acceptodds
Under review as a conference paper at ICLR 2027

Dynamic Audio-Visual Navigation in Continuous Environments with Moving Targets

Abstract

In real-world scenarios, sound-emitting entities such as people, pets, and mobile robots may move continuously or emit sound intermittently, and their sounds may be mixed with those from irrelevant sound sources. While recent studies have extended audio-visual navigation from discrete, non-semantic settings to continuous, semantic settings, these advances assume stationary targets. We therefore introduce dynamic audio-visual navigation in continuous environments (DANCE), a new task in which an embodied agent uses audio-visual observations to navigate toward a single sound-emitting target that may be moving or stationary. Navigating toward a moving target is challenging because previously observed spatial cues can quickly become outdated as the target moves. To address this challenge, we propose DyGMap, which couples audio-visual goal estimation with dynamic goal mapping. The former integrates complementary acoustic and visual evidence to estimate the target's semantic category and location, while the latter projects these estimates onto category-specific goal maps and decays outdated or spurious target evidence. Extensive experiments in clean and noisy settings demonstrate that DyGMap improves the success rate by approximately 11.6 percentage points over the strongest evaluated baseline, demonstrating the importance of robust goal reasoning and dynamic goal mapping for efficient navigation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
audio-visual navigation, continuous environments, sound event localization and detection, moving target

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.40 over 5 reviews, above 68% of the others.
Reviewer_KdzoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces DANCE, a new audio-visual navigation benchmark in continuous 3D environments where an embodied agent must navigate to a semantically grounded acoustic target that may be either moving or stationary. The paper also proposes DyGMap: a GDN extended with visual cross-attention, plus a dynamic goal map with EMA + temporal decay to suppress stale target evidence.

Critical strengths

The paper proposes a well-motivated task with a well-documented method.

Critical weaknesses

The proposed benchmark includes only three semantic categories, which makes the task considerably less challenging than the open-vocabulary grounding setting commonly adopted in recent work. The authors should expand the category set or justify the restricted scope. The method is largely incremental. The GDN is explicitly "building upon MAGNet", with minor adjustments, and the paper oversells the "dynamic" contribution. The experimental setup does not convincingly isolate why DyGMap wins, and several comparisons are of uncertain fairness. For example, the ablation studies do not provide enough factorial isolation of the claimed axes. In addition, the authors should more clearly document how the baselines were adapted to the task and why they fail.

Questions for authors

We kindly ask that the authors address each of the weaknesses raised in the review above, and clarify how the proposed benchmark and experiments account for the concerns noted therein.

4 comments
AuthorsOpenReview ↗

Response to the Reviewer, Part 1/2

We thank the reviewer for recognizing the motivation behind DANCE. We would like to clarify the scope of the benchmark, the contributions beyond MAGNet, and the basis for our comparisons.

1. Scope of the target categories. Our common response Limited Number of Target Categories in DANCE explains our category choices and addresses open-vocabulary grounding in point 4. DANCE places greater emphasis on grounding the target sound in the visible object that emits it than on recognizing the sound category alone. For example, when the target in DANCE is a dog, a dog is present in the scene and its bark originates from the dog's location. As the dog moves through the scene, both the sound source and the visual asset move with it. The agent can therefore use spatially and semantically aligned audio-visual cues to locate the target. This grounding requirement makes expanding the category set more demanding, since each new category requires both suitable sound recordings and corresponding visual assets.

2. Contributions beyond MAGNet and what the ablations isolate. DyGMap builds on MAGNet's goal descriptor network (GDN) by incorporating vision to estimate the target's category and location. In our revised evaluation along the same oracle trajectories, adding vision improves most target estimation metrics.

The larger change is dynamic goal mapping. DyGMap maps obstacles, explored regions, and the agent's trajectory alongside category-specific goal maps. For each map location, it records both the strength of the target evidence and the time of its last update. This allows the policy to distinguish recently supported target locations from earlier estimates while retaining the scene geometry.

In the noisy-setting ablations (Table 5), the first row provides a MAGNet-style reference. The SR gains from adding the geometric map (rows 2–3), the goal map (rows 3–4), and dynamic updates (rows 4 and 6) are 5.23, 1.69, and 2.31 percentage points, respectively. Rows 5 and 6 isolate the effect of incorporating vision into the GDN on navigation performance, showing an SR gain of 5.27 percentage points. These comparisons identify what each addition contributes in the tested configurations.

3. Contribution of dynamic goal-map updates. We agree that the original submission did not adequately establish how the benefit of dynamic updates varies with target motion. Our common response Dynamic Goal-Map Ablations, Decay-Factor Sensitivity, and the Effect of Target Speed reports results on test sets constructed at seven target-to-agent speed ratios from 0.00 to 1.50 in steps of 0.25. The gains from dynamic goal-map updates vary with both target speed and acoustic conditions. Relative to DyGMap w/o dynamic, the largest SR gain is 3.51 percentage points at 0.25× (57.49% to 61.00%) in the clean setting and 4.24 percentage points at 0.75× (from 37.75% to 41.99%) in the noisy setting.

AuthorsOpenReview ↗

Response to the Reviewer, Part 2/2

4. Baseline adaptations, fairness, and failure modes. Appendix D.2 documents how each baseline is implemented and trained. As detailed in our common response Adaptation of DAV-Nav, we retained DAV-Nav's central design and replaced its grid-based waypoint selector with direct action prediction, the action interface used by all learning-based methods in our evaluation. We consider this a relatively fair comparison under this constraint.

The baselines differ in how they use multimodal observations, especially target-related cues.

(1) AV-Nav, SMT + Audio, and DAV-Nav do not explicitly estimate the target's semantic category and location, while SAVi and MAGNet do not use visual observations to improve these estimates. DyGMap uses complementary audio-visual cues to improve target estimation.

(2) AV-Nav, SMT + Audio, and DAV-Nav do not account for intermittent target sounds. When the audio input is all zeros, SAVi retains its previous category estimate and updates its location estimate using the agent's pose. In noisy settings, however, distractor sounds or background noise can keep the audio input nonzero even when the target is silent. SAVi may therefore continue updating its estimate from those sounds. DyGMap uses a memory-augmented GDN that, as in MAGNet, draws on episodic memory to continue estimating the target's category and location during temporary silence and after the target stops emitting sound entirely.

