acceptodds
Under review as a conference paper at ICLR 2027

EchoBelief: Evidence-Anchored Spatial Memory for Active Acoustic Mapping

Abstract

Echoes encode room geometry, and learned models can infer room layouts from room impulse responses (RIRs). However, one RIR mainly reveals nearby walls in direct view, so a robot mapping a whole room must decide where to move, when to measure and when to stop, as every measurement has a cost. Existing acoustic methods take responses as given, and active systems rely on vision; none chooses measurements from a persistent acoustic map. Such a map faces a dilemma: treating learned predictions as measurements erases observed walls, while discarding them loses spatial context. We propose EchoBelief, a system that resolves this dilemma with evidence-anchored spatial memory: an evidence map stores only structure that echoes or executed motion support, and a separate recurrent memory holds learned predictions. A geometry-supervised polar encoder turns each RIR into a map of wall reflections, and a belief-guided PPO policy reads the memory to move, sense or stop under an acquisition budget. We also introduce EchoRooms, a benchmark of 4,026 simulated rooms across 36 room types and five difficulty tiers, with 1.58M posed 12-channel RIRs on a dense 1 m grid, which we will release. On 355 test rooms, with known poses, EchoBelief reaches 90 % wall recall in 70.1 % of episodes, against 49.4–52.0 % for three tuned planners, while sensing half as often. The gain is largest in large rooms, where the planners rarely succeed (52.9% vs. at most 8.2%); in smaller rooms, the planners match or exceed it but sense far more. On 60 rooms with geometry disjoint from EchoRooms, evaluated only once, EchoBelief still outperforms raycast planning, and retraining without the memory pathway costs 22.8 success points.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Active acoustic mapping, active sensing, spatial memory, reinforcement learning, room impulse responses, room geometry reconstruction, robotic exploration

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_RRsLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a system for acoustic-based mapping, named EchoBelief, which uses echos to gradually create the physical map (walls) of the environment, and decide where to move to sense next. It stores map updates from actual observations separately from the estimated map state based on predictions to ensure that information actually observed is not forgotten or erased. A benchmark is created based on simulated data and the results of the proposed system show significant benefits compared to baselines.

Critical strengths

A new benchmark for acoustic mapping with a variety of layouts, from more regular to more challenging ones

A complete system that addresses the sensing, active exploration and map state estimation steps

Significant results in numbers of rooms mapped successfully under a limited sensing budget above a required accuracy threshold

Critical weaknesses
  • Not clear how the proposed and competing algorithms trade the % of successfully mapped environment vs the available time/sensing budget. It would be great to run the experiments with

  • A more detailed ablation study could be added to show where the main benefits come from. There are four steps (polar echo encoder, local geometry mapper, evidence anchored spatial memory, and belief guided exploration. An ablation study replacing each of these four steps with state of the art could show more clearly where most of the benefit comes from.

  • In the simulated data it would have been good to simulate noise from people, furniture etc to show performance a more realistic setting

Questions for authors

Is the main novelty in the overall system architecture or each individual step? It would be worth highlighting if some of the steps in the architecture could be replaced successfully with other state of the art methods, and what would be the overall impact on performance.

What would be the impact of people / furniture and so on on the performance? Is it possible to simulate rooms with more dynamic traffic and show the impact on the accuracy of the mapping?

A more detailed ablation study would be useful to show the impact of the novelty in each step wrt to existing competing algorithms for each step.

5 comments
AuthorsOpenReview ↗

Response (1/3)

Thank you for recognizing the value of EchoRooms, the complete acoustic mapping system, and the results under a limited sensing budget. Your questions concern how performance changes with the available budget, where the component benefits come from, and how noise and obstacles affect mapping. We address each with the comparisons below.

W1. Success versus sensing budget.

To make the tradeoff explicit, we compare EchoBelief with the three original planners and the strongest tested frontier-plus-raycast hybrid at six acquisition caps. The evaluation uses 355 test rooms with three starts each. Success means at least 90% wall recall in the final map. Throughout this response, differences are computed before rounding; brackets give paired room-bootstrap 95% confidence intervals conditional on the trained checkpoints.

