acceptodds
Under review as a conference paper at ICLR 2027

Acoustic World Models for Active Scene Reconstruction

Abstract

Room impulse responses provide an indirect acoustic probe of scene geometry: as an agent moves, the recorded reverberation changes with nearby walls, openings, and free space. However, converting acoustic observations into maps is inherently ambiguous, since different reflector configurations can produce similar responses, especially when observations are sparse or motion is limited. We study active acoustic scene reconstruction, where an agent must choose sensing poses that improve its geometric belief rather than follow a predefined scan. We introduce a cross-modal acoustic world model that encodes histories of RIRs and known poses into a motion-conditioned latent state used for both local occupancy decoding and future acoustic-latent prediction. At test time, candidate trajectories are rolled out in latent space, decoded into imagined occupancy maps, and scored by predicted map-space information gain. We construct a synthetic benchmark of paired acoustic trajectories, poses, and floor-plan geometry. Experiments show improved local acoustic-to-geometry reconstruction over geometric and passive baselines, and closed-loop mapping that matches or improves frontier-based exploration while substantially reducing collisions. Code and datasets will be made available upon publication.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Acoustic Sensing, Navigation, Active-Sensing

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_WawoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies active scene reconstruction from single-channel RIRs and known poses. It proposes an acoustic world model that jointly reconstructs local wall maps and predicts future acoustic representations. These predictions support trajectory selection through anticipated map improvement and collision checking. Experiments evaluate local reconstruction and closed-loop mapping in simulation.

Critical strengths
  1. The integration of acoustic reconstruction, future observation prediction, and closed-loop planning provides a coherent approach to active acoustic mapping. The reported reduction in mean collision rate suggests that the integrated predictive planner can support safer exploration.
Critical weaknesses
  1. The idealized sensing and environment assumptions limit the practical conclusions. (1) The experiments directly use simulator-generated RIRs without evaluating a recording pipeline in which a source signal is convolved with the RIR and corrupted by noise. The existing RIR-level noise experiments do not cover this acquisition and estimation process. (2) Evaluation focuses on empty room or building envelopes, without explicitly considering furniture such as tables and beds or the acoustic effects of doors and windows. The evaluation primarily treats individual rooms as separate reconstruction units, leaving reconstruction across connected multi-room environments underexplored.
  2. The comparative evaluation does not sufficiently establish the proposed method’s advantage or isolate its contribution. (1) The reconstruction baselines comprise two trajectory-only heuristics and one custom EchoScan-style predictor. Relevant acoustic floorplan reconstruction methods, such as the audio-only variant of AV-Map [1], may be considered for adaptation, retraining, and evaluation under matched conditions. (2) Moreover, a baseline with the same reconstruction architecture, trained without the future-prediction objective, is needed to distinguish the benefits of world-model training from those of the encoder and sequence aggregation. (3) The presentation would also benefit from analyses of how reconstruction performance varies across different conditions, such as room type, floor area, or reverberation time. Such analyses would complement the existing results and help clarify the conditions under which the method performs well or struggles. Videos comparing the complete reconstruction process across methods would also help make their differences easier to assess.

[1] S. Purushwalkam et al., "Audio-Visual Floorplan Reconstruction," 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 1163-1172.

Questions for authors
  1. EchoScan’s official implementation is publicly available at https://github.com/inmoyeon/EchoScan. Could the authors update the statement that its code is unavailable and verify the fidelity of their adaptation, including the architectural choices, training procedure, and output conversion used for evaluation?
  2. The main text describes monaural omnidirectional sensing, whereas Appendix A states that acoustic descriptors are computed using the left binaural channel. Binaural rendering includes a head-related transfer function, selecting one ear would retain head-shadow and directional filtering effects and would not be equivalent to an omnidirectional microphone. Please clarify the rendering configuration and discuss the trade-offs involved in choosing monaural sensing over binaural audio, FOA, or a small microphone array.
  3. Could the appendix be condensed and reorganized around the essential methodological details and findings? It contains numerous large reconstruction figures which occupy substantial space and make it difficult to identify the main findings.
  4. Could the authors compare their work with MapLearn in terms of the problems addressed, research scope, and reconstruction capabilities, clarifying which aspects represent advances and which involve a narrower scope or reduced capabilities?
7 comments

We thank the reviewer for the thoughtful review. We wish to address raised concerns.

  1. About acoustic acquisition and environmental scope.

    • Regarding realistic acoustic acquisition.

