Pith. sign in

REVIEW 4 major objections 7 minor 1 cited by

Active Test-time Vision-Language Navigation

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read ATENA turns episodic pass/fail human feedback into mixture-entropy updates that adapt a vision-language navigation policy at test time, reporting gains over entropy-minimization baselines on REVERIE, R2R, and R2R-CE.

desk verdict Practical active TTA for VLN with a promising new combination; the scalar episodic feedback assumption is the main thing I'd want addressed before trusting the gains. read the letter →

arxiv 2506.06630 v1 pith:TOQDQNHB submitted 2025-06-07 cs.RO cs.AI

Heeju Ko , Sungjune Kim , Gyeongrok Oh , Jeongyoon Yoon , Honglak Lee , Sujin Jang , Seungryong Kim , Sangpil Kim This is my paper
classification cs.ROcs.AI
keywords vision-languagenavigationtest-timeadaptationactivelearningentropyminimizationhuman-in-the-loopfeedbackmixtureoptimizationembodiedAIuncertaintysampling
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

The paper proposes ATENA, a test-time adaptation method for vision-language navigation that asks a human for a single binary verdict at the end of an episode, rather than step-by-step demonstrations. It uses that verdict to steer an entropy-based update: minimize a mixture entropy for successful episodes and maximize it for failed ones. A self-prediction head supplies the same verdict in low-uncertainty episodes, so the agent keeps adapting even when no human feedback arrives. The paper reports that this scheme improves success rate, path efficiency, and grounding efficiency over the underlying VLN policies and over entropy-minimization baselines on REVERIE, R2R, and R2R-CE. A reader should care because it suggests that a small amount of intuitive episodic human feedback can substitute for dense supervision in online embodied navigation.

What carries the argument

The load-bearing object is the mixture action distribution $q_{\mathrm{mix}}(a|o_t,I) = \lambda q_{\mathrm{pseudo}}(a|a^{\mathrm{sel}}_t) + (1-\lambda)\pi_\theta(a|o_t,I)$, where $q_{\mathrm{pseudo}}$ is a one-hot distribution on the selected action. Its entropy $H(q_{\mathrm{mix}})$ is averaged over the episode, then minimized in successful episodes and maximized in failed ones via $L_{\mathrm{mix}} = I_{\mathrm{success}} H'(q_{\mathrm{mix}}) - (1 - I_{\mathrm{success}}) H'(q_{\mathrm{mix}})$. Because the pseudo-expert term lifts the selected action's probability above the policy's own value, the entropy signal exerts a stronger and more directional gradient on the chosen action than plain policy entropy. The second mechanism is the self-prediction head $f_\phi$, a binary classifier on the averaged hidden state trained with cross-entropy to supply the success or failure label when average episode entropy falls below a threshold $\delta$.

What would settle it

Measure, on failed REVERIE episodes whose path passes within 3 meters of the target at an intermediate step, whether the probability assigned by the policy to oracle-consistent early actions decreases after ATENA adaptation. If it does, the scalar episode-level credit assignment is actively suppressing early correct decisions, and the claim that sign-switched mixture entropy improves calibration is false for exactly those episodes.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that episodic binary success/failure labels are enough to guide test-time entropy adaptation in VLN if the entropy is computed on a mixture of the policy's action distribution and a pseudo-expert one-hot distribution peaked at the selected action. For a successful episode the paper minimizes the mixture entropy, sharpening the policy toward chosen actions; for a failed episode it maximizes the mixture entropy, explicitly spreading probability away from the selected actions. This sign-switched update is intended to fix the failure mode of plain entropy minimization, which tends to make the agent overconfident in wrong actions. The self-active learning head extends the same signal to confident episodes by predicting success or failure from the averaged hidden state, and the joint objective adapts the policy online with a batch size of one.

Load-bearing premise

The gains rest on applying one episode-level success or failure label to every timestep's entropy term, assuming that in a successful episode every chosen action deserves reinforcement and in a failed episode every chosen action deserves suppression; long trajectories with early correct detours or late single mistakes violate this assumption.