Acquisition cap KEchoBelief success (%)Original planners (%)Hybrid (%)EchoBelief minus hybrid (points; 95% CI)
1629.316.0–17.719.4+9.9 [7.3, 12.5]
3243.527.2–33.136.5+6.9 [4.4, 9.6]
4846.435.8–39.545.2+1.2 [−1.0, 3.5]
6449.439.3–43.549.7−0.3 [−2.8, 2.3]
9665.444.7–48.458.7+6.8 [3.4, 10.1]
12870.149.4–52.069.1+1.0 [−2.7, 4.9]

EchoBelief improves success over the hybrid at 16, 32, and 96 acquisitions, while the differences at 48, 64, and 128 are not statistically resolved. Its success advantage over each of the three original planners is resolved at every tested cap.

The joint exploration and stopping strategy also reduces measurement use. At the 128-acquisition cap, EchoBelief uses 59.3 subsequent SENSE actions versus 122.9 for the hybrid, approximately 52% fewer, with no statistically resolved success difference. At the 64-acquisition cap, it uses 40.9 versus 61.9. These counts exclude the shared initial acquisition; the cap K includes it. Together, the results show both where EchoBelief maps more rooms successfully and where it uses fewer measurements.

For each cap, all methods have a decision limit of 8K. EchoBelief is evaluated separately at each cap. Planner trajectories are truncated at acquisition K or decision 8K because those planners do not read the remaining budget; at K = 64, this reproduces the independently evaluated success outcome in every episode. We will add this table and the recall-versus-acquisitions figure, with final maps carried forward after an episode ends. These evaluate sensing budgets and measurement use; physical execution time was not measured.

AuthorsOpenReview ↗

Response (2/3)

W2 / Q1 / Q3. Contribution and component replacements.

Our contribution is the complete acoustic mapping system, with the geometry-supervised polar echo field and evidence-anchored spatial memory as its central design contributions. The mapper and policy use established neural architectures. The polar field locates first-order wall reflections in range and bearing, providing a geometric signal for mapping. Sparse echoes still leave substantial structure unobserved, so the system also needs predictive context. Support-gated fusion protects the evidence map from unsupported prediction writes, while a separate recurrent memory retains that context for reconstruction and exploration.

To answer where the gains come from and whether components can be replaced successfully, we evaluated each stage through controlled comparisons: alternative encoder supervision, a geometric mapper replacement, a retrained memory ablation, and a hybrid exploration planner. The whole memory pathway produces the largest tested policy-ablation effect: removing it and retraining the belief and policy lowers success from 70.1% to 47.3%, a difference of 22.8 points [18.8, 27.0]. The table identifies the comparison and evaluation level for each stage.

StageComparison and evidenceContribution supported
Polar echo encoderIn a token-only local-mapping probe, near-wall F1 is 71.3% with polar supervision, 42.2% with masked reconstruction, and 35.2% with waveform reconstruction; polar minus masked: +29.1 points [28.4, 29.8].Geometric supervision makes the acoustic features more useful for local mapping. This compares training objectives.
Local geometry mapperReplacing the U-Net with a validation-tuned, non-learned echo-to-occupancy projection gives 72.4% versus 80.5% test near-wall F1 at 0.125 m; U-Net gain: +8.1 points [7.8, 8.4].The learned mapper improves local reconstruction over geometric projection. This is a local substitution.
Evidence-anchored memoryRemoving the whole memory pathway, with belief and policy retrained, gives 47.3% versus 70.1% success.Retaining predictive context alongside supported evidence benefits the complete system. The ablation measures the whole pathway.
Belief-guided explorationReplacing the learned explorer with the hybrid, using the same belief, gives 69.1% versus 70.1% full-budget success, with an unresolved difference; subsequent SENSE actions are 122.9 versus 59.3.The belief supports another successful exploration method, while EchoBelief's joint strategy reduces measurement use.

We also include an external acoustic reconstruction comparison using an adaptation of Nguyen et al. (IROS 2019) under EchoRooms' sensor interface. To isolate reconstruction from the chosen trajectory, we score identical acquisition trajectories with both estimators. Replacing its geometric reconstruction with our frozen belief and decoder raises success from 28.0% to 43.7% (+15.7 points [12.9, 18.7]) and wall F1 by 20.5 points [18.7, 22.4]. Thus, the reconstruction pipeline extracts more useful geometry from the same measurements.