      Our sensing setup assumes an active acoustic source: the robot emits a short acoustic impulse and records the resulting response, providing a direct means of measuring an RIR. Practical excitation signals have finite bandwidth, and recordings are affected by noise and measurement artifacts. Our existing RIR-level noise experiments characterize sensitivity to additive acoustic measurement errors.

      The underlying layouts are derived from real-world scans, panoramic observations, and residential floor plans, while acoustic propagation is simulated using SoundSpaces 2.0, which has previously been evaluated against measured RIRs and shown to reproduce key room-acoustic characteristics in real environments.

      To further evaluate acquisition effects, we introduce a simulated measurement pipeline in which a known exponential sine sweep is convolved with the simulated RIR, recording noise is added at specified SNR levels, and the RIR is recovered through regularized deconvolution. The resulting estimated RIR is provided to the unchanged reconstruction model.

      Recording SNRChamfer (m) ↓DF-MAE (m) ↓[email protected] ↑[email protected] ↑[email protected] ↑
      Clean0.1460.0960.8160.9580.976
      30 dB0.1470.0970.8150.9580.976
      20 dB0.1460.0970.8130.9570.975
      10 dB0.1820.1410.7340.9040.949
      5 dB0.3320.2790.5560.7270.811
      3 dB0.4250.3600.4890.6490.733
      1 dB0.5140.4370.4360.5880.673

      Reconstruction remains close to the clean baseline at recording SNRs of 20-30 dB, with increasing degradation at lower SNRs. This experiment evaluates excitation, noisy recording, and inverse filtering under controlled conditions. It assumes known excitation and linear acoustics.

    • Regarding furniture and multi-room environments.

      Our evaluation includes geometrically diverse room and apartment envelopes, including nonconvex layouts, but does not model furniture, heterogeneous acoustic materials, or the detailed effects of doors and windows. The reconstruction targets focus on room and apartment boundaries, with internal apartment subdivisions omitted.

      Internal clutter introduces additional reflections and occlusions, while connected rooms involve more complex acoustic propagation and topology. We will clarify that the present study evaluates predictive active acoustic mapping under controlled geometric assumptions, with furnished, connected buildings deferred to future work - mostly because of the difficulty in acquiring such diverse data accurately and at scale.

  1. About comparisons, experimental analysis, and presentation.

    • Regarding comparison with AV-Map.

      We appreciate the suggestion to consider AV-Map. Its audio-only configuration is relevant to acoustic floorplan reconstruction, and we will expand the related-work discussion accordingly.

      AV-Map addresses floorplan and semantic-layout inference from collected observations, while our work focuses on RIR-conditioned local wall reconstruction, active sampling, future acoustic-latent prediction, and closed-loop trajectory optimization with predictive collision evaluation.

      We have not performed a comparison with AV-Map, however we do emphasize the extension of AV-Map to our task is non-trivial. A matched empirical comparison would require adapting the acoustic inputs, reconstruction targets, and evaluation protocol that AV-Map is adapted for.

    • Regarding contribution of predictive training.

      We additionally train a geometry-only variant with λpred=0\lambda_{\mathrm{pred}}=0, retaining the acoustic encoder, pose-conditioned temporal model, occupancy decoder, geometric supervision, dataset, and training budget of 300 epochs.

      Removing predictive supervision slightly reduces local reconstruction accuracy: Chamfer increases from 0.146 to 0.150 m, while [email protected] decreases from 0.958 to 0.953. Results show modest reconstruction benefit from predictive supervision, future acoustic prediction is primarily motivated by its use in active planning.

    • Regarding reconstruction under different scene conditions.

      We extend the evaluation with scene-level analysis across floor area, geometric complexity, and reverberation time (RT60). For each characteristic, all 287 held-out ResPlan scenes are divided into low, medium, and high groups, and reconstruction accuracy is aggregated across five runs.

      We observe floor area has little association with reconstruction accuracy: Chamfer remains approximately 0.253 m across all three groups, with [email protected] ranging from 0.687 to 0.698.

      Geometric complexity is more associated with reconstruction difficulty. From low to high complexity, Chamfer increases from 0.232 to 0.283 m, while [email protected] decreases from 0.720 to 0.653.

      Higher estimated RT60 correlates with better reconstruction in this dataset: [email protected] increases from 0.659 to 0.723, while Chamfer decreases from 0.275 to 0.227 m.

    • Regarding Appendix organization.