(3) AV-Nav, SMT + Audio, SAVi, and MAGNet do not use explicit spatial maps, while DAV-Nav lacks an explicit mechanism for updating target-location evidence as the target moves. DyGMap maps obstacles, explored regions, and the agent's trajectory alongside category-specific goal maps. Its dynamic goal-map update mechanism records when the target evidence at each location was last updated and gradually reduces the weight of that evidence if no new estimates support it. This can happen when the target moves away or when a transient prediction error places evidence at an incorrect location. The agent can thus distinguish recently supported target locations from earlier estimates while retaining the scene geometry. These maps also allow the agent to reason about which regions are occupied or have been explored, where targets of each category may be located, and how strong the evidence is at each location.

Our experiments and ablations support these design choices and help explain DyGMap's substantial performance gains over the baselines.

The authors’ response addresses most of my concerns. After considering it, I have decided to keep my original score.

Thank you for carefully considering our responses and for acknowledging that they address most of your concerns. We respect your decision and sincerely appreciate your constructive feedback and suggestions for improving our work.

Reviewer_J97oOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The manuscript introduces DANCE, a benchmark for audio-visual navigation toward semantically grounded targets that may move or remain stationary in continuous environments with intermittent sound emission. It also proposes DyGMap, which combines audio-visual estimation of target category and location with category-specific goal maps, temporal decay, and a goal-update timestamp channel. Experiments evaluate navigation and goal reasoning in clean and noisy settings. The authors report improved navigation performance over the evaluated baselines, with ablations examining visual information and mapping components.

Critical strengths
  1. DANCE addresses a relevant navigation problem by jointly considering target motion, intermittent sound emission, semantic grounding, and continuous environments. This combination gives the benchmark potential value for studying navigation when target evidence changes or becomes unavailable.

  2. The authors report promising navigation results in both clean and noisy settings, with DyGMap outperforming the evaluated baselines. These results support the potential of combining goal estimation with spatial mapping and provide a useful starting point for further study.

Critical weaknesses
  1. Limited methodological novelty. Previous studies have investigated moving-sound navigation [1], continuous semantic audio-visual navigation [2], and audio-visual heatmap-based navigation [3]. The authors should better clarify the methodological advances of DyGMap beyond existing cross-modal attention and temporal decay techniques.

  2. Insufficient evidence for dynamic goal mapping. Removing temporal decay and the timestamp channel reduces success rates by only 1.17 and 2.31 percentage points in clean and noisy settings, respectively (Tables 8 and 5). Their individual contributions remain unclear.

  3. Unclear benefit of visual fusion during acoustic silence. Table 11 shows that adding vision reduces F≤20° from 0.244 to 0.224 (clean) and from 0.146 to 0.120 (noisy), while noisy localization error increases from 45.53° to 50.15°. These findings do not consistently support the claimed benefit of visual information during silence.

  4. Insufficient control of benchmark difficulty. DANCE uses predefined target trajectories and a fixed target speed at half the agent's speed. Table 9 shows higher success rates for moving targets than stationary targets, raising concerns about differences in task difficulty. Better-controlled comparisons are needed to isolate the challenges introduced by target motion.

References

[1] Catch Me If You Hear Me: Audio-Visual Navigation in Complex Unmapped Environments with Moving Sounds. IEEE Robotics and Automation Letters (2023).

[2] Semantic Audio-Visual Navigation in Continuous Environments. CVPR (2026).

[3] Learning to Chase: Adaptive Audio-Visual Navigation for Moving Sounds in Complex Environments. Pattern Recognition Letters (2026).

Questions for authors
  1. What are the substantive differences between DyGMap and the approaches in [1–3], particularly regarding cross-modal fusion and spatial goal representation? What new capability or insight does the dynamic goal map provide?

  2. Could the authors independently ablate temporal decay and the goal-update timestamp channel, and evaluate different decay factors? This would help determine which component accounts for the gains in Tables 5 and 8 and how robust those gains are to the decay setting.

3 comments
AuthorsOpenReview ↗

Response to the Reviewer, Part 1/2

W1 & Q1: Relationship to prior work and methodological differences

These three papers are closely related to our work. Table 1 compares their task settings with ours. Studies [1] and [3] address the DAV-Nav task, which does not consider continuous environments, intermittent sound emission, or semantic grounding. Study [2] addresses the SAVN-CE task, which assumes stationary targets. DANCE jointly considers all four characteristics.

Table 1: Comparison of representative audio-visual navigation tasks.

TaskContinuous environmentIntermittent soundSemantic groundingMoving target
AVN✗✗✗✗
SAVN✗✓✓✗
SAVN-CE✓✓✓✗
DAV-Nav✗✗✗✓
DANCE (proposed)✓✓✓✓

DyGMap combines audio and visual observations with self-motion cues in separate branches and stores the resulting representations in two separate episodic memory sequences. Separate Transformer encoders process the audio and visual memory sequences, and a Transformer decoder fuses them before the MLP predicts the target's category and location. We use these predictions to construct category-specific goal maps and introduce a dynamic update mechanism that records the latest update time at each location and applies temporal decay to unrefreshed map evidence.

Study [1] builds maps of obstacles and explored regions from depth and uses convolutional and deconvolutional layers to project audio spectrograms to the same spatial dimensions. It fuses these representations to learn geometric relationships between local maps and acoustic observations and passes the combined features to the policy.

Building on [1], study [3] combines visual features with audio features before projecting them to the geometric map's dimensions. Three cross-attention branches fuse the resulting map features with visual and audio features in different orders. Their concatenated outputs guide navigation.

MAGNet, proposed in study [2], employs a memory-augmented GDN that integrates audio observations, self-motion cues, and episodic memory to jointly infer the spatial and semantic properties of a single target. Its GDN uses no visual information for goal reasoning, and MAGNet does not construct spatial maps. Although the GDN produces explicit target estimates, MAGNet uses the GDN's intermediate goal embedding to guide navigation.

These three studies use latent spatial goal representations to guide navigation. DyGMap additionally constructs explicit category-specific goal maps from the GDN's target estimates, while retaining the most recent encoded representation from the GDN's audio branch as its goal embedding. None of these three studies applies temporal decay. Although temporal decay is an established technique, DyGMap applies it to category-specific goal maps built from successive target estimates, while preserving the geometric map of the scene. This reduces the influence of evidence remaining at former target locations and of transient prediction errors.