Together, these comparisons identify the benefits of geometric supervision, learned local mapping, and retained predictive context. They also demonstrate a successful explorer replacement: the same belief supports the hybrid planner, while EchoBelief's joint strategy uses fewer measurements. We will add the four-stage table, distinguish local reconstruction tests from complete-system evaluations, and include the strongest hybrid in the main comparison.

AuthorsOpenReview ↗

Response (3/3)

W3 / Q2. People, furniture, and noise.

EchoBelief retains higher success than raycast IG across all 12 tested noise conditions. On 120 test rooms with three starts each, we add white, pink, and motor-like noise to the raw 12-channel RIRs at 40, 30, 20, and 10 dB. All models remain frozen, with no retraining. The success advantage ranges from 10.8 points [4.7, 16.9] to 22.5 points [15.0, 30.0].

Matched clean success is 71.9% for EchoBelief and 52.2% for raycast. At 40 dB, neither method has a statistically resolved success change. At 20 dB, EchoBelief loses 5–16 points and raycast 0–10; at 10 dB, success falls to 20–27% and 8–11%, respectively. Thus, EchoBelief retains an absolute success advantage, while strong noise still substantially degrades mapping. SNR is defined over the 100 ms input window, including the direct path; early-reflection SNR is about 13 dB lower.

For interior obstacles, EchoBelief retains higher structural-wall recall: 84.5% versus 75.3% (+9.3 points [0.6, 19.2]). All three obstacle-study conditions use the same 24 test rooms with one start each and live rendering at every acquisition. The added obstacles are static full-height boxes that reflect sound and block motion. Box sides range from 0.4 to 1.5 m, with one box per 25 m² of free floor and 1–30 boxes per room. The table scores the original structural walls.

ConditionEchoBelief success (%)Raycast IG success (%)
Clean live rendering75.079.2
Static reflecting obstacles50.050.0
Obstacles and intermittent interference33.333.3

Obstacles therefore reduce success for both methods. EchoBelief's reduction is 25.0 points [8.3, 41.7], and its collisions increase from 3.4 to 22.6 per episode. These frozen models were trained without the added obstacles. The higher structural-wall recall shows that EchoBelief recovers more of the room boundary in this test, while the success and collision results identify the remaining difficulty of exploring clutter.

Intermittent interference affects 25% of acquisitions, using babble-like noise or 5–30 ms broadband bursts at 10 dB. With obstacles and interference combined, structural-wall recall remains higher for EchoBelief by 11.0 points [2.1, 20.4], although both methods reach 33.3% success.

These tests address added acoustic noise and static reflecting obstacles. The boxes behave like pillars, and the interference models intermittent sound; ordinary three-dimensional furniture, moving people, and changing occlusion remain untested. We will state this scope explicitly and include both the retained performance and the degradation in the revised paper.

We hope the budget comparisons, component evidence, and noise and obstacle tests address your concerns and clarify the contribution. We will incorporate these results and explanations into the main text so that each comparison and its implications are easy to follow. Please let us know if any concern remains unresolved or would benefit from further clarification. In light of these clarifications and additional results, we would be grateful if you would consider raising your score.

Reviewer_gMz9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EchoBelief maps rooms from echoes using a speaker and 12-microphone ring, without cameras or lidar. Because each RIR reveals only local structure, the robot must choose where to move, when to sense, and when to stop. The method separates echo-supported evidence from learned predictions to prevent wall erasure. A PPO policy reads both memories and selects MOVE, SENSE, or STOP.

Critical strengths
  1. The task (camera-free, active acoustic mapping with an explicit decision about when to sense) is new as far as I know, and the motivation (smoke, dark, glass) is reasonable.

  2. The benchmark is a real contribution. Precomputed dense RIRs with geometric labels at three scales make this kind of RL work possible for others. The release plan is concrete.

  3. Experimental hygiene is very good. The paper tells us the test split was read four times, then builds a separate 60-room cohort, freezes everything, declares the checkpoint, and reads it once. I wish more papers did this.