Editorial extensions

If this is right

  • On REVERIE Test Unseen, ATENA raises GOAT success rate from 57.72 to 62.03, SPL from 40.53 to 46.82, and RGSPL from 26.70 to 31.54, a gain the paper also reports as third place on the official leaderboard at submission time.
  • On R2R Val Unseen, ATENA improves GOAT success rate from 77.91 to 79.01 and SPL from 67.34 to 69.30 while reducing average trajectory length from 13.43 to 12.52.
  • The uncertainty-gated querying strategy reduces the need for human feedback: with MEO and active learning, DUET reaches 63.70 SR on REVERIE Val Unseen with 55.52% active steps, whereas TENT with the same feedback reaches 55.69 SR with 90.34% active steps.
  • The self-prediction head is accurate enough to serve as a pseudo-label source, reaching 82.19% accuracy on REVERIE Val Unseen, and including SAL adds 6.92 percentage points of SR to DUET on that split.
  • On R2R-CE, the same framework improves path efficiency for BEVBert by 6.7% SPL on Val Seen, showing that episodic feedback transfers to continuous-action navigation, though with smaller gains than on REVERIE.

Reading between the lines

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

  • The paper's scalar credit-assignment rule could be stress-tested directly: for failed episodes that pass near the target early but end far away, entropy maximization may suppress early oracle-consistent actions; measuring that suppression would clarify whether an episode-level label is doing per-step harm.
  • The same mixture-entropy update could apply to other sequential decision tasks that supply an episodic win or loss signal, such as object-goal navigation or instruction following in simulation, as long as a one-hot pseudo-expert over actions can be formed; the paper's limitation section explicitly leaves this open.
  • The self-prediction head's 82.19% accuracy suggests a confidence-controlled weighting scheme: instead of a global constant $\gamma$, the agent could down-weight self-labels when the head's predicted confidence is low, a testable extension the paper does not explore.
  • Since $\lambda=0.4$ outperforms $\lambda=0$, part of the benefit comes from distribution sharpening itself rather than from the outcome-conditioned sign; scheduling $\lambda$ from small to larger values during deployment could be a cheap additional gain, though the paper keeps $\lambda$ fixed.
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 / 7 minor

Summary. The paper introduces ATENA, a test-time adaptation framework for vision-language navigation (VLN) that couples episodic binary success/failure feedback (from a human oracle or a learned self-prediction head) with mixture entropy optimization (MEO). MEO forms a convex mixture of the policy's action distribution and a one-hot 'pseudo-expert' distribution centered on the argmax action, then minimizes the mixture entropy on successful episodes and maximizes it on failed episodes. A self-active learning (SAL) mechanism selects, per episode, whether to request human feedback or to use the self-prediction head's own outcome estimate. The method is evaluated on REVERIE, R2R, and R2R-CE across several base policies (HAMT, DUET, BEVBert, ETPNav, GOAT), reporting consistent gains over TENT and FSTTA, along with ablations on the mixture weight, feedback sampling strategy, and the SAL component, plus a leaderboard comparison and computational-cost analysis.

Significance. If the reported results hold up, ATENA is a practical and novel contribution: it demonstrates that sparse, episodic, binary human feedback can be integrated into test-time adaptation for sequential embodied tasks, with low latency overhead and a self-supervision mechanism to reduce human burden. The paper is also commendable for its breadth of evaluation (three benchmarks, five base policies, multiple ablations), its computational-cost analysis, and its inclusion of a confusion-matrix analysis for the self-prediction head. However, the central empirical claim is weakened by a supervision confound in the headline tables and by the lack of statistical significance reporting; the core MEO mechanism also rests on an unexamined credit-assignment assumption. These issues are substantive but addressable with additional analysis and experiments.