      We accept the comment and thank the reviewer for raising it. We have uploaded a revised manuscript that prioritizes experimental protocols, controlled ablations, and principal results while unifying repetitive qualitative reconstruction figures.

    • Regarding qualitative presentation. We recognize the value of side-by-side videos showing reconstruction evolution and agent trajectories. To that end, we uploaded a supplementary zip file containing short episodes across different scenes: one comparing CEM variants, another comparing CEM (WM) against non-CEM baselines, and a final video illustrating CEM (WM) trajectory sampling and hallucinations. We hope these videos provide clearer intuition into reconstruction quality and planner behavior.

  2. About EchoScan implementation.

    We correct our statement regarding code availability. The repository's first public commit is from around a year ago, after development of our baseline had begun – therefore we have apparently missed the official implementation’s release. Our adaptation was developed from the published paper.

    EchoScan is the closest learned acoustic reconstruction baseline to our setting. We implemented the transferable components and verified that we can reproduce the original method's reported results in its native experimental setting.

    In any case it is worth noting that our task requires substantial adaptations because the sensing configurations and reconstruction targets differ. EchoScan uses a fixed multichannel microphone array to infer a filled room floorplan and height. Our setting uses single-channel RIRs acquired along a moving trajectory, explicitly conditions on sensor poses, and predicts local wall occupancy. Accordingly, we adapt the input encoder for single-channel RIRs, incorporate pose embeddings, aggregate information across observations, retain the floorplan branch, and convert its output to the local wall representation used by our evaluator.

    The baseline retains the residual acoustic encoding, multi-aggregation principle, floorplan decoding, and MSE plus Dice supervision from EchoScan. The changes to its input representation, aggregation domain, prediction target, and evaluation conversion are documented in Appendix E.2. Because these adaptations materially change the task, we describe the baseline as an EchoScan-style adaptation and not an exact execution of the official implementation on our data.

    We have corrected and added missing information in the revised manuscript.

  1. About monaural versus binaural sensing.

    We thank the reviewer for highlighting the ambiguity in the sensing description. Our reconstruction and planning experiments use omnidirectional, single-channel RIRs. The left channel of separately binaurally rendered RIRs mentioned in Appendix A is used exclusively for supplementary acoustic statistics and is not provided to the reconstruction model or planner.

    The reconstruction and planning experiments therefore use no binaural head-shadow effects or HRTF-based directional filtering. We have revised Appendix A to distinguish the omnidirectional model inputs from the binaural recordings used for statistical characterization and document their respective rendering configurations.

  2. About comparison with MapLearn.

    MapLearn is an acoustic floorplan reconstruction approach that uses propagation-aware learning to infer geometry from recordings collected along known trajectories.

    Our contribution focuses on active sampling, future acoustic representation prediction and its integration into closed-loop sensing decisions through imagined occupancy maps, collision-aware trajectory selection, and information-seeking exploration.

    We have clarified our positioning away from MapLearn in the revised version.

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

In this paper, the authors propose a world model for indoor geometry reconstruction and prediction using acoustic sensor instead of RGB or LiDAR sensor. The whole pipeline is combined of few modules: an acoustic encoder to extract acoustic feature, a pose encoder to encode sensor position, a casual transformer to integrate acoustic feature and motions, occupancy decoder to decode local geometry, a future latent predictor to predict feature acoustic feature, a backend to build global geometry, and a planner to plan trajectories. In their experiments, they build a synthetic dataset including 3251 scenes for evaluation and validation. Comparing with other acoustic baseline, their method's planner reduce the mean collision rates and preserve comparable reconstruction accuracy.

Critical strengths
  1. It's a novelty direction to using world model for acoustic reconstruction and prediction. It's reasonable contribution for the conditions that camera and LiDAR is not working well such as under water. It have potential practical application in both research and industry to the related area.

  2. In their experiments, they show a large improvement on reducing collision rate compare to the baselines. Also, the synthetic benchmark is meaningful contribution for other future work.

Critical weaknesses
  1. The proposed method is only validated on synthetic dataset which is limited to prove their performance. More comparison experiments are need on diverse real world dataset.

  2. The experiment result is not strong enough. Even it show large improvement on collision rate, but not significant performance on reconstruction coverage. Also, there is no enough ablation study to validate the contribution of each module individually.

