- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
Summary:
This paper proposes the D2AVN framework, decomposing audio-visual navigation into two stages. TargetDOA extracts target-specific acoustic arrival evidence from complex acoustic mixtures conditioned on a target query, supervised by target-only acoustic references that preserve environmental propagation. DREAM then adapts this arrival evidence into a reward-aligned navigation direction preference using local geometry and egomotion-aligned history. Experimental results demonstrate the effectiveness of the proposed method.
Contributions:
-
D2AVN framework: A novel direction-decoupled framework that first isolates target acoustic arrival evidence from acoustic mixtures before translating it into actionable navigation guidance.
-
TargetDOA: Estimates target-specific arrival distributions from acoustic mixtures conditioned on a target query, supervised by paired target-only DOA that excludes competing sources while preserving environmental propagation.
-
DREAM: Transforms these distributions into reward-aligned navigation direction preferences using local depth geometry and egomotion-aligned history, with navigation rewards guiding this transformation through the downstream actor-side objective.
-
System integration and policy optimization: The resulting navigation preferences condition RGB-D fusion for recurrent action prediction, with the policy optimized using PPO while TargetDOA remains frozen.
Critical strengths
-
The supervision design is ingenious. Using "target-only components from the same scene, same pose, and same window length" as supervision both excludes competing sources and preserves propagation effects; the paired dual-query setup forces the model to perform query-dependent discrimination rather than fitting a single pathway.
-
The paper is clearly written, with the methodology presented in a well-articulated manner.
Critical weaknesses
-
The paper's central claim that acoustic arrival supervision outperforms geometric source-direction supervision and yields better navigation is not validated by controlled experiments. In Table 3, all variants are evaluated against the same acoustic reference; the geometric supervision variant wins on but loses on , which is largely an artifact of label definition and demonstrates only label fit, not navigation utility. There is neither a geometrically supervised TargetDOA integrated into the same navigation policy nor an oracle-direction upper bound. Moreover, the cross-label gap between the two models is about 12°, within a single 15° bin, suggesting that the acoustic reference and geometric bearing are highly correlated rather than "diverging significantly." Physically, a window-averaged active intensity in an ideal diffuse field has zero net diffuse component, so it primarily reflects the direct path and early reflections, not the strongly reverberant arrival the paper assumes. In my opinion, this paper should add: (a) a geometric-supervision variant trained under the same navigation objective, (b) an oracle-direction policy bound, and (c) results stratified by the aco-geo angular gap, together with the gap's distribution (mean, variance, multimodality) across occlusion, RT60, and direct-path presence; otherwise, "arrival is not source" remains unsupported.
-
The FOA versus binaural comparison is confounded by training configuration, and the FOA headline result lacks a same-modality baseline. Table 7 shows that the two modalities are trained under substantially different settings: FOA continues from an existing navigation checkpoint whereas binaural is trained from scratch; the learning rates differ markedly across modules; the action-entropy coefficients follow different schedules; and FOA replaces its frozen frontend partway through training while binaural trains it from scratch. The large success-rate gap therefore cannot be attributed to input modality. Moreover, the FOA row in Table 2 has no FOA-input baseline (ENMuS³ is binaural-only), yet the abstract leads with the FOA figure, and Table 1 does not state whether D2AVN uses FOA or binaural. The authors should either rerun both modalities with matched initialization, learning rates, and entropy schedule, or explicitly downgrade the FOA result to a "configuration upper bound" and restrict the main comparison to binaural.
Questions for authors
-
In Table 1, D2AVN's DTG of 6.86 is worse than ENMuS³'s 5.74, yet its SR/SPL is higher; what is the distance distribution of the failed episodes?
-
In Figure 3, D2AVN fails in the second example (SPL 0.000); what is the selection criterion? Is it representative?
-
Will the code for this method be open-sourced? Additionally, do the paper's claims about generality and generalization exceed what its simulated, closed-set, single-configuration evaluation can support?
-
Why are all the citations in the manuscript rendered in pink font? Do the authors modify the official ICLR submission template to use this as a special marker?
5 comments
Response to Reviewer PSxt — Part 2
W1 (continued). Independently retrained navigation controls.
Independently retrained navigation controls.
We independently retrained the navigation policy with a geometrically supervised frontend and DREAM retained, and trained an acoustic Direct control from scratch. Full and Direct share the same frozen acoustic frontend and PPO settings. Direct uses throughout training and inference, bypassing DREAM’s adaptation and history while retaining the downstream policy GRU.
We retrained Target-DOA with all validation scenes excluded from its training data, then retrained navigation using the updated frontend. This explains the change from the originally reported Full result. Current checkpoint results on Fixed500 are:
| Variant | Frontend supervision | DREAM | SR (%) ↑ | SPL ↑ | SNA ↑ | DTG (m) ↓ |
|---|---|---|---|---|---|---|
| Full | Acoustic | Retained | 29.6 | 0.1507 | 0.0925 | 11.104 |
| Direct | Acoustic | Bypassed | 23.2 | 0.1301 | 0.0756 | 11.76 |
| Geometric Full | Geometric | Retained | 5.0 | 0.0170 | 0.0103 | 11.82 |
Geometric Full records 5.0% success rate. The frontend and navigation results together show that geometric supervision did not yield an effective Target-DOA frontend under the evaluated reverberant conditions within our framework, which explicitly models a directional prior for navigation.