major comments (4)
  1. [Sec. 3.3, Eq. (3)] Equation (3) applies the single episodic binary outcome Isuccess to the episode-averaged mixture entropy H'(qmix). This implicitly assumes that every action in a successful episode is worth reinforcing and every action in a failed episode is worth suppressing. In REVERIE/R2R, success is defined by the final stopping distance, and trajectories can be long; failed episodes often contain correct early actions, and successful episodes can contain detours. Under Eq. (3), entropy maximization on a failed episode will suppress correct early decisions, while entropy minimization on a successful episode will sharpen potentially suboptimal detour actions. The paper does not analyze this credit-assignment issue, and the Limitations section only discusses generalization to other tasks. Since MEO is the core novel component, the reported gains may rest on this unexamined premise. Please provide an analysis of per-step gradient contributions or an ablation that relaxes the uniform credit assignment (e.g., last-step-only entropy, temporally weighted entropy, or per-step pseudo-relevance weights).
  2. [Tables 1–3 vs. Table 4] The headline comparisons in Tables 1–3 evaluate ATENA, which consumes episodic success/failure labels, against TENT and FSTTA, which are label-free. The observed improvements could in principle be due entirely to the additional supervision rather than to MEO or SAL. Table 4 attempts to control for this by integrating active learning into TENT and FSTTA, but it does so only for DUET on REVERIE, and the comparison is confounded by different query rates (e.g., on Val Unseen, DUET+TENT* requests feedback on 90.34% of episodes while DUET+MEO* requests on 55.52%). A matched-budget comparison, or a plot of performance against feedback rate across all base policies and datasets, is needed to separate the effect of the mixture entropy objective from the effect of having labels at all.
  3. [Sec. 4.3–4.7 (overall evaluation)] The paper reports only point estimates, despite stating in Appendix .2 that results are averaged over 3 random seeds. Several reported differences are small (e.g., GOAT SR on R2R Val Unseen from 77.91 to 79.01; DUET SR on R2R Val Unseen shown as 75 vs. 72 without decimals) and could plausibly fall within seed-to-seed variance. Without standard deviations, confidence intervals, or significance tests, the reader cannot assess whether the claimed improvements are reliable. Please report variance across seeds for all main tables and, where meaningful, paired statistical tests.
  4. [Sec. 3.4.2, Eqs. (6)–(7)] The self-prediction head is trained with Lself using labels that, when the feedback source is 'Agent', are the head's own hard thresholded outputs (Eq. 6). This is a self-training loop that may lead to confirmation bias or collapse, and the paper does not discuss this risk or provide safeguards beyond the global weight gamma. The confusion matrix in Table 7 reports 82.19% overall accuracy, but the evaluation protocol is unclear: is this accuracy measured after adaptation on the validation set, and how does it evolve over time as the head is updated with its own pseudo-labels? Please clarify the protocol and provide an analysis of self-label accuracy during adaptation, along with a discussion of the potential pitfalls of self-training in this online setting.
minor comments (7)
  1. [Eq. (3)] The loss can be written more compactly as Lmix = (2Isuccess − 1) H'(qmix); the current two-term form obscures that the only difference between success and failure is a sign flip.
  2. [Table 4 caption] The caption states that Active (%) is 'the ratio of navigation steps where feedback is requested', but feedback is requested per episode, not per step; please clarify whether this is the fraction of episodes receiving human feedback.
  3. [Sec. 4.3, REVERIE paragraph] The phrase 'improves the SR metric in the validation unseen split up to 3.19%, 44.98% and 25.72%' uses relative percentages without stating that they are relative; please specify absolute vs. relative changes to avoid misleading the reader.
  4. [Algorithm 1 / Eq. (5)] The average entropy in Eq. (5) sums t=1 to T, whereas Algorithm 1 sums t=0 to T−1; since both are averaged, this is only an indexing inconsistency, but please align them.
  5. [Fig. 2 caption] The caption contains a typo: 'llustration' should be 'Illustration'.
  6. [Appendix .2] The sentence 'The rest of the experimental configurations strictly follow the ones of the pre-trained navigation policies' is grammatically awkward; consider revising to 'All other experimental configurations follow those of the pre-trained navigation policies.'
  7. [Sec. 4.6 / Fig. 3] Figure 3 is said to average results over three seeds, but no error bars are shown; please add them for consistency with the text.

Circularity Check

1 steps flagged · score 4.0 of 10

SAL's self-prediction head trains on its own thresholded outputs in agent-labeled episodes, creating a self-referential label loop; external benchmarks keep the central claim partly independent.

  1. self definitional [Section 3.4.2, Eqs. (6)-(7); Algorithm 1 lines 7-10]
    "(6) Isuccess = 1, if σ(fϕ(savg)) > 0.5, 0, otherwise ... (7) Lself = −[Isuccess log(σ(fϕ(savg)) + (1 − Isuccess) log(1 − σ(fϕ(savg)))] ... if the feedback oracle is the agent itself, this can be interpreted as a self-training paradigm with pseudo label derived from the agent's own assessment of task completion"

    For agent-oracle episodes, Eq. (6) defines the training target Isuccess as the thresholded output of the self-prediction head fϕ itself, and Eq. (7) then trains fϕ against that same self-generated label; Algorithm 1 also feeds this same Isuccess into Eq. (3) to set the entropy-minimize/maximize direction. Substituting Eq. (6) into Eq. (7) gives Lself = -log(max(σ(fϕ), 1-σ(fϕ))), which only pushes the sigmoid away from 0.5 and carries no independent evidence about navigation outcome. The SAL improvement in Table 5 therefore rests partly on a self-confirming confidence-sharpening loop rather than on an external success/failure signal. Table 7's offline confusion matrix provides an external anchor, so the loop is not completely ungrounded.