AuthorsOpenReview ↗

Response to the Reviewer, Part 2/2

W2 & Q2: Component ablations and the temporal decay factor

Our common response Dynamic Goal-Map Ablations, Decay-Factor Sensitivity, and the Effect of Target Speed reports both experiments. Compared with the variant with both components disabled, the goal-update timestamp channel alone increases SR by 0.83 percentage points in the clean setting and 1.16 in the noisy setting, while temporal decay alone increases SR by 1.23 and 1.85 percentage points, respectively. On moving-target episodes, with model weights held fixed, SR varies by only 0.53 percentage points in the clean setting and 0.39 in the noisy setting across the tested β values from 0.80 to 0.95, but increasing β from 0.95 to 1.00 reduces SR by 3.70 and 5.55 percentage points, respectively.

W3: Benefits of visual integration during acoustic silence

As detailed in Does Visual Integration Benefit Target Estimation?, we evaluated all GDNs along the same oracle trajectories. Adding vision improves most target estimation metrics in both the clean and noisy settings during both the sound-emitting and silent periods. During silence, the F-score increases from 0.292 to 0.325 in the clean setting and from 0.230 to 0.243 in the noisy setting.

W4: Target motion and benchmark difficulty

Our common response Dynamic Goal-Map Ablations, Decay-Factor Sensitivity, and the Effect of Target Speed reports how navigation performance varies across seven target speeds. At each speed, we compared DyGMap with and without dynamic updates. In the clean setting, the largest gain is 3.51 percentage points (from 57.49% to 61.00%) at 0.25 times the agent's speed. In the noisy setting, the largest gain is 4.24 percentage points (from 37.75% to 41.99%) at 0.75 times the agent's speed.

For a given episode, the predefined target trajectory is the same in both the stationary and moving settings. A simulator parameter determines whether the target remains at the trajectory's starting position or moves along it. The default mixed setting samples stationary and moving targets with equal probability. In the stationary setting, each target stays at the starting position of its trajectory. In the moving setting, each target follows its predefined trajectory.

Appendix C.3 suggests that moving targets can be slightly easier for the oracle policy to reach. For a learned policy, target motion can create new opportunities to observe the target while also making it harder to track. The difficulty therefore depends in part on how well each method can pursue a moving target. This made it difficult to anticipate and balance the difficulty of stationary- and moving-target episodes during dataset construction. We therefore evaluate the methods on stationary targets and on targets moving at different speeds to examine how motion affects navigation performance.

Dear Reviewer J97o,

With the rebuttal period coming to an end, we would appreciate any feedback on our responses. We would be happy to address any remaining concerns before the deadline.

Thank you for your time and consideration. We look forward to your reply.

Reviewer_Um38OpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces DANCE, which studies audio-visual navigation toward intermittently sound-emitting, moving or stationary targets in continuous environments. The DyGMap method combines audio-visual target estimation, persistent geometric maps, and dynamic goal maps, achieving success rates of 55.20% and 42.37% in the clean and noisy settings, respectively, improving by approximately 11.6 percentage points over the strongest baseline in each setting.

Critical strengths
  1. DANCE combines continuous environments, semantic targets, intermittent sound emission, and target motion, requiring the agent to continue localizing the target when acoustic evidence is outdated or temporarily unavailable, thereby extending the evaluation scope of stationary sound-source navigation. Training, validation, and test use different scenes, target assets, and target and distractor sound recordings, which helps reduce the influence of directly memorizing specific scenes or assets.

  2. The method preserves obstacles, explored regions, and the agent trajectory, while independently updating and decaying category-specific goal maps, clearly distinguishing persistent scene structure from time-varying target evidence. In the noisy setting, adding the scene map increases the success rate from 33.14% to 38.37%, and further adding the goal map and dynamic updates reaches 42.37%, supporting the role of explicit spatial representations.

  3. DyGMap outperforms the compared methods in both clean and noisy settings, with advantages reflected in success rate, SPL, and SNA rather than relying on a single metric. The paper also provides analyses of target motion state, category, acoustic interference strength, and failure types, as well as multiple oracle variants.

Critical weaknesses
  1. The contributions should be better summarized. The statement “We conduct comprehensive experiments in both clean and noisy settings and demonstrate the effectiveness of DyGMap, which improves the success rate by approximately 11.6 percentage points over the strongest baseline in each setting.” cannot be considered as a contribution. Conducting experiments is a means to verify the effectiveness of your technical innovations, not an end in itself.

  2. DAV-Nav's waypoint-selection mechanism is replaced with direct action prediction; the current results reflect the influence of both the original method and the adaptation, and the effect of the replacement itself still needs to be verified. DyGMap adopts a three-stage training strategy, and Table 7 does not include all prior training and initialization costs, so it cannot be used to determine whether the main methods received comparable cumulative training investment.

  3. “Removing dynamic updates” disables temporal decay and removes the timestamp at the same time; the existing ablation can only determine their joint effect and cannot attribute it solely to temporal decay. This joint benefit corresponds to success-rate increases of 2.31 percentage points in the noisy setting and 1.17 percentage points in the clean setting. If the importance of the decay mechanism itself is emphasized, temporal decay and the timestamp should be tested separately, and the sensitivity of the results to β should be reported.

  4. Visual integration has condition-dependent benefits: full and no-vision models perform similarly in clean conditions (55.20% vs. 55.28%), while vision improves success under noise (42.37% vs. 37.10%). Table 11, Fig. 15, and Appendices F.2/F.5 provide preliminary perception-level evidence and explanations. However, evaluation along each policy’s own trajectories limits controlled comparison, leaving unclear how much of the navigation gain comes from improved target estimation.

  5. The reported standard deviations do not reflect the stability of independent training. The appendix states that the five runs are repeated evaluations on selected checkpoints, intended to reflect fluctuations in the evaluation process such as acoustic simulation, rather than five independent training runs. This does not negate the relatively large overall performance gaps, but it is insufficient to judge whether smaller component gains are stable, especially the 1–3 percentage-point improvement brought by dynamic updates. It is recommended to supplement independent training results at least for the full model, the strongest baseline, and key ablations.

  6. Targets include only three animal categories, move at a fixed speed equal to half that of the agent, and follow predefined paths; current evidence mainly supports this motion protocol. Table 9 demonstrates that the method can intercept moving targets under this setting, but performance under different speeds, motion patterns, or longer silence intervals has not been verified. Therefore, conclusions about dynamic-target navigation capability should be limited to the current protocol and not directly generalized to more general target motion scenarios.

  7. The clean-trained model achieves a success rate of about 10% in the noisy evaluation in Figure 13, indicating that robustness highly depends on interference training and that the applicable scope across acoustic distributions needs to be limited.