Full improves SR by 6.4 percentage points over Direct, with higher SPL and SNA. Since Direct is independently trained with DREAM bypassed, this provides evidence that DREAM’s benefit is not solely due to inference-time distribution shift. The Full–Direct comparison uses one PPO training seed; the three-seed statistics concern the frontend diagnostic.
W2. Choice of FOA and modality-matched evaluation.
We select FOA for its explicit spatial representation: its omnidirectional and three directional components provide a structured input for target-direction estimation and navigation adaptation. This makes FOA a suitable sensing option for our framework, which explicitly models directional evidence for navigation.
To enable modality-matched comparisons with existing binaural RL baselines, we additionally trained a binaural D2AVN variant and evaluated all methods under the same soundscape and Fixed500 protocol:
| Binaural method | SR (%) ↑ | SPL ↑ | SNA ↑ |
|---|---|---|---|
| ENMuS³, adapted | 12.4 | 0.0655 | 0.0361 |
| MAGNet, adapted | 14.8 | 0.1132 | 0.0596 |
| D2AVN | 15.8 | 0.1142 | 0.0685 |
D2AVN achieves higher SR, SPL, and SNA than both adapted baselines, supporting the effectiveness of our perception–adaptation framework with binaural input. MAGNet uses the official SAVN-CE distractor checkpoint, categorical goal input, and a 1-s audio window. These are task-adaptation comparisons with different total training budgets.
The binaural experiments provide our main external comparison. We report the FOA configuration separately and use the independently retrained Direct model as an internal control for DREAM under the same FOA input. The original FOA and binaural runs used different training configurations, so their performance gap does not isolate a modality effect. We will clarify the distinct purposes of these comparisons in the revision.
Response to Reviewer PSxt — Part 3
Q1. Why is DTG worse despite higher SR and SPL?
The failure-distance distribution reveals a heavier long-distance tail for D2AVN–PhysicalDOA. Across the same 994 episodes:
| Statistic | D2AVN–PhysicalDOA | ENMuS³ |
|---|---|---|
| Successful episodes | 200 | 191 |
| Failed-episode DTG: mean | 8.582 m | 7.102 m |
| Failed-episode DTG: median | 3 m | 4 m |
| Failed-episode DTG: 90th percentile | 24 m | 17 m |
Our failures have a lower median DTG but a heavier distant tail, which raises the mean. This is compatible with higher SR and success-weighted path efficiency. Failed episodes receive zero SPL/SNA, so short failed trajectories do not explain the efficiency gains.
Q2. Figure 3’s selection and failed example.
We apologize for the lack of clarity in Figure 3. In the highlighted example, it is the baseline that fails and therefore receives an SPL of 0.000, whereas D2AVN successfully completes the task. The trajectories were selected to illustrate navigation behavior and potential advantages; they are not a random or statistically representative sample. We will clarify the selection criterion and explicitly identify each method’s success or failure in the caption.
Q3. Code availability and scope of generalization.
We will make our code publicly available. We are currently conducting additional experiments to address the reviewers’ valuable feedback.
Q4. Citation color.
Thank you for pointing this out. The pink coloring carries no special meaning; it simply highlights hyperlinks and references for readability. Similar formatting appears in previously accepted ICLR papers. For example, Statistical Advantage of Softmax Attention: Insights from Single-Location Regression (ICLR 2026) uses blue to highlight hyperlinks.
Response to Reviewer PSxt — Part 1
Thank you for your careful review and for recognizing the target-specific supervision design and clarity of our presentation. We have added reference-level analyses, matched frontend experiments, and independently retrained navigation controls.
W1. Acoustic–geometric differences and supervision targets.
The approximately 12° difference between model errors does not measure the angular gap between the two ground-truth references. The relatively small cross-reference MAE gap between differently supervised Target-DOA models was mainly due to the data distribution: acoustic arrival and geometric bearing can be close in some scenes. We therefore compared paired references for the same target, receiver pose, and audio window across 59 scenes in our custom reverberant simulation environment:
The gap remains substantial at target-reachable positions:
The mean gap is also larger in the architecture-blocked subset. These measurements support treating acoustic arrival and geometric bearing as distinct supervision targets.
Physical interpretation. We agree that an ideal isotropic diffuse component has zero expected net active intensity. Our supervision uses a finite-window net acoustic direction computed from target-only rendered audio; directional propagation components, including early reflections, can make this direction differ from the geometric source bearing.
Matched frontend retraining. On a separate diagnostic set enriched for large angular gaps, we retrained acoustic- and geometric-supervised frontends with matched inputs, architecture, initialization, and training budgets. Each model was evaluated against its own ground truth. Both tables report scene-macro MAE as mean ± SD over three seeds.
The acoustic model retains lower own-target error across all splits, with substantially higher geometric error even on the training split.
Performance by acoustic–geometric angular gap. We further grouped localization results on the combined test set—481 samples across eight scenes—by the angular gap between ground-truth acoustic arrival and geometric bearing:
The ≥90° subset is included in the ≥45° group. On samples with substantial acoustic–geometric disagreement, geometric supervision yields much larger own-target errors, while acoustic arrival direction remains accurately estimable under the same model and training budget.