acceptodds
Under review as a conference paper at ICLR 2027

D2AVN: Efficient Direction-Decoupled Audio-Visual Navigation in Multi-Source and Reverberant Soundscapes

Abstract

Audio-Visual Navigation (AVN) empowers embodied agents to leverage the target-specific acoustic cues to complement visual perception, enabling them to locate sound sources that are visually occluded or outside their immediate field of view. Previous AVN methods predominantly derive directional guidance by estimating the geometric source direction—the direct spatial vector from the agent to the target. However, this paradigm struggles in authentic indoor environments, where multi-source interference obscures target cues and reverberation causes the perceived acoustic Direction of Arrival (DOA) to diverge significantly from the direct source path. To this end, we propose D2AVN, a novel direction-decoupled framework that explicitly isolates target acoustic arrival evidence before translating it into actionable navigation guidance. Importantly, our D2AVN employs a novel target-conditioned DOA (TargetDOA) module that extracts target-specific arrival distributions directly from complex acoustic mixtures, utilizing paired target-only acoustic references that exclude competing sources while preserving environmental propagation. Then, our Depth-conditioned Reward-Aligned Egocentric Affordance Modeling (DREAM) module reshapes these arrival distributions into reward-aligned navigation direction preferences by incorporating local geometry and egomotion-aligned history. To facilitate robust policy learning in these complex spaces, we implement a training pipeline with effective online acoustic rendering and reinforcement learning. Extensive validation demonstrates that D2AVN achieves a 33.0% success rate with First-Order Ambisonics (FOA) and 15.8% with binaural audio. Notably, the binaural configuration improves Success weighted by Path Length (SPL) and action efficiency (SNA) by 74% and 90%, respectively, over the fine-tuned ENMuS³ baseline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Audio-Visual Navigation, Embodied AI, Multimodal Learning, Spatial Audio, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_PSxtOpenReview ↗
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:

  1. D2AVN framework: A novel direction-decoupled framework that first isolates target acoustic arrival evidence from acoustic mixtures before translating it into actionable navigation guidance.

  2. 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.

  3. 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.

  4. 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
  1. 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.

  2. The paper is clearly written, with the methodology presented in a well-articulated manner.

Critical weaknesses
  1. 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 MAEgeoMAE_{geo} but loses on MAEacoMAE_{aco}, 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.

  2. 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
  1. 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?

  2. In Figure 3, D2AVN fails in the second example (SPL 0.000); what is the selection criterion? Is it representative?

  3. 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?

  4. 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
AuthorsOpenReview ↗

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:

GT angular gapNumber of positionsProportion
<15°2,56122.65%
15–30°1,70815.11%
30–45°1,38512.25%
45–90°3,15827.93%
>90°2,49522.07%

The gap remains substantial at target-reachable positions:

Receiver subsetValid nodesGT gap: mean ± SDGap >45°Gap >90°
Target-reachable4,35261.60 ± 50.49°53.65%27.37%
Target-reachable, graph distance 6–50 m3,22968.87 ± 50.71°60.51%32.30%

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.

SplitSamples (scenes)Proportion with gap ≥45°Acoustic MAE (°) ↓Geometric MAE (°) ↓
Train3,763 (43)51.9%6.99 ± 0.6233.24 ± 1.72
Dev762 (8)52.5%9.99 ± 1.9455.25 ± 4.73
Test481 (8)53.8%11.25 ± 0.4953.21 ± 0.33

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:

GT angular gapSamplesAcoustic-supervised MAE (°) ↓Geometric-supervised MAE (°) ↓
≥45°25914.24 ± 0.4972.10 ± 2.47
≥90°12514.46 ± 1.3079.35 ± 7.66

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.

AuthorsOpenReview ↗

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 qt=ptq_t=p_t 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:

VariantFrontend supervisionDREAMSR (%) ↑SPL ↑SNA ↑DTG (m) ↓
FullAcousticRetained29.60.15070.092511.104
DirectAcousticBypassed23.20.13010.075611.76
Geometric FullGeometricRetained5.00.01700.010311.82

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 methodSR (%) ↑SPL ↑SNA ↑
ENMuS³, adapted12.40.06550.0361
MAGNet, adapted14.80.11320.0596
D2AVN15.80.11420.0685

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.

AuthorsOpenReview ↗

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:

StatisticD2AVN–PhysicalDOAENMuS³
Successful episodes200191
Failed-episode DTG: mean8.582 m7.102 m
Failed-episode DTG: median3 m4 m
Failed-episode DTG: 90th percentile24 m17 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.

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

This paper proposes D2AVN, an audio-visual navigation framework designed for multi-source and reverberant indoor environments. The key idea is to separate acoustic sound arrival direction from the actual navigation direction, since sound reflections and obstacles can make these directions different. The method introduces two main components: TargetDOA, which extracts target-specific acoustic direction information from mixed sounds using target-only acoustic supervision, and DREAM, which combines this information with depth geometry and navigation history to learn more effective navigation directions. The authors also develop a multi-source acoustic simulation environment using SoundSpaces 2.0 and train the navigation policy with reinforcement learning. Experiments compare D2AVN against existing navigation methods, particularly ENMuS3, using success rate, path efficiency, and action efficiency. Results show improved navigation performance, especially with FOA audio, while component analyses demonstrate the contributions of target conditioning, spatial acoustic cues, and navigation direction adaptation.

