acceptodds
Under review as a conference paper at ICLR 2027

FireNav: Decision-Oriented Reconstruction Evidence Routing for Vision-Language Navigation in Indoor Fire Environments

Abstract

In real-world indoor fire environments, visual degradation severely impairs instruction grounding, waypoint prediction, and long-horizon exploration in vision-language navigation (VLN). Although image restoration offers a natural means to mitigate such visual degradation, existing methods primarily optimize reconstruction fidelity, which does not necessarily translate into better navigation decisions and may even corrupt task-critical cues when applied indiscriminately. In this paper, we present FireNav, a decision-oriented reconstruction evidence routing framework that aligns visual restoration with navigation utility in VLN. To advance this task, we first contribute FireVLN-Bench, a paired indoor fire navigation benchmark with clean references preserving task semantics, safety supervision, and matched settings for continuous navigation and robot execution with simulated dynamics. Our FireNav retains native visual tokens and treats latent reconstruction features as auxiliary evidence, where a lightweight router leverages visual features, task context, and prior execution states to determine whether such evidence should be admitted and how strongly it should contribute to the visual tokens. To learn when reconstruction is beneficial, we further introduce decision utility calibration, which compares action prediction losses with and without candidate evidence under the same state, thereby supervising evidence selection according to its contribution to waypoint, heading, and arrival prediction. Experiments demonstrate that FireNav consistently outperforms state-of-the-art baselines in both navigation success and path efficiency across evaluation settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision-Language Navigation, Embodied Navigation, Indoor Fire Navigation, Visual Reconstruction, Evidence Routing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ExTqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces FireVLN-Bench, comprising 60 indoor scenes with paired clean/degraded observations and two simulated evaluation tracks: continuous navigation and execution with robot dynamics. FireNav freezes the navigator and reconstruction encoder, uses differences in action prediction loss to supervise selective fusion of reconstruction features, and limits interference through bounded residuals and clean-observation retention. Compared with Restored RGB + Adapter, it improves SPL by 4.9/6.0 percentage points on PE/CE, supported by component ablations, evidence interventions, and efficiency analyses.

Critical strengths
  1. The method directly addresses the stated problem and receives targeted empirical support.

Preserving the native visual pathway while admitting reconstruction evidence according to navigation utility reasonably addresses the risk of reconstruction corrupting task-relevant cues. Removing utility calibration reduces PE/CE SPL by 4.0/4.9 points.

  1. The closed-loop gains are substantial, with favorable computational overhead.

Under a shared navigation backbone and execution setup, FireNav outperforms reconstructed-RGB and navigation-adapter baselines, while routing improvements extend across multiple reconstruction encoders.

Critical weaknesses
  1. Decision utility calibration is not fully specified for the slow planning branch.

Equation (4) and Appendix A.3 define evidence utility through waypoint, heading, and arrival prediction losses, whereas the slow branch produces subgoals; the paper does not explain how an evidence intervention in this branch translates into action predictions, including whether subgoal selection and the fast policy are rerun. Since routing across both branches is part of the proposed method, this omission prevents reproduction of a central training procedure rather than merely leaving out an implementation detail.

  1. The evaluation does not establish whether the gains reflect safer navigation decisions in fire environments.

The benchmark explicitly provides risk-aware paths accounting for fire proximity, dense smoke, and blocked routes, yet the paper does not specify which paths supply expert targets or how SPL accounts for safety-driven detours, while Table 9 reports only collisions, falls, stuck episodes, and false stops. Consequently, the results cannot distinguish improved hazard avoidance from more efficient goal reaching under visual degradation, leaving the benchmark’s fire-specific value insufficiently validated.

Questions for authors
  1. How is utility calibration implemented for the slow planning branch? After intervening on its evidence, do you regenerate the subgoal and run the fast policy to compute the action loss, and how are gradients handled through discrete subgoal selection?

  2. Does isolated single-view utility predict the benefit of joint evidence fusion? Please report the fraction of admitted evidence with positive utility on held-out states, the rejection rate of harmful evidence, and whether utility signs remain consistent when other evidence is already admitted.

  3. How are language instructions, expert trajectories, and safety evaluation aligned? Please clarify instruction quality control, whether expert targets follow shortest or risk-aware paths, and how SPL is normalized; can existing trajectories provide fire-proximity or smoke-exposure measurements?

  4. Are the reported gains reproducible and statistically robust? Please provide the training and evaluation configurations, baseline parameter counts, and uncertainty estimates across training seeds and test scenes.

Reviewer_MhtWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

FireNav addresses VLN under simulated indoor fire conditions. It introduces FireVLN-Bench and augments a frozen navigator with selectively routed reconstruction features. Experiments report improved navigation performance over restoration frontends and task-trained adapters, supported by component ablations and comparisons across reconstruction encoders.

Critical strengths
  1. An underexplored problem with practical relevance. The paper studies VLN under complex visual degradation in indoor fire environments, a practically important setting that remains underexplored.

  2. A valuable benchmark for the community. FireVLN-Bench provides paired clean/degraded observations and matched continuous-navigation and simulated robot-execution tracks. This design supports studies of how visual degradation and restoration affect navigation, offering a useful platform for future research on robust embodied navigation.

Critical weaknesses
  1. Lack of comparison with mainstream VLN methods. Table 1 compares restoration frontends and adapters built on the same OmniNav navigator, without evaluating representative VLN baselines such as NaVILA, NaVid and StreamNav. While these comparisons are useful for isolating the proposed modules, they do not establish FireNav’s effectiveness across different navigation architectures. This omission is particularly important for a paper introducing a new VLN benchmark.