full rationale

ATENA's headline results are evaluated on held-out REVERIE, R2R, and R2R-CE splits against Tent and FSTTA, so the central empirical claim is externally anchored rather than derived from the method's own equations. I found no load-bearing self-citation chain and no imported uniqueness theorem: the paper cites standard TTA/AL work and its baselines without leaning on the authors' own prior results. The mixture entropy objective (Eqs. 1-3) is also not circular by itself: the pseudo-expert is a one-hot at the policy's own argmax, but the sign of the loss is set by an external episodic success/failure label, so it acts as a reward-modulated entropy regularizer rather than a prediction that is its own input. The reported Eq. (3) does assume that one binary outcome applies to every timestep's entropy; that is a credit-assignment assumption, not a definitional identity. The one genuine self-referential loop is in SAL: when the entropy-based selector routes an episode to the agent oracle, Eq. (6) defines Isuccess as the thresholded output of fϕ itself, and Eq. (7) then trains fϕ with that same self-generated label (and Eq. (3) uses the same label to set the entropy direction). For such episodes, Lself reduces to -log(max(σ(fϕ),1-σ(fϕ))), which only pushes the model's confidence away from 0.5 and supplies no independent evidence about navigation success. Table 5 attributes part of ATENA's gain to SAL, so this self-labeling loop is load-bearing, although Table 7's offline confusion matrix (82.19% accuracy against true labels) provides an external check that keeps the loop from being completely unanchored. Overall, the method is a test-time learning procedure whose main results are empirically validated on held-out splits, so the circularity is partial and localized to the self-supervision branch.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The method depends on outcome labels that are external for human-queried episodes and self-generated for agent-labeled episodes; the self-referential pseudo-expert and self-prediction head add internal feedback loops. The scalar credit-assignment axiom, where one final label guides every step's entropy update, is the least examined assumption.

free parameters (4)
  • lambda (mixture weight) = 0.4 (best in REVERIE sweep)
    Controls how strongly the pseudo-expert one-hot sharpens the mixture action distribution in Eq. 1; swept over 0.0 to 1.0 in 0.1 steps and selected by validation performance.
  • delta (uncertainty threshold) = 0.1 (used in Tables 4 and 5)
    Decides whether an episode receives human feedback or self-generated feedback in Eq. 5; searched over {0.1, 0.2, 0.3}.
  • gamma (self-loss weight) = not reported
    Balances Lself in Eq. 8; the paper mentions reducing gamma to mitigate imperfect self-labels but does not give the value used in experiments.
  • Adaptation learning rates = chosen from {5e-6, 1e-6, 5e-7} for ATENA; modified sets for FSTTA
    Learning rates for online policy updates are tuned per method; no final per-dataset values are reported in the paper.
assumptions (5)
  • domain assumption Pre-trained VLN policies provide hidden states and action distributions that remain informative under distribution shift
    The entire adaptation procedure operates on internal states s_t and action distribution pi_theta of HAMT, DUET, BEVBert, ETPNav, or GOAT without offline retraining.
  • domain assumption A binary episodic success/failure label is an accurate and available test-time signal that can be supplied by a human oracle or a learned self-oracle
    Eq. 5 and Algorithm 1 rely on this label to choose entropy minimization versus maximization; the paper does not quantify human labeling cost or label noise.
  • ad hoc to paper The same episodic outcome can be used as credit assignment for every timestep's entropy in the episode
    Lmix in Eq. 3 applies one Isuccess to the average mixture entropy over all steps, reinforcing or suppressing every selected action uniformly on the basis of final success.
  • ad hoc to paper The policy's own argmax action in a successful episode is a reasonable proxy for an expert action
    qpseudo in Eq. 1 is a one-hot on asel_t, the action already chosen by the policy, so the pseudo-expert distribution is self-referential.
  • domain assumption A self-prediction head can learn a reliable success predictor from online episodes and can safely label confident episodes from the start of streaming
    SAL uses f_phi(s_avg) as the label source for low-entropy episodes even though phi is initialized at test time; the confusion matrix is only measured post hoc.
invented entities (2)
  • Self-prediction head f_phi
    purpose: Predicts episodic success/failure from averaged hidden states to generate pseudo-labels for confident episodes (Eq. 6) and is trained with binary cross-entropy (Eq. 7).
    It is a new test-time module with no external supervision or falsifiable prediction outside the adaptation loop; its reliability is only evaluated on the same benchmark after adaptation.
  • Pseudo-expert distribution qpseudo
    purpose: A one-hot distribution on the policy's selected action used to sharpen the mixture action distribution in Eq. 1.
    It is a construct defined from the policy itself, not an external expert or dataset, so it cannot independently validate the chosen action.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Active Test-time Vision-Language Navigation." pith.science (2026). https://pith.science/paper/TOQDQNHB