  4. The controls in Table 36 address obvious concerns: giving raycast a stopping rule, giving it the policy's move set, removing acoustic input (success goes to zero), and live rendering at the true pose instead of the cache (69.0% vs 70.1%).

  5. The replay study (Table 19), where planner runs are replayed through beliefs with and without memory, is a good way to separate the effect of memory on map quality from its effect on decisions.

Critical weaknesses
  1. Simulation only, and an easy simulation. RIRs are noiseless, rooms are empty extruded 2D floor plans, there is no diffraction or transmission, microphones are ideal omnidirectional sensors, and poses are exact. Real echoes are much messier (furniture, people, robot ego-noise, speaker directivity, clutter, multipath from tables and chairs). The authors state this clearly, which I appreciate, but it still limits what we can conclude. Even a small real-world test, or at least adding measured noise, furniture-like scatterers, and pose drift during the full episode (not only in the local mapper study, where 5 cm / 2 degree noise costs 2 points), would make the paper much stronger.

  2. All baselines share EchoBelief's own belief and information contract. There is no learned external comparator, which the authors admit. The planners collapse in large rooms (4.3% to 8.2% success on rooms with at least 600 nodes), which makes me ask whether the planners are limited by a short candidate horizon or by the decision limit, rather than by the information they use. A planner that combines frontier-based global goals with local raycast information gain is a standard approach and might close much of the large-room gap. A PPO policy that sees the belief but with a simple feature set (no one-step predictor, no per-move features) would also help. The "reference recipe" arm is a partial version of this, but it bundles many changes together.

  3. Keeping observed occupancy separate from predicted or anticipated occupancy is common in visual exploration (for example occupancy anticipation work and MapEx-style uncertainty maps, which the paper cites). The new part is mostly applying this to acoustics, where predictions are much less reliable. That is fine, but the related work should say clearly how the two-store design differs from those systems.

Questions for authors
  1. Why do the planners fail so badly on large rooms? Is it the candidate radius, the decision limit, or path planning?

  2. What happens with pose noise during full episodes, not just in the local mapper?

  3. How does performance change with added sensor noise at realistic SNR?

  4. Can you show recall versus acquisitions curves for all methods?

  5. How long does training take end to end (encoder, mapper, belief, PPO), and on what hardware?

4 comments
AuthorsOpenReview ↗

Response (1/4)

Thank you for recognizing the value of EchoRooms, the experimental controls, and the crossed-replay study. Your questions concern the strength of the planning comparisons, the acoustic memory contribution, and performance under more challenging sensing conditions. We address these with the additional comparisons below.

Success means at least 90% structural-wall recall in the final map. Unless otherwise stated, brackets give paired room-bootstrap 95% confidence intervals; differences are computed before rounding and reported in percentage points. SENSE counts exclude the shared initial acquisition.

W2 / Q1. The suggested frontier–raycast hybrid.

We implemented your suggested combination of global frontier goals and local raycast information gain. Its switching threshold and goal rule were selected on validation rooms and frozen before evaluation on 355 test rooms with three starts each.

At the standard 128-acquisition / 1,024-decision cap, EchoBelief uses 59.3 SENSE actions versus 122.9, approximately 52% fewer. Success is 70.1% versus 69.1%, with no statistically resolved difference (+1.0 points [−2.7, 4.9]). EchoBelief also incurs fewer collisions (5.6 versus 20.3) and travels less (348 versus 661 m). These are whole-episode outcomes under the same caps; the methods choose different numbers of measurements.

The hybrid substantially improves on the original raycast planner's 49.4% success. EchoBelief retains a resolved advantage in the largest rooms, while the hybrid succeeds more often in the two intermediate-size groups:

Room sizeEchoBelief successHybrid successEchoBelief minus hybrid [95% CI]
100–400 nodes87.6%94.3%−6.7 [−10.1, −3.9]
400–600 nodes75.8%98.3%−22.5 [−33.3, −12.5]
≥600 nodes52.9%40.2%+12.7 [6.3, 19.4]

The comparison also depends on the available sensing budget. With acquisition cap K and decision cap 8K, the success differences are:

KEchoBelief minus hybrid [95% CI]
16+9.9 [7.3, 12.5]
32+6.9 [4.4, 9.6]
48+1.2 [−1.0, 3.5]
64−0.3 [−2.8, 2.3]
96+6.8 [3.4, 10.1]
128+1.0 [−2.7, 4.9]

Thus, EchoBelief improves success at 16, 32, and 96 acquisitions, and uses fewer measurements at every cap above 16. At K = 16, sensing counts are nearly equal (14.1 versus 14.2). The hybrid does not read the remaining budget, so its smaller-cap results use trajectory prefixes; a separate 64/512 run reproduces success in every episode. EchoBelief is evaluated separately at each cap because budget is a policy input.

