C. Liu, Y . Liu, D. Liu, S. Liu, D. Lu, R. Luo, C. Lv, R. Men, L. Meng, X. Ren, X. Ren, S. Song, Y . Sun, J. Tang, J. Tu, J. Wan, P. Wang,...

    Pith/arXiv arXiv 2025

  31. [32]

    Efficient-vln: A training-efficient vision-language navigation model,

    D. Zheng, S. Huang, Y . Li, and L. Wang, “Efficient-vln: A training-efficient vision-language navigation model,”arXiv preprint arXiv:2512.10310, 2025

    Pith/arXiv arXiv 2025

  32. [33]

    Beyond the 80/20 rule: High-entropy minority tokens drive effective reinforcement learning for LLM reasoning,

    S. Wang, L. Yu, C. Gao, C. Zheng, S. Liu, R. Lu, K. Dang, X.-H. Chen, J. Yang, Z. Zhang, Y . Liu, A. Yang, A. Zhao, Y . Yue, S. Song, B. Yu, G. Huang, and J. Lin, “Beyond the 80/20 rule: High-entropy minority tokens drive effective reinforcement learning for LLM reasoning,” in NeurIPS, 2025

    2025

  33. [34]

    No prompt left behind: Exploiting zero-variance prompts in LLM reinforcement learning via entropy-guided advantage shaping,

    T.-L. V . Le, M. Jeon, K. Vu, V . D. Lai, and E. Yang, “No prompt left behind: Exploiting zero-variance prompts in LLM reinforcement learning via entropy-guided advantage shaping,” inICLR, 2026

    2026

  34. [35]

    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

    Pith/arXiv arXiv 2017

  35. [36]

    Habitat 2.0: Training home assistants to rearrange their habitat,

    A. Szot, A. Clegg, E. Undersander, E. Wijmans, Y . Zhao, J. Turner, N. Maestre, M. Mukadam, D. Chaplot, O. Maksymets, A. Gokaslan, V . V ondrus, S. Dharur, F. Meier, W. Galuba, A. Chang, Z. Kira, V . Koltun, J. Malik, M. Savva, and D. Batra, “Habitat 2.0: Training home assistants to rearrange their habitat,” inNeurIPS, 2021

    2021

  36. [37]

    Learning to navigate unseen environ- ments: Back translation with environmental dropout,

    H. Tan, L. Yu, and M. Bansal, “Learning to navigate unseen environ- ments: Back translation with environmental dropout,” inProceedings of NAACL-HLT, 2019

    2019

  37. [38]

    Scaling data generation in vision-and-language navigation,

    Z. Wang, J. Li, Y . Hong, Y . Wang, Q. Wu, M. Bansal, S. Gould, H. Tan, and Y . Qiao, “Scaling data generation in vision-and-language navigation,” inICCV, 2023

    2023

  38. [39]

    Bridging the gap between learning in discrete and continuous environments for vision-and- language navigation,

    Y . Hong, Z. Wang, Q. Wu, and S. Gould, “Bridging the gap between learning in discrete and continuous environments for vision-and- language navigation,” inCVPR, 2022

    2022

  39. [40]

    Waypoint models for instruction-guided navigation in continuous environments,

    J. Krantz, A. Gokaslan, D. Batra, S. Lee, and O. Maksymets, “Waypoint models for instruction-guided navigation in continuous environments,” inICCV, 2021

    2021

  40. [41]

    Sim-2-sim transfer for vision-and-language navigation in continuous environments,

    J. Krantz and S. Lee, “Sim-2-sim transfer for vision-and-language navigation in continuous environments,” inECCV, 2022

    2022

  41. [42]

    Gridmm: Grid memory map for vision-and-language navigation,

    Z. Wang, X. Li, J. Yang, Y . Liu, and S. Jiang, “Gridmm: Grid memory map for vision-and-language navigation,” inICCV, 2023

    2023

  42. [43]

    Etpnav: Evolving topological planning for vision-language navigation in continuous environments,

    D. An, H. Wang, W. Wang, Z. Wang, Y . Huang, K. He, and L. Wang, “Etpnav: Evolving topological planning for vision-language navigation in continuous environments,”arXiv preprint arXiv:2304.03047, 2023

    Pith/arXiv arXiv 2023

  43. [44]

    Topological planning with transformers for vision-and-language nav- igation,

    K. Chen, J. K. Chen, J. Chuang, M. V ´azquez, and S. Savarese, “Topological planning with transformers for vision-and-language nav- igation,” inCVPR, 2021

    2021

  44. [45]

    Language-aligned waypoint (law) supervision for vision-and- language navigation in continuous environments,

    S. Raychaudhuri, S. Wani, S. Patel, U. Jain, and A. X. Chang, “Language-aligned waypoint (law) supervision for vision-and- language navigation in continuous environments,”arXiv preprint arXiv:2109.15207, 2021

    Pith/arXiv arXiv 2021

  45. [46]

    Cross-modal map learning for vision and language navigation,

    G. Georgakis, K. Schmeckpeper, K. Wanchoo, S. Dan, E. Miltsakaki, D. Roth, and K. Daniilidis, “Cross-modal map learning for vision and language navigation,” inCVPR, 2022

    2022

  46. [47]

    Weakly-supervised multi-granularity map learning for vision-and- language navigation,

    P. Chen, D. Ji, K. Lin, R. Zeng, T. H. Li, M. Tan, and C. Gan, “Weakly-supervised multi-granularity map learning for vision-and- language navigation,”arXiv preprint arXiv:2210.07506, 2022

    Pith/arXiv arXiv 2022

  47. [48]

    Sim-to-real transfer via 3d feature fields for vision-and-language navigation,

    Z. Wang, X. Li, J. Yang, Y . Liu, and S. Jiang, “Sim-to-real transfer via 3d feature fields for vision-and-language navigation,”arXiv preprint arXiv:2406.09798, 2024

    Pith/arXiv arXiv 2024

  48. [49]

    Embodied navigation foundation model,

    J. Zhang, A. Li, Y . Qi, M. Li, J. Liu, S. Wang, H. Liu, G. Zhou, Y . Wu, X. Li,et al., “Embodied navigation foundation model,”arXiv preprint arXiv:2509.12129, 2025

    arXiv 2025

  49. [50]

    Omninav: A unified framework for prospective exploration and visual-language navigation,

    X. Xue, J. Hu, M. Luo, S. Xie, J. Chen, Z. Xie, K. Quan, W. Guo, M. Xu, and Z. Chu, “Omninav: A unified framework for prospective exploration and visual-language navigation,” inICLR, 2025

    2025

  50. [51]

    Janusvln: Decoupling semantics and spatiality with dual implicit memory for vision-language navigation,

    S. Zeng, D. Qi, X. Chang, F. Xiong, S. Xie, X. Wu, S. Liang, M. Xu, and X. Wei, “Janusvln: Decoupling semantics and spatiality with dual implicit memory for vision-language navigation,” inICLR, 2026

    2026

  51. [52]

    Ground slow, move fast: A dual-system foundation model for generalizable vision-and-language navigation,

    M. Wei, C. Wan, J. Peng, X. Yu, Y . Yang, D. Feng, W. Cai, C. Zhu, T. Wang, J. Pang,et al., “Ground slow, move fast: A dual-system foundation model for generalizable vision-and-language navigation,” inICLR, 2025

    2025