Questions for authors
  1. After replacing DAV-Nav's waypoint-selection mechanism, how is the reliability of the adaptation verified? Are the initialization sources and cumulative training investment of the methods comparable?

  2. Can the independent contributions of temporal decay and the timestamp be verified separately, and can the sensitivity of the dynamic-update benefit to the decay factor β be clarified?

  3. Are there results from independent training seeds that can support the advantage of the full model over the strongest baseline and the stability of smaller component gains in key ablations?

3 comments
AuthorsOpenReview ↗

Response to the Reviewer, Part 1/2

W1: Summary of contributions

Thank you for pointing this out. We agree that the original statement emphasized DyGMap's performance gain without explaining what readers could learn from the evaluation. We have revised it as follows. "We systematically evaluate representative methods on DANCE under clean and noisy conditions and analyze how their navigation performance varies with acoustic availability, navigation distance, acoustic interference, target speed, and target category."

W2 & Q1: Adaptation of DAV-Nav, model training, and initialization

In our common response Adaptation of DAV-Nav, we describe how we adapted DAV-Nav for evaluation on DANCE. We did not verify the effect of replacing the waypoint selector. We retained DAV-Nav's central design while replacing its waypoint selector with direct action prediction, the action interface used by all learning-based methods in our comparison. We consider the comparison informative within the stated action-prediction setting. We discuss model training and initialization in our common response Model Training, Weight Initialization, and Independent Training Runs.

W3 & Q2: Dynamic goal-map updates and sensitivity to the temporal decay factor

In our common response Dynamic Goal-Map Ablations, Decay-Factor Sensitivity, and the Effect of Target Speed, we report the separate contributions of temporal decay and the goal-update timestamp channel and examine sensitivity to β. Compared with the variant with both components disabled, the goal-update timestamp channel alone increases SR by 0.83 percentage points in the clean setting and 1.16 in the noisy setting, while temporal decay alone increases SR by 1.23 and 1.85 percentage points, respectively. On moving-target episodes, SR varies by only 0.53 percentage points in the clean setting and 0.39 in the noisy setting across the tested β values from 0.80 to 0.95, but increasing β from 0.95 to 1.00 reduces SR by 3.70 and 5.55 percentage points, respectively.

W4: Visual integration and target estimation

Please see our common response Does Visual Integration Benefit Target Estimation? We evaluated all GDNs along the same oracle trajectories. Across the different task settings, DyGMap's vision-integrated GDN outperforms its audio-only counterpart on most target estimation metrics.

In DyGMap, acoustic information about the target reaches the navigation policy through three inputs:

  • the audio embedding (etBe_t^B), produced directly by the audio encoder,
  • the goal embedding (etGe_t^G), which is the most recent encoded representation from the audio branch of the GDN, and
  • the map embedding (etMe_t^M), which encodes a local map that includes the goal map constructed from the GDN's target estimates.

The policy uses these inputs jointly, making it difficult to isolate how much of the navigation gain is attributable to improved target estimation. Our results suggest that adding vision to the GDN has little effect on navigation SR in the clean setting but increases SR in the noisy setting. We hypothesize that, in the clean setting, the first two inputs already provide much of the information the policy needs about the target. In the noisy setting, the policy may rely more heavily on the GDN's visually informed target estimates, which reach it through the goal map and provide cues that the first two inputs cannot reliably supply.

AuthorsOpenReview ↗

Response to the Reviewer, Part 2/2

W5 & Q3: Independent training runs

We agree that the reported standard deviations capture variability across evaluations rather than variability across independent training runs. Please see our common response Model Training, Weight Initialization, and Independent Training Runs. Training each model takes one to two weeks, and the available computing resources and time did not allow us to train the models with multiple random seeds.

In our evaluations, we compared the navigation performance of the different methods and conducted detailed ablation studies in both clean and noisy settings. Together, these results support the consistency and reliability of our conclusions across the two settings. The quantitative comparisons and qualitative analyses in the appendix provide further evidence of our model's effectiveness. Compared with DAV-Nav, DyGMap achieves SR gains of 11.62 percentage points in the clean setting and 11.60 points in the noisy setting. These gains, together with improvements in other navigation metrics, support DyGMap's overall performance advantage in both settings.

We agree that the smaller gains from dynamic goal-map updates warrant a more cautious interpretation. We examined the contribution of this mechanism in three ways.

  • Component ablations. We evaluated temporal decay and the goal-update timestamp channel separately to assess the contribution of each component in both the clean and noisy settings.

  • Decay-factor sensitivity and target speed. With the training value fixed at β = 0.95, we varied β at test time to assess sensitivity to temporal decay. We also compared DyGMap with and without dynamic updates at different target speeds to examine how the navigation gains vary with target motion.

  • Goal-map visualizations. Figure 4 shows how the goal map evolves during navigation, with the full video provided in the supplementary materials. Figure 10 compares dynamic and static goal maps along the same trajectory, illustrating how dynamic updates attenuate evidence at previously occupied target locations.

W6: Target categories, motion patterns, and silence intervals

We discuss the choice of target categories in Limited Number of Target Categories in DANCE. We have also added results at seven different target speeds, as described in Dynamic Goal-Map Ablations, Decay-Factor Sensitivity, and the Effect of Target Speed. To generate a target trajectory, we follow the shortest path between the start and end positions using actions from the agent's action space. We preserve the brief pauses that occur when the target turns, so the trajectory includes periods of both movement and rest. This approach produces simple, controllable trajectories. We have not yet explored other motion patterns.

If by silence intervals the reviewer means how long the agent must navigate after the target stops emitting sound, we examine a related measure in the Acoustic Availability analysis in Appendix F.3. We report cumulative success rates for all methods as a function of the action ratio, defined as the number of steps along the oracle trajectory divided by the number of steps during which the target is acoustically active. A higher action ratio suggests that silence intervals may account for a greater proportion of the time the agent spends navigating.

If the reviewer means the pauses between successive sounds within the emission period, their duration depends on the source recording. Each target category includes multiple recordings with different pause durations. Our current setup does not allow us to vary these pauses independently, which makes it difficult to isolate their effect on navigation performance.

W7: Performance under acoustic interference

We designed this experiment to assess how well a model trained under clean conditions performs in a noisy environment that more closely resembles real-world conditions, and whether the resulting performance loss would be acceptable for deployment on a physical robot. The clean-trained model's performance in the noisy setting fell below what we consider acceptable for deployment. This result underscores the importance of accounting for acoustic interference and background noise when preparing models for physical robots.