We will include the hybrid in the main comparison and revise the abstract and conclusion to distinguish sensing efficiency, budget-dependent gains, and room-size effects.

AuthorsOpenReview ↗

Response (2/4)

Q1, continued. Why the original planners struggle in large rooms.

The additional controls support the importance of selecting global exploration goals. Increasing the raycast decision allowance from 1,024 to 4,096 changes overall success only from 49.4% to 49.5%. Increasing its candidate radius from 8 m to 16 m, 32 m, or the whole window, with the route horizon scaled as 2r and costs held fixed, gives 49.6%, 50.6%, and 51.3% success against a 49.5% reference. None of these radius comparisons has a paired interval excluding zero; success in ≥600-node rooms remains 5.5–7.1%.

Retuning costs and adding validation-selected 4 m sensing spacing raises success to 59.4%, including 20.2% in ≥600-node rooms. The frontier–raycast hybrid reaches 69.1% and 40.2%, respectively. A post-hoc trajectory analysis is consistent with this pattern: the original planners place 59–63% of new acquisitions within 2 m of an earlier one, versus 16.7% for EchoBelief. These controls support the importance of global goal selection within the tested settings. We will include the validation grids and reproduction checks with these comparisons.

W3. The acoustic memory contribution.

Our distinction is how acoustic evidence and predictive context are maintained and used. First-order echo estimates and executed motion gate writes to the evidence map, while a separate recurrent memory retains local predictions for reconstruction and exploration. This addresses the need to preserve supported structure while still using predictions where echoes are sparse. We will clarify this relationship to OccAnt's uncertainty-filtered occupancy aggregation and MapEx's observed maps and prediction-guided frontier scoring in related work; separating observations and predictions is established in visual exploration.

The crossed replay you highlighted tests reconstruction independently of action selection. On identical raycast trajectories, the primary belief, before support-gated fine-tuning, improves wall F1 at 64 acquisitions by 1.6 points [1.1, 2.1] over a geometric-only belief. Removing the whole memory pathway and retraining the belief and policy costs 22.8 success points [18.8, 27.0]. These results support retained predictive context and the complete memory pathway, respectively.

W2. Simpler learned features and external comparison.

The predictor ablation gives 68.8% success versus 70.1% for the full policy (+1.3 points [−1.7, 4.3]), so the benefit of that component alone is not statistically resolved. It retains per-move features; removing both feature sources, as you suggest, is a separate control that has not been completed. The shared-belief planner comparisons test exploration with the same reconstruction model. They do not supply an external learned comparator, which remains untested. We will distinguish these comparison types explicitly.

AuthorsOpenReview ↗

Response (3/4)

W1 / Q3. Sensor noise and clutter.

EchoBelief retains higher success than raycast IG across all 12 tested noise conditions. We add white, pink, and motor-like noise at 40, 30, 20, and 10 dB to the raw 12-channel RIRs before the frozen encoder, evaluating 120 test rooms with three starts each. Clean success is 71.9% for EchoBelief and 52.2% for raycast. The table reports EchoBelief / raycast success (%).

Noise30 dB20 dB10 dB
White72.5 / 50.666.9 / 52.220.3 / 9.4
Pink68.6 / 54.763.3 / 48.326.9 / 7.8
Motor-like68.6 / 55.855.8 / 41.926.9 / 11.4

At 40 dB, neither method has a statistically resolved success change. EchoBelief retains higher absolute success across the tested noise levels, although strong noise substantially degrades both methods.