Critical strengths
  • I think the idea of separating acoustic arrival direction from navigation direction is interesting and makes sense, especially in environments with multiple sound sources, reverberation, and obstacles. The proposed TargetDOA and DREAM modules are designed to address these challenges, providing a meaningful approach to improving navigation in complex environments.

  • The authors evaluate the method in both single-source and multi-source environments and show improvements over ENMuS3, particularly in path and action efficiency. I also appreciate the component analyses, which examine the effects of target conditioning, spatial audio cues, and navigation history. These experiments help support the proposed design choices and make the overall contribution more convincing, although further comparisons would strengthen the results.

Critical weaknesses
  • The comparison with recent audio-visual navigation methods could be more comprehensive. While the paper includes ENMuS3 as a strong baseline, the main multi-source navigation experiment only compares against this method. Other recent approaches, such as SAVN-CE, RAVN, and BDATP, are closely related to the proposed spatial reasoning and navigation framework. Including some of these methods under the same evaluation setting would help establish how D2AVN compares with more recent developments in the field.

  • The current DREAM ablations are performed by disabling components of an already trained model at inference time. While these experiments provide useful insights, they do not fully establish whether the proposed architecture is better than simpler alternatives or the degradation is simply due to feature distribution shift. Retraining a model without DREAM under the same training budget would make the comparison more convincing.

Questions for authors

Please see weakness.

1 comment
AuthorsOpenReview ↗

Response to Reviewer cXtA

We thank the reviewer for recognizing the motivation for separating acoustic arrival evidence from navigation guidance and for the suggestions on recent baselines and retrained ablations. We have completed additional comparisons addressing both concerns.

W1. Comparison with recent audio-visual navigation methods.

We agree that comparisons beyond ENMuS³ would strengthen the evaluation. We have adapted MAGNet from the official SAVN-CE repository, starting from its distractor checkpoint. The results below use binaural audio and the same soundscape and Fixed500 episodes as our existing comparison.

MethodSR (%) ↑SPL ↑SNA ↑
ENMuS³, adapted12.40.06550.0361
MAGNet (SAVN-CE)14.80.11320.0596
D2AVN, binaural15.80.11420.0685

D2AVN achieves higher SR, SPL, and SNA than the adapted MAGNet checkpoint. MAGNet uses categorical goal input and a 1-s audio window. These results evaluate task adaptation on the reused Fixed500 development suite.

At the time of this response, we could not locate an available implementation of RAVN. For BDATP, we do not yet have results under the same protocol and will acknowledge this limitation in baseline coverage. We will include the completed MAGNet comparison and its adaptation details in the revision.

W2. Retraining without DREAM under the same budget.

We agree that inference-time removal alone cannot isolate DREAM’s benefit from a mismatch with the trained policy. We therefore trained Full and Direct independently from scratch, using the same frozen Target-DOA frontend, soundscape, initialization seed, fresh Adam optimizers, and PPO settings. Direct uses qt=ptq_t=p_t throughout training and inference, bypassing DREAM’s adaptation and history while retaining the downstream policy GRU. Its directional interface is thus consistent between training and evaluation.

VariantSR (%) ↑SPL ↑SNA ↑DTG (m) ↓
Full29.60.15070.092511.104
Direct23.20.13010.075611.76

The updated Full result differs from the original report because Target-DOA was retrained with all evaluation scenes excluded from its training data.

Full outperforms the independently retrained Direct policy across SR, SPL, and SNA, improving SR by 6.4 percentage points. Because Direct is trained with the same bypassed interface used at evaluation, this comparison provides evidence that DREAM’s benefit is not solely an artifact of inference-time distribution shift, complementing the original ablations in Table 4.

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

This paper proposes a method for target-conditioned av navigation in multi-source, reverberant indoor simulations. It implements a TargetDOA module which predicts the target-specific acoustic arrival distribution from mixed audio and text query. And the training references computed from the targets' isolated rendered audio. It proposes another module called DREAM which combines this evidence with depth and pose-aligned history to produce a navigation preference learned through the actor's RL objective. It then conditions RGB-D fusion and a recurrent PPO policy with TargetDOA module remains frozen. Evaluation includes Matterport3D multi-source navigation, SAVI-derived single-source transfer, standalone acoustic controls, and DREAM inference interventions. It report on the custom 500 episodes with success rates 33% with FOA and 15.8% with binaural audio. The modality-matched binaural comparison improves SPL and SNA by 74% and 98%.

Critical strengths

The authors propose well motivated supervision of target acoustic arrival. The proposed TargetDOA module utilize target-only reverberant signal instead of direct arrival to capture environmental propagation. This is a sensible intermediate representation and provides explicit target-specific evidence for downstream navigation. The DREAM module then adapts the arrival of distribution with current depth and pose-aligned history, incorporating context for obstacles and temporal continuity. Together, these two modules provide a coherent and interpretable interface between acoustic perception and navigation

The authors make meaningful component controls and comparative navigation results. The binaural comparison matches the audio modality and reports improvements over ENMus, while the results in table 1 provides complimentary evidence of efficiency gains.