Questions for authors
  1. While the experiments in the paper show reduced collision rate, but also show lower reconstruction coverage. Could author have deeper analyze on this tradeoff?
  2. How sensitive is the proposed method to different level of inaccurate poses?
  3. How is the running time and GPU memory consumption of the the proposed method?
4 comments

We thank the reviewer for the helpful evaluation of our work. We wish to address the concerns through additional experiments and several clarifications.

  1. About synthetic-only validation.

    We acknowledge that simulation-only evaluation limits conclusions about real-world deployment. Nonetheless, under the challenges in real-world large-scale testing of acoustic mapping, we argue our experiments are a reasonably reliable proxy to results in a real scene – our benchmark is grounded in real-world architectural geometry. Matterport3D and AtlantaLayout provide room layouts derived from scans and panoramic observations of actual indoor environments, ResPlan provides residential floor plans derived from real-estate listings. We convert these layouts into 3D acoustic scenes and simulate sound propagation using SoundSpaces 2.0, a physics-based acoustic simulator previously evaluated against measured room impulse responses in real environments. To strengthen this claim we also include several noise experiments discussed in our rebuttal.

    Our benchmark contains ~ 3,200 scenes with diverse geometric configurations, with active planning evaluated on 287 held-out ResPlan scenes. This provides a controlled setting for comparing planning strategies across varied architectural layouts. Our acoustic-noise and pose-uncertainty experiments further characterize sensitivity to imperfect observations.

    We acknowledge simulations use simplified room geometry and homogeneous acoustic material. Real-world validation across acoustic materials, clutter, hardware effects, and environmental conditions is deferred to future work, mostly because of the difficulty in acquiring such diverse data at scale.

  2. About safety-coverage trade-off.

    We extend the analysis with per-episode collision statistics, collision-free episode rates, traveled distance, and safe high-coverage success, defined as the percentage of episodes achieving at least 90% coverage without collisions.

    PlannerCoverage (%) ↑Collisions/100m ↓Collision-free (%) ↑Safe@90 (%) ↑Distance (m)
    CEM (No-WM)95.943.88663.3856.4149.69
    CEM (Hybrid)96.031.19879.7270.7750.42
    CEM (WM)93.650.77982.5463.1749.15

    Full WM reduces the mean collision rate by approximately 80% relative to No-WM and increases the collision-free episode rate by 19.2 %. Coverage decreases a little, while safe high-coverage success noticeably improves. The similar travel distances indicate that the lower collision rate is not explained simply by shorter trajectories.

    Hybrid retains No-WM's frontier-proximity objective and adds the predictive collision checks used by full WM. It achieves a 69% reduction in mean collisions with comparable coverage and travel distance. Its safe high-coverage success increases by 14.4 %. These results isolate the contribution of predictive collision checking under the same scoring objective.

    Full WM achieves the lowest mean collision rate, while Hybrid provides a favorable balance between safety and coverage. Reconstruction accuracy remains broadly comparable across the CEM variants.

  3. About additional module ablations.

    We also evaluate a geometry-only model trained with λpred=0\lambda_{\mathrm{pred}}=0, retaining the same architecture and 300-epoch training budget, and a latent-persistence variant that replaces predicted future acoustic latents with the latest observed latent. Together with Hybrid, these controls distinguish predictive supervision, future acoustic prediction, and predictive collision evaluation. They complement our existing context-length, rollout-horizon, and latent-representation analyses.

  4. About sensitivity to pose uncertainty.

    We evaluate local geometric reconstruction under independent Gaussian position noise and accumulating odometry drift at inference time, without retraining. Drift is parameterized by σ10m\sigma_{10m}, the per-coordinate standard deviation of accumulated position error after 10 m of travel.

    Noise std. (m)Independent Chamfer ↓Independent [email protected] ↑Drift Chamfer ↓Drift [email protected] ↑
    Clean0.1460.9580.1460.958
    0.050.1520.9510.1470.957
    0.100.1650.9300.1500.953
    0.200.2180.8480.1610.935
    0.300.2970.7500.1790.905
    0.400.3720.6670.2000.869
    0.500.4410.5960.2250.831

    Reconstruction degrades gradually with increasing localization error. Performance remains close to the clean baseline under small perturbations, with larger degradation at higher noise levels. These experiments characterize local reconstruction sensitivity.

We split our answer due to comment length constraints: 5. About runtime and GPU memory.

We profile mapper inference, complete planning latency, and peak PyTorch GPU memory allocation under the same hardware and profiling protocol.