SNR is measured over all channels and the first 100 ms, including the direct path; median early-reflection SNR is approximately 12–13 dB lower in the encoder and mapper analyses. These are synthetic sensitivity tests, without calibration to measured robot noise. We will include the SNR definition with the results so the nominal levels can be interpreted correctly.

For clutter, a separate live-rendering study compares clean and cluttered conditions on 24 rooms with matched starts. We add static full-height reflecting boxes. EchoBelief retains higher structural-wall recall, with a 9.3-point advantage [0.6, 19.2]. Structural-wall success falls from 75.0% to 50.0% for EchoBelief and from 79.2% to 50.0% for raycast. These boxes test added reflecting obstacles; they do not reproduce full furniture geometry or measured robot acoustics. All models remain frozen. We will include the retained recall advantage and the success degradation, and keep the claims within these evaluated conditions.

Q2. Pose noise throughout the episode.

We evaluated cumulative pose drift over full episodes on 120 rooms with three starts each. Each executed move of length ℓ adds independent translation noise with per-axis standard deviation kℓ and heading noise with standard deviation σθ√ℓ. The three levels are (k, σθ) = (1%, 0.5°/√m), (3%, 1°/√m), and (5%, 2°/√m). Estimated poses drive map registration and policy/planner inputs; true poses drive motion, collisions, and sensing legality. Maps are scored in the initial true frame without realignment or SLAM correction.

Drift levelEchoBelief successRaycast with policy move set
Clean71.9%52.5%
Low12.5%6.1%
Medium5.3%1.7%
High0.6%0.3%

Accumulated registration error substantially reduces success for both methods. Even at the lowest setting, EchoBelief's mean final error reaches 2.65 m / 6.5°. Its remaining success advantage is +6.4 points [2.2, 10.6], but its drop from clean is larger. High-drift results also involve frequent registration clamps; all episodes remain included. These results establish accurate localisation as an operating assumption, which we will state explicitly. The noise and drift comparisons use the raycast variants specified above, with the same acoustic front end and belief as EchoBelief.

AuthorsOpenReview ↗

Response (4/4)

Q4. Recall versus acquisitions.

We have consolidated the main comparison curves in one figure: EchoBelief, the three original planners, the policy-move raycast control, belief-frontier and random walkers, tuned raycast, the plain and quiet-stop hybrids, both Nguyen adaptations, and the reference checkpoint. The main panels retain all 355 rooms and the 400–600 and ≥600-node groups. The reference checkpoint, evaluated with STOP floor 128, appears in a separate matched 246-room panel, with the other trajectory methods restricted to the same room/start pairs.

The horizontal axis counts distinct acquired locations, including the start; repeated SENSE actions are excluded. At each k, we score the last decision with at most k distinct locations, carry the final map forward after termination, and retain every episode in the denominator. We also show the fraction reaching each count. This trajectory analysis complements the separate cap-K evaluations, which account for the budget supplied to EchoBelief's policy.

At k = 16/32/64/96/128, recall is 52.6/67.3/81.3/88.4/90.7% for EchoBelief and 52.4/65.7/78.8/85.6/89.6% for the hybrid. Tuned raycast leads in overall recall up to about 60 distinct locations; EchoBelief leads from about 64 onward. We will include these curves and their counting convention in the revised paper.

Q5. Training time and hardware.

The five final fitting runs took approximately 67.3 GPU-hours, each on one NVIDIA A40:

StageTime
Polar echo encoder2.8 h
Local geometry mapper21.8 h
Evidence-anchored memory, base fit5.2 h
Support-gated fine-tune2.2 h
PPO policy, 15 M-transition budget35.3 h

These durations cover the complete final fitting runs, including in-run validation. They exclude separate post-training calibration, cache construction, corpus rendering, and the broader development and ablation campaign. PPO reads precomputed RIRs and needs no online acoustic simulation. We will report this timing scope alongside the table.

We hope the stronger planning comparison, memory evidence, and noise, clutter, and drift evaluations address your concerns and clarify the contribution. We will revise the main text to make the method, comparisons, and operating assumptions easier to follow. Please let us know if any concern remains unresolved or would benefit from further clarification. In light of these clarifications and additional results, we would be grateful if you would consider raising your score.