In the standard noisy setting, the distractor SIR ranges from 0 to 10 dB, and the background-noise SNR ranges from 0 to 20 dB. In the evaluation shown in Figure 13, we varied each ratio separately from −20 to 60 dB in 5 dB increments. We held the SNR at 20 dB when varying the SIR and the SIR at 10 dB when varying the SNR. We have revised the manuscript to limit our conclusions to the acoustic conditions tested.

Dear Reviewer Um38,

With the rebuttal period coming to an end, we would appreciate any feedback on our responses. We would be happy to address any remaining concerns before the deadline.

Thank you for your time and consideration. We look forward to your reply.

Reviewer_1E9TOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Contributions: This paper introduces DANCE, a new audio-visual navigation task in continuous environments where an agent uses intermittent acoustic cues and visual observations to navigate toward a single sound-emitting target that may be moving or stationary; proposes DyGMap, which combines audio-visual goal estimation with a dynamic map that preserves scene geometry while maintaining time-varying target information; and conducts comprehensive experiments in both clean and noisy settings, demonstrating that DyGMap improves the success rate by approximately 11.6 percentage points over the strongest baseline in each setting.

Critical strengths
  1. The motivation is clear and interesting.

  2. The figures and text are described in considerable detail.

Critical weaknesses
  1. The paper has not yet sufficiently established that the core dynamic mapping mechanism is the primary driver of the overall performance gains. Compared with the variant lacking the dynamic component, the full model improves the success rate by only 1.17% and 2.31% in the clean and noisy settings, respectively, indicating that most of the gains are already achieved without this component. Moreover, the ablation simultaneously disables temporal decay and removes the goal-update timestamp channel, making it impossible to isolate their individual contributions or to attribute the improvement specifically to the suppression of outdated target evidence. Although the visualizations show fewer stale traces after applying decay, they demonstrate a change in the map representation rather than provide direct evidence that this change improves navigation performance.

  2. This paper directly interprets higher SPL/SNA as higher navigation efficiency, but these metrics are simultaneously affected by success rate and the efficiency of successful trajectories, so their improvement alone cannot prove that paths are shorter or that fewer actions are taken. For example, under the noisy setting, DyGMap's SPL/SR is approximately 0.766, slightly lower than DAV-Nav's 0.773, indicating that its higher overall SPL may not necessarily stem from an improvement in the average path efficiency of successful trajectories. However, because the sets of successful episodes differ between the two methods, this also cannot conversely prove that DyGMap is less efficient. In my view, this work should distinguish between success rate improvement and trajectory efficiency improvement, and compare path length, number of actions, or interception time on commonly successful episodes to support conclusions regarding efficiency.

Questions for authors
  1. Does the adaptation of DAV-Nav affect its performance as a baseline?

  2. Do the gains from dynamic mapping come from temporal decay or the goal-update timestamp channel?

2 comments
AuthorsOpenReview ↗

Response to the Reviewer

The summary paragraph describing D2AVN, both items under Critical Strengths, and the first item under Critical Weaknesses appear to concern a different paper. We respectfully ask the reviewer to revisit these comments and reassess our manuscript. We address the remaining two weaknesses and both questions below.

W2 & Q2: Contributions of temporal decay and the goal-update timestamp channel

1. We agree that DyGMap with a static goal map already achieves most of the full model's SR gain over DAV-Nav. Our common response Dynamic Goal-Map Ablations, Decay-Factor Sensitivity, and the Effect of Target Speed examines how much further dynamic updates improve SR at each target speed. In the clean setting, the largest gain is 3.51 percentage points (from 57.49% to 61.00%) at 0.25 times the agent's speed. In the noisy setting, the largest gain is 4.24 percentage points (from 37.75% to 41.99%) at 0.75 times the agent's speed.

2. In the same common response, we also assess the individual contributions of temporal decay and the goal-update timestamp channel. Compared with the variant with both components disabled, enabling only the goal-update timestamp channel increases SR by 0.83 percentage points in the clean setting and 1.16 percentage points in the noisy setting. Enabling only temporal decay increases SR by 1.23 and 1.85 percentage points, respectively. With both components enabled, the corresponding gains are 1.17 and 2.31 percentage points. Each component therefore improves SR on its own in both settings.

3. We agree that visualizations alone cannot establish how temporal decay affects navigation performance. In the same common response, we also tested whether a policy trained with β = 0.95 relies on decay by varying only β at test time in the moving-only setting. Disabling temporal decay (β = 1.00) reduces SR by 3.70 percentage points in the clean setting and 5.55 percentage points in the noisy setting relative to β = 0.95. SPL and SNA also decrease, and the final distance to the target increases in both settings. These results suggest that the policy has learned to use temporal decay to distinguish recently supported target locations from locations supported only by older observations. When decay is disabled at test time, evidence at previously occupied locations can persist, making it harder for the policy to identify the target's likely current location.

W3: Navigation efficiency on episodes where both methods succeed

We agree that overall SPL and SNA reflect both success rate and efficiency on successful trajectories. Comparing SPL/SR or SNA/SR across methods also requires care when methods succeed on different episodes. Following the reviewer's suggestion, for each method (DAV-Nav and DyGMap), we identified the episodes in which it succeeded in at least three of the five evaluation runs. We then took the intersection of the two sets. For each method, we computed the mean SPL, SNA, and SWM using only its successful runs on episodes in the intersection.

Table 1: Navigation performance on episodes in which both DAV-Nav and DyGMap succeeded.

SettingMotion state# EpisodesMethodSPL ↑SNA ↑SWM ↑
CleanStationary678DAV-Nav79.8167.23–
DyGMap86.7981.29–
Moving739DAV-Nav78.4970.3670.57
DyGMap83.9281.0081.44
NoisyStationary409DAV-Nav78.9266.37–
DyGMap79.2074.48–
Moving465DAV-Nav77.2468.9172.93
DyGMap76.3274.5278.57

Although DyGMap achieves higher SNA in all four groups and higher SPL in three of the four, indicating generally better navigation efficiency on these shared episodes, we have revised the manuscript to avoid implying that higher overall SPL and SNA scores alone establish greater navigation efficiency.

Q1: Effect of adapting DAV-Nav

As detailed in Adaptation of DAV-Nav, we retained DAV-Nav's central design: fusing audio and visual information in a spatial feature space to learn the geometric relationships between local maps and acoustic observations. We replaced DAV-Nav's original waypoint selector with direct action prediction, the interface used by all learning-based methods in our evaluation, including DyGMap.