@misc{pith2026250606630,
  author       = {Pith},
  title        = {Pith review of: Active Test-time Vision-Language Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TOQDQNHB}},
  note         = {Machine review of arXiv:2506.06630}
}
read the original abstract

Vision-Language Navigation (VLN) policies trained on offline datasets often exhibit degraded task performance when deployed in unfamiliar navigation environments at test time, where agents are typically evaluated without access to external interaction or feedback. Entropy minimization has emerged as a practical solution for reducing prediction uncertainty at test time; however, it can suffer from accumulated errors, as agents may become overconfident in incorrect actions without sufficient contextual grounding. To tackle these challenges, we introduce ATENA (Active TEst-time Navigation Agent), a test-time active learning framework that enables a practical human-robot interaction via episodic feedback on uncertain navigation outcomes. In particular, ATENA learns to increase certainty in successful episodes and decrease it in failed ones, improving uncertainty calibration. Here, we propose mixture entropy optimization, where entropy is obtained from a combination of the action and pseudo-expert distributions-a hypothetical action distribution assuming the agent's selected action to be optimal-controlling both prediction confidence and action preference. In addition, we propose a self-active learning strategy that enables an agent to evaluate its navigation outcomes based on confident predictions. As a result, the agent stays actively engaged throughout all iterations, leading to well-grounded and adaptive decision-making. Extensive evaluations on challenging VLN benchmarks-REVERIE, R2R, and R2R-CE-demonstrate that ATENA successfully overcomes distributional shifts at test time, outperforming the compared baseline methods across various settings.

Figures

Figures reproduced from arXiv: 2506.06630 by the authors.

Figure 1
Figure 1. Overview of the ATENA adaptation framework. At each navigation step, the agent stores state and entropy information in its memory. Once the episode ends, the stored entropy is used to determine the feedback source: human oracle for uncertain episodes, and self oracle for certain episodes. Self oracle utilizes a self-prediction head, trained during online test-time, enabling the agent to autonomously predict navigati… view at source ↗
Figure 2
Figure 2. llustration of Mixture Entropy Optimization (MEO). (a) The Mixture Action Distribu￾tion is constructed by combining the action distribution (yellow) with a pseudo-expert distribution (red). (b) Mixture entropy is minimized for successful episodes to encourage the correct actions, and maximized for failures to penalize incorrect ones. where At is the set of all possible actions at step t. We average the entropy over … view at source ↗
Figure 3
Figure 3. Effect of the combination weight λ. Performance comparison with different combination weight λ in Mixture Entropy Optimization. λ = 0 corresponds to vanilla entropy without distribution mix. The results are averaged across three experiments with different seeds. benefit of distribution sharpening. The performance peaks at λ = 0.4, where the balance between the predicted distribution and the pseudo-expert distributio… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Different Episode Sampling Strategies. Our uncertainty-based sampling outperforms baselines and remains competitive against full-feedback settings. We compare the uncertainty-based active learning strategy with two different sam￾pling baselines: (1) Random Episodes, wh…
Figure 5
Figure 5. Figure 5: In Step 3, Trial 1 selected an incorrect action with 94.5% confidence. After ATENA’s adaptation, Trial 2 correctly [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: In Step 2, Trial 1 incorrectly selected an action with 85.4% confidence. After ATENA’s adaptation, Trial 2 correctly [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: In Step 3, Trial 1 incorrectly selected an action with 98.1% confidence. After adaptation with ATENA, Trial 2 [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. VTM-Nav: Harnessing Cross-Episode Experience for Object-Goal Navigation with Hierarchical Visual-Topological Memory

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A hierarchical room-and-object memory that persists across independent ObjectNav episodes yields small success-rate gains, but most of the gain comes from within-episode memory rather than the cross-episode component.