Critical weaknesses

Appendix A.3 states that its scenes participate in binaural frontend selection, which brings uncertainty if Tables 2 and 4 report performance on a development suite?

Table 4 changes one full-model checkpoint at inference without retraining? It brings concern that bypassing DREAM could create a mismatch with the policy's training recipe.

Questions for authors

In Appendix D.1, author states that "fixed 500" was repeatedly used for development and checkpoint selection. Is the "fixed 500" the same as the held out sessions for testing?

Are there multi-source results on scenes and episodes excluded from frontend selection, navigation checkpoint selection, and iterative development?

How does navigation performance change when TargetDOA uses geometric rather than acoustic supervision, and DREAM is bypassed and the downstream policy is retrained?

2 comments
AuthorsOpenReview ↗

Response to Reviewer iQGU — Part 1

Thank you for your careful review and for recognizing the motivation of separating target-specific acoustic evidence from navigation adaptation. We agree that the evaluation split and the distinction between inference interventions and independently trained controls need to be stated more clearly.

W1 and Q1–Q2. Evaluation split and independent testing.

Yes. Fixed500 in Tables 2 and 4 was repeatedly used for development and checkpoint selection and should be described as a development evaluation, not an untouched test set. Our reconstruction of the historical V9 frontend split indicates that eight Fixed500 scenes were included in frontend training. The navigation policy’s separate training/evaluation scene split therefore did not establish scene independence for the complete system. We will correct the split description and qualify the associated generalization claims.

Regarding Q2, we have retrained Target-DOA with all validation scenes excluded from its training data and retrained the navigation models using this updated frontend. Navigation training and validation scenes are also strictly disjoint. The updated results therefore measure transfer to scenes excluded from both frontend and policy training. We believe this protocol provides meaningful evidence of generalization across scenes for these comparisons. We will identify these as validation results and clarify Fixed500’s role in checkpoint selection and iterative development in the revised paper.

W2 and Q3.(a). Independent retraining without DREAM.

Table 4 reports inference-time interventions on a trained Full model. We agree that these alone cannot separate DREAM’s benefit from a mismatch with the trained policy.

We therefore trained Full and Direct independently from scratch, using the same frozen Target-DOA frontend, soundscape, initialization seed, fresh Adam optimizers, and PPO settings. Direct uses qt=ptq_t=p_t throughout training and inference, bypassing DREAM’s adaptation and history while retaining the downstream policy GRU. Its directional interface is thus consistent between training and evaluation.

The following results use Fixed500 and one PPO training seed:

Independently trained configurationSR (%) ↑SPL ↑SNA ↑DTG (m) ↓
Full29.60.15070.092511.104
Direct: DREAM bypassed throughout training23.20.13010.075611.76

The Full result differs from the original report because it uses the retrained Target-DOA frontend described above.

Full exceeds Direct by 6.4 percentage points in SR, with higher SPL and SNA. Since Direct was independently trained with DREAM bypassed, this comparison provides evidence that DREAM’s benefit is not solely an artifact of inference-time distribution shift, complementing the original Table 4 ablations.

AuthorsOpenReview ↗

Response to Reviewer iQGU — Part 2

Q3.(b). Geometric supervision and independently retrained controls.

We conducted separate retraining controls for geometric supervision, supported by the following reference analysis and frontend experiments.

Acoustic–geometric reference differences. Our census covers 11,307 valid graph nodes, without selecting nodes by angular gap. The paired acoustic and geometric references have the following distribution:

GT angular gapNumber of positionsProportion
<15°2,56122.65%
15–30°1,70815.11%
30–45°1,38512.25%
45–90°3,15827.93%
>90°2,49522.07%

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.

SplitSamples (scenes)Proportion with gap ≥45°Acoustic MAE (°) ↓Geometric MAE (°) ↓
Train3,763 (43)51.9%6.99 ± 0.6233.24 ± 1.72
Dev762 (8)52.5%9.99 ± 1.9455.25 ± 4.73
Test481 (8)53.8%11.25 ± 0.4953.21 ± 0.33

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:

GT angular gapSamplesAcoustic-supervised MAE (°) ↓Geometric-supervised MAE (°) ↓
≥45°25914.24 ± 0.4972.10 ± 2.47
≥90°12514.46 ± 1.3079.35 ± 7.66

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.

Independently retrained navigation. We independently retrained the downstream policies for geometric supervision with DREAM retained. Results from the currently evaluated checkpoints on Fixed500 are:

VariantFrontend TypeDREAMSR (%) ↑SPL ↑SNA ↑DTG (m) ↓
FullAcousticRetained29.60.15070.092511.104
Geometric FullGeometricRetained5.00.01700.010311.820

The geometric Full model records 5.0% success rate. The frontend and navigation results together show that, within our framework of explicitly modeling a directional prior for navigation, geometric supervision did not yield an effective Target-DOA frontend under the evaluated reverberant conditions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.