Under this common action interface, DyGMap outperforms the adapted DAV-Nav by 11.62 percentage points in success rate in the clean setting and 11.60 percentage points in the noisy setting. We consider this comparison informative within the same action-prediction setting. The effect of replacing the selector itself remains unverified. If waypoint selection proves beneficial in continuous semantic audio-visual navigation, DyGMap could naturally use the cell with the highest goal-map value as a long-term goal.

Dear Reviewer 1E9T,

With the rebuttal period coming to an end, we would appreciate any feedback on our responses. We would be happy to address any remaining concerns before the deadline.

As noted in our earlier response, the summary paragraph describing D2AVN, both items under Critical Strengths, and the first item under Critical Weaknesses appear to concern a different paper. We respectfully ask you to revisit these comments, consider our responses to the remaining concerns, and reassess our manuscript accordingly.

Thank you for your time and consideration. We look forward to your reply.

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

This paper introduces Dynamic Audio-Visual Navigation in Continuous Environments. The task involves navigating continuously, understanding what you are looking at, hearing sounds now and then, and following targets that might move. DyGMap uses audio-visual goal descriptors, category-specific goal maps, scene geometry that doesn't change over time, and update timestamps.

DyGMap is successful 55.20% of the time in quiet settings and 42.37% of the time in noisy settings. These results are 11.62 and 11.60 percentage points better than the strongest evaluated baseline, respectively.

Critical strengths

1.I find DANCE a useful way to figure out where to go when there is no sound or the sound is bad. It combines continuous control, semantic grounding, and target motion. Earlier settings don't cover these aspects together. The splits also separate scenes, target recordings and target assets. This helps test how well they can apply what they know to situations that are different from the ones they've seen before.

2.The full system shows significant improvements in the main navigation metrics. The oracle comparisons and component ablations help us to understand the roles of target information and spatial memory. The acoustic interference tests and failure diagnostics provide useful background information. The design is the main reason for my positive rating, as are the results.

Critical weaknesses

1.I still find it hard to tell what the dynamic update rule contributes on its own. The experiments in Tables 5 and 8 enable temporal decay and timestamp encoding together. SR improves by 1.17 percentage points in clean conditions and 2.31 percentage points under noise. But the motion analysis doesn’t show consistent benefits for moving targets. The diagnostic for stopping at previous target positions also doesn’t consistently improve (Tables 9 and 12). Figure 10 already compares maps along identical trajectories. Quantitative results from this setup would help show whether the update rule reduces evidence at obsolete target locations. The paper also reports five evaluations of selected checkpoints. These evaluations don’t resolve variation across independent training runs. That uncertainty matters when the gains are relatively small.

2.The closest baseline comparisons still involve some important adaptation differences. The authors replace DAV-Nav’s waypoint selection with direct action prediction (Appendix D.2). Appendix D.3 also describes different training stages and initialization histories. The internal ablations partly address these concerns, but their exact relationship to MAGNet needs clarification. These differences make it hard to attribute the full comparative gain solely to the proposed representation.

Questions for authors

1.How much does the passage of time affect the evaluation when it is done separately from the encoding of timestamps? How much does timestamp encoding contribute? If we look at the same paths in Figure 10, does updating the target location make it more accurate or reduce the evidence at locations that are no longer useful?

2.What is the difference between MAGNet and DyGMap in terms of how they are built, how they are set up, and how they are trained? What evidence is there that replacing DAV-Nav's waypoint selection keeps the baseline effective?

3.Do the gains over the basic static map control happen in the same way across policies that have been trained separately?

2 comments
AuthorsOpenReview ↗

Response to the Reviewer

W1 & Q1: Dynamic goal-map updates and their evaluation

1. Our common response Dynamic Goal-Map Ablations, Decay-Factor Sensitivity, and the Effect of Target Speed reports separate ablations. Compared with the variant with both components disabled, enabling only the goal-update timestamp channel increases SR by 0.83 percentage points in the clean setting and 1.16 percentage points in the noisy setting. Enabling only temporal decay increases SR by 1.23 and 1.85 percentage points, respectively. With both components enabled, the corresponding gains are 1.17 and 2.31 percentage points. Each component therefore improves SR on its own in both settings.

2. We used the stopping at previous target positions diagnostic to assess whether the agent could distinguish current from past target positions. On revisiting our analysis, we identified two limitations that we had overlooked.

  • We originally pooled stationary- and moving-target episodes but now analyze them separately. Among moving-target episodes, the rate of stopping at a previous target position is 7.49% for DyGMap w/o dynamic and 7.17% for DyGMap in the clean setting. In the noisy setting, the corresponding rates are 9.87% and 7.58%.

  • The diagnostic records whether the agent stops near the target's past trajectory. Two behaviors can affect the reported rate: (1) An agent stops at a past target position because it mistakes that position for the target's current location. This is the behavior we intended to detect. (2) An agent that continues pursuing the target but ultimately fails to reach it may finish with a lower distance to goal (DTG) and have more opportunities to stop near the target's past trajectory than an agent that gives up early.

From the stopping position alone, we cannot reliably distinguish a stop during an unsuccessful pursuit from one caused by mistaking a past target location for the current one. We therefore recommend interpreting this rate alongside other navigation metrics such as DTG.

3. We currently lack a direct quantitative measure of goal-map quality that would allow us to determine whether dynamic updates reduce outdated target evidence and, if so, by how much. Our common response Does Visual Integration Benefit Target Estimation? evaluates the goal descriptor networks (GDNs) along the same oracle trajectories, showing that incorporating visual observations improves target estimation on most metrics. DyGMap separates target estimation from goal-map construction. The GDN estimates target category and location, and the mapping module uses these estimates to construct goal maps and attenuates evidence at locations that receive no new estimates. We assess the quality of goal-map construction through visualizations (e.g., Figures 4 and 10) and, indirectly, through quantitative navigation metrics.

W2 & Q2: Relationship to MAGNet, differences in training, and adaptation of DAV-Nav

Our common responses Model Training, Weight Initialization, and Independent Training Runs and Adaptation of DAV-Nav detail the training stages, weight initialization, and changes to DAV-Nav. We build DyGMap on MAGNet's memory-augmented goal descriptor network (GDN) and make two key technical contributions.

  • We incorporate visual observations to complement the acoustic input and improve target estimation on most metrics.
  • The most substantial difference from MAGNet is the introduction of explicit spatial maps that record obstacles, explored regions, the agent's trajectory, and the distribution of target evidence by category. Each map cell also records when its target evidence was last updated. These maps give the policy access to the locations of obstacles and explored regions, the agent's trajectory, and the location, strength, and recency of evidence for each target category as the agent explores.