PlannerMapper (ms)Planning (ms)Total (ms)P95 total (ms)Peak GPU (GiB)
CEM (No-WM)13.34386.82400.79473.850.17
CEM (Persistent)12.723163.593176.903306.550.71
CEM (Hybrid)11.782908.472920.822984.570.71
CEM (WM)11.823175.753188.153338.720.71

Mapper inference takes approximately 12-13 ms across CEM variants. Predictive planning increases total computation time by approximately eightfold relative to No-WM, while peak PyTorch GPU allocation remains below 1 GiB. We acknowledge this computational overhead and have uploaded a revised manuscript with clearer discussion of this limitation – we thank the reviewer for raising this matter.

Dear reviewer jNmT, Thank you once again for the time you have invested in evaluating our work. We hope our rebuttal has successfully addressed the points you raised. As the discussion window draws to a close, we would be very grateful for your feedback. Should any outstanding questions remain, we would be very happy to discuss If you find that our responses have resolved your concerns, we kindly ask that you consider adjusting your score to reflect this

Thank you for the detailed rebuttal and additional experiments. They rebuttal addresses some of my concerns, but:

  1. Still lacking real environment experiment results to support the proposed method.
  2. The author provide the running time result, but long latency raised another concern about real-time deployment.

Overall, I will keep my rating as 2(weak rejection).

Reviewer_mJJmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies active acoustic mapping from single-channel room impulse responses and known sensor poses. It proposes a world model that reconstructs local occupancy from acoustic histories and predicts future acoustic latents for closed-loop planning. On a synthetic benchmark, the full planner reduces the reported mean collision rate by approximately 80% relative to a matched CEM control, while maintaining broadly comparable reconstruction quality at slightly reduced coverage.

Critical strengths

S1. The problem is well-motivated: observations collected through motion can help resolve ambiguities in single-channel acoustic mapping. Combining acoustic prediction with closed-loop planning is a promising direction. S2. The results suggest a meaningful safety benefit while largely preserving reconstruction quality relative to the matched CEM control.

Critical weaknesses

W1. The reconstruction experiments do not isolate the contribution of predictive training. Table 1 shows strong performance for the complete mapper, but the EchoScan-style baseline differs in architecture, output representation, and supervision: it predicts a global filled-floor mask, whereas the proposed model predicts local wall occupancy with different losses. Despite matched acoustic histories, poses, and a common evaluation frame, this comparison cannot determine whether future acoustic-latent prediction improves reconstruction beyond the pose-conditioned architecture and local geometry supervision alone. A geometry-only training ablation of the proposed mapper is missing.

W2. The planning experiments do not disentangle information seeking from predictive collision avoidance. As described in Appendix G.4, moving from CEM (No-WM) to CEM (WM) changes both the trajectory-scoring objective and the predictive collision checks. Table 2 supports improved safety with broadly comparable reconstruction: Chamfer distance is nearly unchanged (0.254 versus 0.253 m), F1 improves modestly, and DF-MAE worsens slightly. However, it remains unclear how much of the benefit comes from the information-seeking objective versus the collision checks. Frontier+A* (WM) provides useful evidence for predictive safety checks, but its different planning mechanism leaves the analogous comparison within CEM unresolved.

Questions for authors
  1. How does the proposed mapper perform on Table 1 when retrained with lambda_pred = 0, keeping the acoustic encoder, pose-conditioned temporal model, occupancy decoder, geometric supervision, data, and training budget fixed?

  2. How does the full planner compare with a CEM variant that retains frontier-proximity scoring but uses the same predictive collision evaluation? Please explain how collision-penalty weights are calibrated and how the boundary-discovery bonus is controlled.

3 comments

We thank the reviewer for the constructive feedback. We wish to address concerns raised.

  1. About contribution of predictive training.

    The EchoScan-style comparison involves architectural and supervision differences that prevent it from isolating the contribution of predictive training. We therefore trained a geometry-only variant of our mapper with λpred=0\lambda_{\mathrm{pred}}=0, retaining the same acoustic encoder, pose-conditioned temporal model, occupancy decoder, geometric losses, dataset, and training budget of 300 epochs. Results are in the table below.

    ModelChamfer (m) ↓DF-MAE (m) ↓[email protected] ↑[email protected] ↑[email protected] ↑
    Full model0.1460.0960.8160.9580.976
    λpred=0\lambda_{\mathrm{pred}}=00.1500.0990.8100.9530.972

    Predictive supervision provides a modest improvement across the reconstruction metrics – so its motivation is enabling imagined future acoustic observations for active planning.