Reviewer_C99kOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper addresses active acoustic mapping problem based on just acoustic support, no visual or lidar inputs, with memory to better perform room reconstruction, decide where to move, when to sense and where to stop under a measurement budget.

Critical strengths
  1. The motivation to incorporate spatial evidence memory is well.
  2. The work also introduces a useful EchoRooms benchmark for such a problem.
Critical weaknesses
  1. Missing baselines: There are no existing acoustic mapping baselines in the experiments. The paper majorly compares against raycast information gain and two planners that imagine an observation, along with frontier and random walkers (that are good for comparing within EchoBelief itself) but does not compare against other relevant acoustic mapping baselines that the paper introduced as a motivation itself. Furthermore, it’s closest competitor [1], also plans the path with microphones and a source to estimate the goemetry of room. It is not cited and it majorly weakens the claim in introduction that ‘No existing method explores this way.’ I believe missing the baselines for comparison that the work introduced as a motivation part itself does not fully support the scope of paper and leaves the comparison mostly against the work’s heuristics itself. Ref: [1] ‘Can a Robot Hear the Shape and Dimensions of a Room?’ this is an IROS 2019 paper that also explores acoustic mapping. The paper does not mention such relevant works.
  2. Different budget measurements might benefit EchoBelief: The paper’s different acquisition results, and tuned-stop raycast comparison address whether early stopping alone explain EchoBelief’s gain. However, the main baselines are compared under the same maximum budget rather than at matched actual numbers of measurements, and there is no rate-matched periodic-sensing control. I would like to see comparing methods at matched measurement counts, to clarify how much of the gain comes from sensing efficiency versus the rest of the method.
Questions for authors

Address the missing baseline and mentioned weakness.

2 comments
AuthorsOpenReview ↗

Response (1/2)

Thank you for recognizing the motivation for spatial evidence memory and the value of EchoRooms. Your comments identify two comparisons needed to assess the contribution: an existing acoustic mapping method and controls for the number and timing of measurements. We have added both, as detailed below.

Prior work and contribution. Thank you for highlighting Nguyen et al. (IROS 2019), which we will cite and discuss explicitly. Their work establishes acoustic-guided exploration through geometric wall estimation and verification, and we will revise our broad statement accordingly. EchoBelief learns a geometry-supervised polar echo field that locates first-order wall reflections in range and bearing, providing an explicit geometric signal for mapping. Sparse acoustic observations nevertheless leave substantial geometry unobserved, so exploration benefits from learned spatial predictions. Directly fusing unsupported predictions into the evidence map can erase observed walls. EchoBelief blocks such unsupported writes through support-gated fusion, while a separate recurrent memory retains predictive context. A learned policy reads both stores to guide movement, sensing, and stopping under an acquisition budget. EchoRooms enables evaluation of this complete process across diverse layouts.

Comparison with an existing acoustic mapping method. To address the missing baseline, we adapted Nguyen et al. to EchoRooms and evaluated it on the same 355 test rooms with three starts per room. We tested both its geometric reconstruction and a version with its path parameters retuned on validation rooms for EchoBelief reconstruction. This second comparison gives both systems the same final-map reconstruction model.

MethodSuccess (%)Mean SENSE actions
EchoBelief70.159.3
Nguyen adaptation with EchoBelief reconstruction (retuned path)46.990.3
Nguyen adaptation with geometric reconstruction28.0107.3

All methods have a cap of 128 acquisitions. Success means at least 90% wall recall; SENSE counts exclude the shared initial acquisition. Differences are computed before rounding; intervals are paired room-bootstrap 95% confidence intervals.

With final-map reconstruction shared, EchoBelief improves overall success by 23.3 points [18.7, 27.9], while using fewer measurements. The improvement also holds at the 64-acquisition cap: 11.3 points [8.2, 14.6]. The adapted controller still selects actions from its geometric estimates, so this comparison evaluates the acquisition strategies with a common reconstruction model.

To measure the reconstruction contribution separately, we also scored identical controller trajectories with both estimators. Replacing geometric reconstruction with EchoBelief's belief raises success from 28.0% to 43.7%, a gain of 15.7 points [12.9, 18.7], and wall F1 by 20.5 points [18.7, 22.4]. Thus, EchoBelief's belief and decoder produce a better reconstruction from the same acquired evidence. Together, these comparisons support the reconstruction pathway and the complete exploration system.

