{"id":"64545a1b-4b54-49ad-82fb-dacd6ba3c8d1","arxiv_id":"2412.09878","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A six-microphone array on a robot arm, combined with a learned audio and motion model, localizes contact points on the arm to within 0.4 to 2.2 cm, including on novel objects and human strikes.","lead":"SonicBoom is a pipe-shaped robot arm attachment with six contact microphones that learns where it was hit from the sound of the collision. It maps vibrations to a contact point within a few centimeters, even for objects and motions never seen in training, which could help robots navigate occluded farms.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated mesh-intersection labels in Sec. V-B make the 0.43 cm in-distribution MED potentially a measure of label bias rather than true localization accuracy.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing issue: the ground-truth label pipeline in Sec. V-B is unvalidated. I agree that this is the most serious threat to the central claim, and I do not see a more fundamental flaw that would overturn the system's overall plausibility. The paper otherwise presents a coherent learning-based approach with a large dataset, sensible ablations, and a real-robot demonstration. The robot-stationary human-strike experiment with visually marked ground truth provides independent support for cm-level accuracy in a zero-shot setting, which is genuine evidence that the acoustic features carry spatial information. However, the abstract's headline 0.43 cm in-distribution number is measured against the same mesh-intersection heuristic used to generate training labels; if that heuristic is systematically biased, the model can score artificially well on Test Set 1 without localizing physical contacts any better. My proposed check directly measures label bias and would settle whether the sub-centimeter claim is meaningful. Until that check is performed, a conditional verdict is appropriate, and no change to the reader's verdict is needed.","tokens_in":10494,"tokens_out":5324,"duration_ms":64806,"concrete_test":"Re-run the evaluation on a held-out set of roughly 100 strikes while independently recording the true contact point using two calibrated high-speed cameras observing the rod and end-effector, or an instrumented impact sensor. Compare the Sec. V-B mesh-derived labels to these independent measurements. If the mean or median label discrepancy is below 0.5 cm, the 0.43 cm in-distribution result is supported. If the discrepancy is comparable to or larger than 0.43 cm, the reported in-distribution error is dominated by label bias and the central claim needs to be restated in terms of true physical contact accuracy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of sub-centimeter in-distribution localization depends entirely on the ground-truth labels produced in Sec. V-B. Labels are generated by loading robot joint states at collision time into a mesh representation, finding near-zero mesh intersection points, averaging them, and projecting the average onto the SonicBoom surface. No independent optical, tactile, or other measurement validates this label pipeline. If the robot penetrates the object before the controller stops, the centroid of the intersection volume is not generally the first physical contact point, and the projection step can introduce systematic shifts along z or theta. A supervised model can learn to reproduce this systematic label bias, so an MED of 0.43 cm on Test Set 1 may reflect consistency with a biased label generator rather than true physical localization precision. Test Set 1 is also described as the validation set used during model development, which further inflates the headline number. The 2.22 cm robot-stationary result is less vulnerable because ground truth was provided by visually marked surface positions, but that does not validate the sub-centimeter in-distribution claim. The concern is not that the model is wrong, but that the measurement standard for its headline accuracy is unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents SonicBoom, a hardware and learning pipeline that localizes impulsive single-point contacts on a PVC end-effector link using an array of six piezoelectric contact microphones and robot proprioception. The system maps mel spectrograms, GCC-PHAT features, and one-second end-effector trajectories to cylindrical coordinates (z, θ) via a multimodal transformer. The authors report mean Euclidean distances of 0.43 cm on in-distribution data, 1.01 cm on novel rod geometries, 2.01 cm in robot-active haptic mapping, and 2.22 cm when a human strikes the stationary robot. They also provide extensive ablations of input modalities, preprocessing, and augmentation, plus real-world demonstrations in a mock canopy.","tokens_in":10751,"tokens_out":3222,"duration_ms":34460,"significance":"If the reported accuracy is genuine, this is a practical contribution to tactile sensing for occluded manipulation: it shows that a small array of cheap contact microphones can turn a rigid robot link into a contact-aware surface, with generalization to novel objects, materials, and human strikes. The strengths of the paper include a large autonomously collected dataset (18,000 events), open-sourced hardware and software, thorough feature and ablation comparisons, and real-robot demonstrations. The phase-based features (GCC-PHAT) are shown to be particularly important for out-of-distribution generalization, which is a useful empirical finding. However, the headline sub-centimeter accuracy depends entirely on an unvalidated label-generation procedure, and the paper lacks baselines against simpler analytical or classical methods, so the contribution of the learning pipeline over conventional TDOA approaches is not yet established.","major_comments":[{"comment":"The ground-truth contact labels are generated by averaging near-zero mesh-intersection points and projecting the average onto the SonicBoom surface, but this procedure is never validated against an independent measurement such as an optical marker, a tactile sensor, or manual annotation. The centroid of a mesh-intersection volume is not generally the first physical contact point, and the projection step can systematically shift labels along z or θ. Since