  2. About information seeking vs. predictive collision avoidance.

    We add a variant CEM (Hybrid), that retains the frontier-proximity scoring objective of CEM (No-WM) while applying the same world-model-based predictive collision checks used by full CEM (WM). The variants share other components (e.g. learned mapper, occupancy fusion backend, CEM sampling and optimization budget, current-map feasibility checks, frontier-based fallback mechanism) - so the hybrid variant isolates predictive collision evaluation under the same frontier-scoring objective.

    We also evaluate another variant CEM (Persistent), which replaces predicted future acoustic latents with the most recent observed latent while retaining future pose conditioning, the Transformer, occupancy decoder, planning objective, and predictive collision checks.

    PlannerCoverage (%) ↑Collisions/100m ↓
    CEM (No-WM)95.9413.886
    CEM (Hybrid)96.0251.198
    CEM (Persistent)95.2180.920
    CEM (WM)93.6460.779

    Hybrid reduces mean collisions by almost 70% relative to No-WM while maintaining similar coverage. Full WM achieves an 80% reduction relative to No-WM, accompanied by a modest decrease in coverage. We draw that predictive collision evaluation accounts for a substantial portion of the safety improvement. The information-gain-based objective achieves a further reduction in mean collisions with the same predictive collision checks.

    The persistence control also shows that pose-conditioned occupancy inference with repeated acoustic features retains much of the planning benefit. Full autoregressive prediction achieves the lowest observed mean collision rate.

  3. About planning objective and collision handling.

    The planner separates trajectory feasibility from information-gain optimization. Candidate trajectories crossing cells with occupancy probability above a threshold (we set it to 0.85) are rejected, including checks against imagined future occupancy. Such candidates receive a score of −∞-\infty, implementing a hard feasibility constraint without a finite collision-penalty weight. Hybrid and full WM use the same predictive collision-screening procedure.

    When an executed action collides with a wall in the simulator, the contact location is incorporated into the global occupancy map as certain occupied evidence. This observation is treated without modeled measurement uncertainty.

    For feasible trajectories, full CEM (WM) maximizes predicted map-space entropy reduction, with a twofold weighting for predicted wall cells whose occupancy probability exceeds 0.8. This favors information acquisition near predicted occupied structure. Hybrid retains the negative geodesic distance to the nearest frontier used by No-WM. The thresholds and weights are always fixed across evaluations.

    The collision threshold and information-gain weighting were selected through standard hyperparameter tuning on a held-out validation set, using the safety-coverage trade-off to guide selection. The selected parameters were fixed for all test evaluations. We have added clarificaitons in the revised version.

Dear reviewer mJJm, thank you again for your time and effort.

We hope that our rebuttal has addressed your concerns. With the discussion period nearing its end, we would appreciate hearing your thoughts.

We would be happy to address any remaining concerns or provide further clarification. If our responses have resolved your concerns, we would greatly appreciate it if you could reconsider your score.

Thank you for the additional experiments and clarifications. The new controls address my original ablation requests. However, the results leave me unconvinced about the additional contribution of future acoustic-latent prediction and information-gain scoring. Predictive supervision provides only a modest reconstruction improvement, Persistent retains much of the safety benefit, and full WM achieves lower mean collisions at the cost of reduced coverage and lower Safe@90 than Hybrid.

While I acknowledge the value of predictive collision checking, the demonstrated additional benefits are not strong enough to change my overall assessment.

Comments on the paper

1 comment

We thank all reviewers for their thorough and helpful evaluation and constructive feedback.

We have uploaded a revised version of the manuscript incorporating the corrections and additions discussed in our rebuttal. The main changes include a matched ablation isolating the contribution of predictive supervision, additional controlled planning comparisons examining the effects of predictive collision checking and future acoustic prediction, and further safety and coverage metrics. We also added robustness evaluations for simulated acoustic acquisition and pose uncertainty, analyses across scene characteristics, and computational cost measurements. Finally, we clarified the planning objective, dataset and simulation assumptions, and baseline adaptations, and reorganized the appendix to improve clarity and reproducibility.

We hope these revisions and our rebuttal in the comments address the major concerns raised during the review. If the reviewers deem their main concerns addressed, we'd greatly appreciate consideration of a score change in light of the additional evidence and clarifications. In the case any major concerns remain underadressed, we're happy to discuss further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.