Reference graph

Works this paper leans on

58 extracted references · 40 canonical work pages · cited by 1 Pith paper

  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 Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3674–3683, 2018

    work page 2018

  2. [2]

    Counterfactual vision-and-language navigation: Unravelling the unseen

    Amin Parvaneh, Ehsan Abbasnejad, Damien Teney, Javen Qinfeng Shi, and Anton Van den Hengel. Counterfactual vision-and-language navigation: Unravelling the unseen. Advances in neural information processing systems, 33:5296–5307, 2020

    work page 2020

  3. [3]

    Envedit: Environment editing for vision-and-language navigation

    Jialu Li, Hao Tan, and Mohit Bansal. Envedit: Environment editing for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15407–15417, 2022

    work page 2022

  4. [4]

    Think global, act local: Dual-scale graph transformer for vision-and-language navigation

    Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16537–16547, 2022

    2022

  5. [5]

    History aware multimodal transformer for vision-and-language navigation

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

    work page 2021

  6. [6]

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

    Dong An, Hanqing Wang, Wenguan Wang, Zun Wang, Yan Huang, Keji He, and Liang Wang. Etpnav: Evolving topological planning for vision-language navigation in continuous environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    work page 2024

  7. [7]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020

    arXiv 2006

  8. [8]

    Continual test-time domain adaptation

    Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7201–7211, June 2022

    2022

Show all 58 references
  1. [9]

    Test time adaptation via conjugate pseudo-labels

    Sachin Goyal, Mingjie Sun, Aditi Raghunathan, and J Zico Kolter. Test time adaptation via conjugate pseudo-labels. Advances in Neural Information Processing Systems, 35:6204–6218, 2022

    2022

  2. [10]

    Towards test time adaptation via calibrated entropy minimization

    Hao Yang, Min Wang, Jinshen Jiang, and Yun Zhou. Towards test time adaptation via calibrated entropy minimization. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3736–3746, 2024

    2024

  3. [11]

    Test-time adaptation on noisy data via model-pruning- based filtering and flatness-aware entropy minimization

    Xingzhi Zhou, Zhiliang Tian, Boyang Zhang, Yibo Zhang, Ka Chun Cheung, Simon See, Hao Yang, Yun Zhou, and Nevin L Zhang. Test-time adaptation on noisy data via model-pruning- based filtering and flatness-aware entropy minimization. In Proceedings of the AAAI Conference on Arti...

    2025

  4. [12]

    Active learning literature survey

    Burr Settles. Active learning literature survey. 2009

    2009

  5. [13]

    A survey on deep active learning: Recent advances and new frontiers

    Dongyuan Li, Zhen Wang, Yankai Chen, Renhe Jiang, Weiping Ding, and Manabu Okumura. A survey on deep active learning: Recent advances and new frontiers. IEEE Transactions on Neural Networks and Learning Systems, 2024

    2024

  6. [14]

    Entropy-based active learning for object detection with progressive diversity constraint

    Jiaxi Wu, Jiaxin Chen, and Di Huang. Entropy-based active learning for object detection with progressive diversity constraint. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9397–9406, 2022

    2022

  7. [15]

    Active learning helps pretrained models learn the intended task

    Alex Tamkin, Dat Nguyen, Salil Deshpande, Jesse Mu, and Noah Goodman. Active learning helps pretrained models learn the intended task. Advances in Neural Information Processing Systems, 35:28140–28153, 2022

    2022

  8. [16]

    Margin-based sampling in high dimensions: When being active is less efficient than staying passive

    Alexandru Tifrea, Jacob Clarysse, and Fanny Yang. Margin-based sampling in high dimensions: When being active is less efficient than staying passive. InInternational Conference on Machine Learning, pages 34222–34262. PMLR, 2023. 10

    2023

  9. [17]

    Reverie: Remote embodied visual referring expression in real indoor environments

    Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages...

    2020

  10. [18]

    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 Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVIII 16, page...

    2020

  11. [19]

    Vision-language navigation with embodied intelligence: A survey

    Peng Gao, Peng Wang, Feng Gao, Fei Wang, and Ruyue Yuan. Vision-language navigation with embodied intelligence: A survey. arXiv preprint arXiv:2402.14304, 2024

    2024 arXiv

  12. [20]

    Vision-language navigation: a survey and taxonomy

    Wansen Wu, Tao Chang, Xinmeng Li, Quanjun Yin, and Yue Hu. Vision-language navigation: a survey and taxonomy. Neural Computing and Applications, 36(7):3291–3316, 2024

    2024

  13. [21]

    Vision-and-language navigation: A survey of tasks, methods, and future directions

    Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Eric Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. arXiv preprint arXiv:2203.12667, 2022

    2022 arXiv

  14. [22]

    Neighbor-view en- hanced model for vision and language navigation

    Dong An, Yuankai Qi, Yan Huang, Qi Wu, Liang Wang, and Tieniu Tan. Neighbor-view en- hanced model for vision and language navigation. InProceedings of the 29th ACM International Conference on Multimedia, pages 5101–5109, 2021

    2021

  15. [23]

    Vision-language navigation policy learning and adaptation

    Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Vision-language navigation policy learning and adaptation. IEEE transactions on pattern analysis and machine intelligence, 43(12):4205–4216, 2020

    2020

  16. [24]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  17. [25]

    V olumetric environment representation for vision- language navigation

    Rui Liu, Wenguan Wang, and Yi Yang. V olumetric environment representation for vision- language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16317–16328, 2024

    2024

  18. [26]

    Towards learning a generic agent for vision-and-language navigation via pre-training

    Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13137–13146, 2020

    2020

  19. [27]

    A dual semantic-aware recurrent global-adaptive network for vision-and-language navigation

    Liuyi Wang, Zongtao He, Jiagui Tang, Ronghao Dang, Naijia Wang, Chengju Liu, and Qijun Chen. A dual semantic-aware recurrent global-adaptive network for vision-and-language navigation. arXiv preprint arXiv:2305.03602, 2023

    2023 arXiv

  20. [28]

    Multimodal attention networks for low-level vision-and-language navigation

    Federico Landi, Lorenzo Baraldi, Marcella Cornia, Massimiliano Corsini, and Rita Cucchiara. Multimodal attention networks for low-level vision-and-language navigation. Computer vision and image understanding, 210:103255, 2021

    2021

  21. [29]

    Robust navigation with language pretraining and stochastic sampling

    Xiujun Li, Chunyuan Li, Qiaolin Xia, Yonatan Bisk, Asli Celikyilmaz, Jianfeng Gao, Noah Smith, and Yejin Choi. Robust navigation with language pretraining and stochastic sampling. arXiv preprint arXiv:1909.02244, 2019

    1909 arXiv

  22. [30]

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

    Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641–7649, 2024

    2024

  23. [31]

    Discuss before moving: Visual language navigation via multi-expert discussions

    Yuxing Long, Xiaoqi Li, Wenzhe Cai, and Hao Dong. Discuss before moving: Visual language navigation via multi-expert discussions. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 17380–17387. IEEE, 2024

    2024

  24. [32]

    Navgpt-2: Unleashing navigational reasoning capability for large vision-language models

    Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. Navgpt-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2024. 11

    2024

  25. [33]

    Fast-slow test-time adaptation for online vision-and- language navigation

    Junyu Gao, Xuan Yao, and Changsheng Xu. Fast-slow test-time adaptation for online vision-and- language navigation. In International Conference on Machine Learning, pages 14902–14919. PMLR, 2024

    2024

  26. [34]

    Context-guided entropy minimization for semi-supervised domain adaptation

    Ning Ma, Jiajun Bu, Lixian Lu, Jun Wen, Sheng Zhou, Zhen Zhang, Jingjun Gu, Haifeng Li, and Xifeng Yan. Context-guided entropy minimization for semi-supervised domain adaptation. Neural Networks, 154:270–282, 2022

    2022

  27. [35]

    Advent: Ad- versarial entropy minimization for domain adaptation in semantic segmentation

    Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Ad- versarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2517–2526, 2019

    2019

  28. [36]

    Entropy minimization versus diversity maximization for domain adaptation

    Xiaofu Wu, Suofei Zhang, Quan Zhou, Zhen Yang, Chunming Zhao, and Longin Jan Latecki. Entropy minimization versus diversity maximization for domain adaptation. IEEE Transactions on Neural Networks and Learning Systems, 34(6):2896–2907, 2021

    2021

  29. [37]

    Semi-supervised learning by entropy minimization

    Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004

    2004

  30. [38]

    Mixmatch: A holistic approach to semi-supervised learning

    David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019

    2019

  31. [39]

    Semi-supervised semantic segmentation via entropy minimization

    Jiawei Wu, Haoyi Fan, Xiaoqing Zhang, Shouying Lin, and Zuoyong Li. Semi-supervised semantic segmentation via entropy minimization. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1–6. IEEE, 2021

    2021

  32. [40]

    A comprehensive survey on test-time adaptation under distribution shifts

    Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, 133(1):31–64, 2025

    2025

  33. [41]