The comparison evaluates an adaptation to EchoRooms: its fixed array and 16 kHz sampling replace the original rotating, extendable microphones and 96 kHz sampling. The adaptation receives admissible measurement nodes and 400 ms of audio per acquisition, versus 100 ms for EchoBelief; validation selected local wall segments and disabled closure stopping. These results concern transfer to EchoRooms, rather than replication of the original apparatus. We will include the adaptation settings, rectangle implementation checks, and evaluation history with the comparison.

AuthorsOpenReview ↗

Response (2/2)

Comparison at matched actual measurement counts. To address your budget concern directly, we score the map immediately after exactly k acquisitions, counting the initial acquisition and every repeated SENSE. We compare EchoBelief with a stronger planner that combines global frontier goals and local raycast information gain. Both methods have a 4,096-decision allowance and collision-memory capacity, with STOP disabled. We retained the frozen EchoBelief trajectories under this protocol and reevaluated the hybrid with the matching allowance, keeping its belief and planner parameters fixed. All 1,065 episodes reach each reported count. The primary comparison is at 64 acquisitions, with 16 and 32 as secondary comparisons.

Exact acquisitions kSuccess difference, EchoBelief minus hybrid (points; 95% CI)Mean decisions, EchoBelief / hybrid
16+2.4 [0.5, 4.4]68 / 120
32+6.9 [4.7, 9.3]149 / 233
64+3.5 [1.9, 5.2]403 / 435

At the primary count of 64 acquisitions, EchoBelief improves success by 3.5 points [1.9, 5.2], wall recall by 3.1 points [2.3, 3.8], and wall F1 by 2.2 points [1.7, 2.8]. These gains hold with the same actual measurement count and decision allowance, directly addressing whether the improvement depends on unequal sensing budgets. This controlled evaluation disables early stopping and uses an extended decision allowance; the normal-operation comparison below measures the savings when stopping is enabled.

Under the standard 128-acquisition cap, EchoBelief reaches 70.1% success using 59.3 SENSE actions, compared with 69.1% success and 122.9 SENSE actions for the hybrid. Thus, EchoBelief uses approximately 52% fewer measurements, with no statistically resolved difference in success (+1.0 points [−2.7, 4.9]). This demonstrates the measurement savings of the joint exploration and stopping strategy. The exact-count comparison above separately evaluates reconstruction quality when both methods receive the same number of measurements.

Rate-matched periodic sensing. To address your concern about the missing periodic-sensing control, we trained a controller with the same belief, reward, training budget, and checkpoint-selection rule. It learns movement and stopping and acquires a measurement after every six MOVE decisions, deferring acquisition when sensing is illegal. The period was fixed before training from the validation sensing rate, and each acquisition consumes a decision. The achieved rates are closely matched: 6.03 moves per sense for the periodic controller and 5.98 for EchoBelief.

In the rate-matched comparison, EchoBelief uses 59.3 versus 64.9 SENSE actions (difference −5.6 [−7.3, −4.0]) and incurs 5.7 versus 13.3 collisions, with no statistically resolved difference in success (+1.2 points [−1.8, 4.2]). These results demonstrate the efficiency benefit of the learned joint exploration and stopping strategy: fewer measurements and fewer collisions.

Presentation and AI-use clarification. We will organize the main-text explanation around the evidence map, predictive memory, and exploration policy, add a short glossary, and move development labels and secondary experiments to the appendix. As disclosed in the submission, LLM assistance was used for manuscript drafting and editing, experiment/figure/table scripts, and log summaries. The authors executed the code, checked outputs against logged artifacts, and take responsibility for the results. The deployed acoustic mapper and PPO policy do not use an LLM. We hope this clarification addresses the AI-use concern and would appreciate reconsideration of the ethics flag if it does.

We hope the added acoustic baseline, the reconstruction comparison on identical trajectories, and the sensing controls address your concerns and clarify the contribution. We will also revise the main text throughout to make the method and experimental findings easier to follow. Please let us know if any concern remains unresolved or would benefit from further clarification. In light of these clarifications and additional results, we would be grateful if you would consider raising your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.