  2. The comparisons and training procedure are insufficiently documented. Although the adapters are described as capacity-matched, their architectures and optimization details are unspecified. Calibration settings are also missing. Furthermore, it is unclear how the slow planner is trained using an action-level utility objective through subgoal selection. This makes the experimental results difficult to be convincing and reproducible.

  3. The benchmark’s fire-related objective and reference semantics remain unclear. The benchmark provides both shortest and risk-aware trajectories, but does not identify which defines the expert targets used for utility calibration or how fire/smoke risks affect success and SPL. Table 9 measures physical failures without resolving this distinction. Moreover, Figure 1 appears to remove visible flames from some clean examples, despite the stated preservation of fire cues. These ambiguities affect what “beneficial reconstruction” actually means.

  4. The presentation makes the core idea difficult to follow. The manuscript introduces numerous concepts without clearly explaining. In particular, the slow-branch utility computation, its training supervision, and the expert targets used for calibration remain unclear even after reading the appendix. These gaps require readers to infer implementation and hinder understanding and reproducibility.

Questions for authors
  1. How does FireNav compare with mainstream VLN methods on FireVLN-Bench?

  2. What are the adapter architectures, parameter counts, training budgets, and key calibration settings? Are the baselines and ablations matched in training data and optimization details?

  3. For the slow branch, how are the utility labels in Eq. (4) generated? Are subgoals recomputed before evaluating the downstream action loss? Please also specify the objective used to train the slow evidence interface during candidate learning.

Reviewer_trREOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies vision-language navigation under simulated indoor fire conditions. The authors introduce FireVLN-Bench, a paired indoor fire navigation benchmark with clean references that preserve task semantics, safety supervision, and matched settings for continuous navigation and robot execution with simulated dynamics. They further propose FireNav, a decision-oriented reconstruction evidence routing framework that retains native visual tokens, treats latent reconstruction features as auxiliary evidence, and uses decision utility calibration to learn when this evidence should be admitted. Results in the main paper and appendix show that FireNav improves over the original navigator and several restoration or adaptation baselines under the evaluated settings. I find the problem and benchmark valuable; my main reservations concern the incremental algorithmic contribution and whether the benchmark's fire-specific reliability and intended scope are sufficiently established.

Critical strengths
  1. FireVLN-Bench addresses a meaningful and practically motivated problem. Its paired clean/degraded observations at matched poses and environment states provide a useful basis for studying how visual degradation and reconstruction affect downstream decisions, while the range of degradation factors, tasks, and execution settings reflects substantial dataset and engineering effort. 2. Considering the main paper together with the appendix, the empirical coverage is reasonably complete. The trends across baselines, ablations, and evidence interventions are broadly consistent and support a clear takeaway: reconstructed evidence is not uniformly beneficial, and a navigator should use it selectively according to its decision utility.
Critical weaknesses
  1. FireNav's frozen visual branches, cross-attention fusion, contextual gating, bounded residual injection, clean-retention objective, and caching are closely related to established conditional fusion or routing paradigms. Decision utility calibration is the most distinctive component, but the current evidence does not fully separate gains from its utility-based supervision from those obtainable with a more conventional task-driven gate or from other training differences. The paper reports a w/o utility calibration ablation and a Quality Gate comparison, but does not sufficiently specify how those alternatives are supervised and trained. This does not call into question FireNav's effectiveness in the reported setting, but it directly affects whether the algorithmic contribution clears the acceptance bar. A convincing rebuttal on this point could change my assessment.
  2. FireVLN-Bench appears complete in terms of data modalities, paired structure, and task coverage, and the paper states that every configuration is checked for deterministic replay, connectivity, collision consistency, and pair alignment. However, it does not report the criteria or aggregate outcomes of these checks. The construction and validation of the risk-aware trajectories, safety labels, and information-gain labels - central to the benchmark's fire-specific positioning - are also under-specified. In addition, the basis for the simulated smoke, fire intensity, illumination, glare, and sensor-noise parameters is unclear. The current evidence therefore supports FireVLN-Bench as a useful controlled benchmark for synthetic visual degradation, but it is less clear how strongly it supports claims about fire risk, safety-aware navigation, or real-world indoor fire environments.
Questions for authors
  1. In the w/o utility calibration variant in Table 2, what signal is used to train the admission gate? In Table 5, does Quality Gate use the same routing architecture, parameter count, training data, and optimization budget as FireNav, with only the gate supervision changed? This distinction affects whether the observed gain can be attributed specifically to decision utility calibration.
  2. Relative to learning a gate directly from the navigation objective, what additional information is provided by the offline single-view utility labels? Do the authors view the contribution primarily as a new supervision signal, a new routing mechanism, or a training strategy?
  3. The paper states that each configuration is checked for deterministic replay, connectivity, collision consistency, and clean/degraded pair alignment. What are the main acceptance criteria and aggregate outcomes of these checks? Were any configurations rejected, corrected, or excluded?
  4. The paper states that FireVLN-Bench provides risk-aware trajectories, safety labels, and information-gain labels, and that risk-aware paths account for collision risk, fire proximity, dense smoke, blocked routes, and excessive turning. How are these factors defined and combined, and what validation supports that the resulting trajectories and labels capture the intended risk or safety properties?
  5. What is the basis for the ranges and temporal evolution of smoke density, fire intensity, illumination, glare, and sensor noise? Were they calibrated using a physical model, real data, or expert knowledge? Given the current level of calibration, should FireVLN-Bench be understood primarily as a controlled synthetic indoor-fire visual-degradation benchmark, or as a benchmark representative of navigation in real-world indoor fire environments?
  6. What values of ρ_b are used for the fast and slow branches, and how were these values selected?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.