Our ablations examine how these components contribute to navigation. The first row of Table 5 provides a MAGNet-style reference. The full system improves SR over MAGNet by approximately 11.8 percentage points in the clean setting and 12.9 points in the noisy setting.

W1 & Q3: Independent training and consistency of the results

As noted in Model Training, Weight Initialization, and Independent Training Runs, time and computational constraints prevented us from repeating training with multiple random seeds. We have evaluated all models at seven target speeds, and these results provide further evidence that the SR gains from dynamic goal-map updates are consistent across the tested speeds. Qualitative comparisons of static and dynamic goal maps and ablation studies in both the clean and noisy settings further support these findings.

Dear Reviewer eQEU,

With the rebuttal period coming to an end, we would appreciate any feedback on our responses. We would be happy to address any remaining concerns before the deadline.

Thank you for your time and consideration. We look forward to your reply.

Comments on the paper

7 comments
AuthorsOpenReview ↗

Response from the Authors

We thank all five reviewers for their careful reading of our manuscript and their constructive suggestions. We appreciate their recognition of the motivation behind DANCE and their thoughtful questions about the benchmark and our evaluation of DyGMap. Their feedback prompted us to examine the contributions of individual model components more closely and to clarify how the observed gains vary with target speed and acoustic conditions.

To avoid repetition, we first address five concerns raised by multiple reviewers.

1. Dynamic Goal-Map Ablations, Decay-Factor Sensitivity, and the Effect of Target Speed.

2. Adaptation of DAV-Nav.

3. Model Training, Weight Initialization, and Independent Training Runs.

4. Limited Number of Target Categories in DANCE.

5. Does Visual Integration Benefit Target Estimation?

After these common responses, we address each reviewer's remaining concerns directly below their comments.

Summary of manuscript revisions

We have revised the manuscript in response to the reviewers' feedback. Revisions are marked in red. The main revisions are as follows.

1. Dynamic update mechanism and sensitivity to the decay factor. We added component ablations in both the clean and noisy settings to evaluate the individual contributions of temporal decay and the goal-update timestamp channel. We also added evaluation results for five values of the temporal decay factor β (0.80, 0.85, 0.90, 0.95, and 1.00).

2. The effect of target speed. We expanded the original comparison of stationary and moving targets in the subsection Motion State to cover seven target-to-agent speed ratios ranging from 0 to 1.50 in increments of 0.25. We renamed the subsection Target Speed to reflect this change.

3. Separate evaluations for stationary and moving targets. We updated the analyses of Acoustic Availability, Navigation Distance, and Failure Analysis to report results separately for stationary and moving targets. The original analyses used a mixed setting in which the two motion states were sampled with equal probability. Evaluating them separately prevents variation in the mix of stationary and moving episodes from affecting the comparisons.

4. Target estimation along the same oracle trajectories. We evaluated the goal descriptor networks (GDNs) of the different methods along the same trajectories generated by an oracle navigation policy. This allows us to compare their target estimation performance while controlling for differences in the paths taken by their own navigation policies.

5. Clarifications to the text. We have revised the wording of several passages in response to the reviewers' comments.

AuthorsOpenReview ↗

Common Response 1 of 5, Part 1/2: Dynamic Goal-Map Ablations, Decay-Factor Sensitivity, and the Effect of Target Speed

In the submitted manuscript, we ablated temporal decay and the goal-update timestamp channel together because they play complementary roles in accounting for changes in the target's location. However, the joint ablation did not show how much each component contributes or whether each is beneficial on its own. We therefore conducted additional ablation experiments to assess their individual contributions.

1. Ablations of the dynamic goal-map update mechanism. We evaluated two additional variants: one retains temporal decay but removes the goal-update timestamp channel, and the other retains the timestamp channel but disables temporal decay. The results are as follows.

Table 1: Ablation study of the proposed dynamic update mechanism in the clean and noisy settings.

SettingTimestampTemporal decaySR ↑SPL ↑SNA ↑DTG ↓SWS ↑SWM ↑DSR ↓
Clean✗✗54.03 ± 0.4945.66 ± 0.3943.59 ± 0.325.50 ± 0.0620.95 ± 0.3621.26 ± 0.38–
✓✗54.86 ± 0.5245.67 ± 0.5044.26 ± 0.475.46 ± 0.0522.13 ± 0.7221.19 ± 0.31–
✗✓55.26 ± 0.7045.80 ± 0.4844.32 ± 0.535.48 ± 0.1022.02 ± 0.8821.44 ± 0.36–
✓✓55.20 ± 0.9447.08 ± 0.7744.77 ± 0.725.19 ± 0.1021.94 ± 0.7722.03 ± 0.61–
Noisy✗✗40.06 ± 0.7629.11 ± 0.4327.87 ± 0.367.40 ± 0.1220.01 ± 0.6314.20 ± 0.590.54 ± 0.12
✓✗41.22 ± 0.4430.17 ± 0.2628.50 ± 0.227.48 ± 0.1121.56 ± 0.5613.92 ± 0.580.36 ± 0.08
✗✓41.91 ± 0.7830.74 ± 0.6229.57 ± 0.577.29 ± 0.1820.47 ± 0.6514.57 ± 0.470.27 ± 0.13
✓✓42.37 ± 0.8532.45 ± 0.6330.97 ± 0.587.33 ± 0.1118.82 ± 0.7714.60 ± 0.320.49 ± 0.07

In both clean and noisy settings, the results indicate that each component can improve navigation performance on its own. Temporal decay attenuates map evidence that has not been refreshed, which may reduce the influence of outdated target locations and transient prediction errors. The timestamp channel records the time of the most recent update at each location, potentially helping the policy decide how much to rely on retained estimates. Combining these mechanisms may be particularly beneficial when unreliable estimates accumulate.

2. Sensitivity to the temporal decay factor β. To examine sensitivity to the temporal decay factor β, we evaluated a DyGMap model trained under the standard configuration (β = 0.95), varying β only at test time in the moving-only setting. The results are shown below.

Table 2: DyGMap navigation performance with respect to the temporal decay factor β in the moving-only setting.