    Beyond model adaptation at test time: A survey

    Zehao Xiao and Cees GM Snoek. Beyond model adaptation at test time: A survey. arXiv preprint arXiv:2411.03687, 2024

    2024 arXiv

  34. [42]

    Unified entropy optimization for open-set test-time adaptation

    Zhengqing Gao, Xu-Yao Zhang, and Cheng-Lin Liu. Unified entropy optimization for open-set test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23975–23984, 2024

    2024

  35. [43]

    Note: Robust continual test-time adaptation against temporal correlation

    Taesik Gong, Jongheon Jeong, Taewon Kim, Yewon Kim, Jinwoo Shin, and Sung-Ju Lee. Note: Robust continual test-time adaptation against temporal correlation. Advances in Neural Information Processing Systems, 35:27253–27266, 2022

    2022

  36. [44]

    Uncertainty-calibrated test-time model adaptation without forgetting

    Mingkui Tan, Guohao Chen, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Peilin Zhao, and Shuaicheng Niu. Uncertainty-calibrated test-time model adaptation without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  37. [45]

    Entropy is not enough for test-time adaptation: From the perspective of disentangled factors

    Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. arXiv preprint arXiv:2403.07366, 2024

    2024 arXiv

  38. [46]

    A survey of deep active learning

    Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning. ACM computing surveys (CSUR), 54(9):1–40, 2021

    2021

  39. [47]

    A survey on active learning and human- in-the-loop deep learning for medical image analysis

    Samuel Budd, Emma C Robinson, and Bernhard Kainz. A survey on active learning and human- in-the-loop deep learning for medical image analysis. Medical image analysis, 71:102062, 2021

    2021

  40. [48]

    A wholistic view of continual learning with deep neural networks: Forgotten lessons and the bridge to active and open world learning

    Martin Mundt, Yongwon Hong, Iuliia Pliushch, and Visvanathan Ramesh. A wholistic view of continual learning with deep neural networks: Forgotten lessons and the bridge to active and open world learning. Neural Networks, 160:306–336, 2023

    2023

  41. [49]

    Few-shot continual active learning by a robot.Advances in Neural Information Processing Systems, 35:30612–30624, 2022

    Ali Ayub and Carter Fendley. Few-shot continual active learning by a robot.Advances in Neural Information Processing Systems, 35:30612–30624, 2022. 12

    2022

  42. [50]

    Active learning for continual learning: Keeping the past alive in the present

    Jaehyun Park, Dongmin Park, and Jae-Gil Lee. Active learning for continual learning: Keeping the past alive in the present. arXiv preprint arXiv:2501.14278, 2025

    2025 arXiv

  43. [51]

    Transferable query selection for active domain adaptation

    Bo Fu, Zhangjie Cao, Jianmin Wang, and Mingsheng Long. Transferable query selection for active domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7272–7281, 2021

    2021

  44. [52]

    Active test-time adaptation: Theoretical analyses and an algorithm

    Shurui Gui, Xiner Li, and Shuiwang Ji. Active test-time adaptation: Theoretical analyses and an algorithm. arXiv preprint arXiv:2404.05094, 2024

    2024 arXiv

  45. [53]

    Effortless active labeling for long-term test-time adaptation

    Guowei Wang and Changxing Ding. Effortless active labeling for long-term test-time adaptation. arXiv preprint arXiv:2503.14564, 2025

    2025 arXiv

  46. [54]

    Bevbert: Multimodal map pre-training for language-guided navigation

    Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Multimodal map pre-training for language-guided navigation. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023

    2023

  47. [55]

    Vision- and-language navigation via causal learning

    Liuyi Wang, Zongtao He, Ronghao Dang, Mengjiao Shen, Chengju Liu, and Qijun Chen. Vision- and-language navigation via causal learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13139–13150, 2024

    2024

  48. [56]

    Rrex- bot: Remote referring expressions with a bag of tricks

    Gunnar A Sigurdsson, Jesse Thomason, Gaurav S Sukhatme, and Robinson Piramuthu. Rrex- bot: Remote referring expressions with a bag of tricks. in 2023 ieee. In RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5203–5210

    2023

  49. [57]

    Matterport3d: Learning from rgb-d data in indoor environments, 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, 2017

    2017

  50. [58]

    Turn on the faucet in the bathroom with a dark green hand towel hanging from the towel rack

    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. In Proceedings of the IEEE/CVF International Conferen...

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.