SettingDecay factor βSR ↑SPL ↑SNA ↑DTG ↓SWS ↑SWM ↑DSR ↓
Clean0.8056.27 ± 0.2947.18 ± 0.3545.30 ± 0.384.83 ± 0.0322.43 ± 0.6943.97 ± 0.14–
0.8556.52 ± 0.6447.53 ± 0.7045.76 ± 0.694.78 ± 0.0522.05 ± 0.4744.26 ± 0.63–
0.9056.28 ± 0.8647.30 ± 0.6745.54 ± 0.634.76 ± 0.0722.05 ± 0.7244.02 ± 0.71–
0.9556.80 ± 0.3747.66 ± 0.4446.09 ± 0.464.76 ± 0.0622.47 ± 0.5144.58 ± 0.48–
1.0053.10 ± 0.9744.35 ± 0.6943.02 ± 0.654.95 ± 0.0919.41 ± 0.5042.12 ± 0.75–
Noisy0.8043.20 ± 0.4332.55 ± 0.4031.68 ± 0.336.88 ± 0.1120.04 ± 0.4830.73 ± 0.600.30 ± 0.07
0.8543.56 ± 0.6132.67 ± 0.4731.99 ± 0.516.83 ± 0.1119.57 ± 0.4430.78 ± 0.690.23 ± 0.10
0.9043.18 ± 0.7032.56 ± 0.5431.89 ± 0.566.85 ± 0.0519.38 ± 0.7430.95 ± 0.590.35 ± 0.07
0.9543.57 ± 1.0833.00 ± 0.9132.28 ± 0.956.86 ± 0.1119.48 ± 0.6731.09 ± 0.840.43 ± 0.07
1.0038.02 ± 0.4229.20 ± 0.3028.47 ± 0.297.32 ± 0.1115.39 ± 0.1428.12 ± 0.580.30 ± 0.09

Navigation performance changes little across the tested values from β = 0.80 to 0.95. Disabling temporal decay (β = 1.00), however, leads to a clear decline in both settings. Changing β from 0.95 to 1.00 at test time reduces SR by 3.70 and 5.55 percentage points in the clean and noisy settings, respectively. SPL and SNA also decrease, and the final distance to the target increases in both settings. These results suggest that the trained policy benefits from temporal decay while remaining relatively insensitive to its exact value over the tested range below 1. With β = 1.00, an erroneous estimate or a previously correct estimate of a target that has since moved can therefore persist in the goal map, making it harder for the policy to distinguish the target's likely current location from locations supported only by older observations.

AuthorsOpenReview ↗

Common Response 1 of 5, Part 2/2: Dynamic Goal-Map Ablations, Decay-Factor Sensitivity, and the Effect of Target Speed

3. The effect of target speed. Several reviewers noted that the gains from dynamic goal-map updates were modest and that their magnitude varied with the target's motion state and the acoustic setting. However, the original analysis considered only two conditions, one with stationary targets (0.00×) and the other with targets moving at half the agent's speed (0.50×). To examine how the benefit of dynamic goal-map updates varies with target speed, we constructed five additional test sets with target speeds of 0.25, 0.75, 1.00, 1.25, and 1.50 times the agent's speed. To isolate the effect of speed, we adjusted the interpolation only along the moving portions of each predefined target trajectory and kept all parameters unrelated to speed unchanged. We summarize the results below and report the detailed results for all methods in the appendix.

Table 3: Navigation performance with respect to target speed in the clean setting.

MethodSpeedSR ↑SPL ↑SNA ↑DTG ↓SWS ↑SWM ↑
DyGMap w/o dynamic0.00×53.4746.1043.105.9722.00–
0.25×57.4947.3546.254.8021.8655.60
0.50×55.2846.1445.014.9720.7343.10
0.75×53.0044.2843.155.3921.0427.22
1.00×51.6142.7341.925.7421.7814.84
1.25×51.5442.7441.945.7921.689.91
1.50×51.1542.3441.465.9221.817.28
DyGMap (Ours)0.00×55.5748.0144.935.6023.33–
0.25×61.0050.6249.034.3725.3559.06
0.50×56.8047.6646.094.7622.4744.58
0.75×53.8945.5143.795.1722.9927.49
1.00×52.3444.3542.865.4222.8115.70
1.25×52.1744.0242.355.5722.829.80
1.50×52.5044.0742.585.6623.077.44

Table 4: Navigation performance with respect to target speed in the noisy setting.

MethodSpeedSR ↑SPL ↑SNA ↑DTG ↓SWS ↑SWM ↑DSR ↓
DyGMap w/o dynamic0.00×39.3328.6126.607.9820.07–0.39
0.25×45.9932.6832.146.2822.8943.370.47
0.50×40.8329.5629.096.8220.2728.220.44
0.75×37.7527.8927.477.4719.2014.980.40
1.00×37.2027.1726.677.6620.217.970.40
1.25×35.9726.1825.807.9119.505.270.42
1.50×36.8426.5426.237.8419.793.960.48
DyGMap (Ours)0.00×40.9031.6629.547.9517.99–0.48
0.25×48.2635.9035.156.2421.7045.950.37
0.50×43.5733.0032.286.8619.4831.090.43
0.75×41.9932.4931.667.2119.6118.910.32
1.00×39.5230.2029.487.6519.779.840.30
1.25×38.6329.2928.647.9018.766.120.41
1.50×39.4029.7128.987.8020.064.060.30

The two tables show that the gains from dynamic goal-map updates vary with target speed. In the clean setting, as target speed increases from 0.00× to 1.50×, the SR gains are 2.10, 3.51, 1.52, 0.89, 0.73, 0.63, and 1.35 percentage points, respectively. The changes in SWM at speeds from 0.25× to 1.50× are 3.46, 1.48, 0.27, 0.86, −0.11, and 0.16 percentage points, respectively. Dynamic updates therefore provide the clearest benefit for slow-moving targets (0.25×) in the clean setting, with less pronounced gains at higher speeds.

In the noisy setting, the SR gains at the same seven speeds are 1.57, 2.27, 2.74, 4.24, 2.32, 2.66, and 2.56 percentage points, respectively. The SWM gains at speeds from 0.25× to 1.50× are 2.58, 2.87, 3.93, 1.87, 0.85, and 0.10 percentage points, respectively. Here, dynamic updates provide the greatest benefit at an intermediate target speed, with both SR and SWM gains peaking at 0.75×.

Furthermore, comparing each method's performance across target speeds reveals a consistent pattern across most metrics in both the clean and noisy settings. Compared with the stationary condition, low target speeds (0.25× and 0.50×) generally yield better navigation performance, while speeds above 0.75× generally yield worse performance. We hypothesize that these patterns reflect two competing effects of target motion: it can create additional opportunities to observe the target while also making the target harder to track.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.