the reported MEDs are computed relative to these labels, the 0.43 cm in-distribution error on Test Set 1 may partly reflect consistency with a biased label generator rather than true physical localization accuracy. The authors should validate the label pipeline on a subset of collisions with independent ground truth and report the resulting label error, and ideally also report localization error against that independent ground truth.","section":"V-B"},{"comment":"Test Set 1 is explicitly described as the validation set used during model development. Reporting it as the headline 'in-distribution' performance is therefore not a measure of held-out generalization: the model architecture and hyperparameters were selected using this set, so the 0.43 cm MED may be optimistically biased. The authors should evaluate on a separate held-out test set collected under the same conditions as the training distribution, or at least clearly label Test Set 1 as validation and move the headline number to a true test set.","section":"VII-A"},{"comment":"The paper claims 'state-of-the-art contact localization' (Sec. II-B) but provides no comparison to any baseline, such as analytical TDOA triangulation from GCC-PHAT peaks, a simpler classical model, or a non-transformer architecture. Without such a baseline, it is unclear how much of the performance comes from the learned model versus the information already present in the microphone array. Additionally, all reported MEDs appear to come from a single trained model with no error bars or multiple-seed runs, so the statistical significance of differences between ablations (e.g., Fig. 7) is not established. The authors should add at least one classical TDOA baseline on the same test sets and report mean and variance over multiple training runs.","section":"VII-B"}],"minor_comments":[{"comment":"The abstract reports a 0.42 cm in-distribution error, while the body text (Sec. VII-A and conclusion) reports 0.43 cm; this inconsistency should be corrected.","section":"Abstract and Introduction"},{"comment":"The robot-stationary human-strike experiment reports a 2.4 cm mean error in Sec. VII-C2, while Test Set 4 in Sec. VII-A reports 2.22 cm for the same scenario; the discrepancy should be clarified.","section":"VII-C2"},{"comment":"The term 'zero-shot transfer' for Test Set 4 is ambiguous: the robot is stationary, so proprioceptive data are uninformative but still available to models that use them. The paper should clarify whether the test uses the audio-only model or feeds a stationary trajectory to the full model, and should avoid implying that the proprioceptive input channel is absent.","section":"VII-A"},{"comment":"The description of GCC-PHAT features says 'full representation of GCC-PHAT vector' is used, but the dimensionality and time-lag range of the vector fed to the MLP are not specified, making the architecture hard to reproduce.","section":"VI-B"},{"comment":"Equations (3) and (4) are duplicated and the second is typeset incorrectly; one equation with the circular-loss substitution for θ would suffice.","section":"VI-C"},{"comment":"The preprocessing ablation is reported only on Test Set 1 (the validation set), which limits the conclusions about generalization; stating this explicitly and reporting on at least one out-of-distribution set would strengthen the claim.","section":"VII-B2"}],"recommendation":"major_revision","confidential_remarks":"The central concern is the unvalidated label pipeline, which the stress-test note correctly identifies. This is addressable with additional experiments rather than being a fundamental flaw, so I recommend major revision. I also noticed that the 'first work' claim for microphone-array contact localization on manipulators may need a more careful literature check, since related acoustic tomography and contact-estimation works exist in closely adjacent venues. The paper is otherwise well-executed and the open-sourcing is a plus."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid systems paper that probably does what it claims at the few-centimeter level, but the sub-centimeter headline number in the abstract is not something I'd trust yet. The 0.43 cm in-distribution result depends entirely on a label-generation heuristic (Sec. V-B) that averages near-zero mesh-intersection points and projects the average onto the cylinder. There is no independent check against vision or tactile ground truth. If the robot penetrates the object before stopping, the centroid of the intersection volume is not the first contact point, and the projection can shift the label along z or theta. A supervised model can learn that bias, so a low MED on Test Set 1 may just mean the model is consistent with the label generator.\n\nThat said, the paper is not resting on that one number. Test Set 4, where a human strikes the stationary end-effector and ground truth comes from visually marked positions, shows 2.2 cm error in a zero-shot transfer. That result is much less vulnerable to the label problem, and it tells me the method genuinely localizes contacts to within a couple centimeters. The overall claim, that an array of contact microphones plus learned features can localize single-point impulsive contacts on a curved link, is credible and worth building on.\n\nWhat is actually new: prior work mostly did coarse region classification or flat-surface analytical triangulation. This is the first combination I know of that puts an array of contact mics on a cylindrical end-effector and learns a continuous mapping to (z, theta), with explicit generalization tests to novel objects, exploratory motions, and a different striking agent. The ablation study is thoughtful and honestly reported—they show proprioception helps angle but not height, and that phase features carry the generalization. The dataset is large (18k events, 108k audio files) and the platform is open-sourced. That is real evidence.\n\nThe weak spots, in order of importance: (1) label validation is missing and should be added, even a small subset with visual markings on the cylinder would settle the bias question; (2) no baselines—at minimum an analytical TDOA triangulation and a single-microphone model would make the learning and array claims concrete; (3) no error bars or multiple seeds anywhere, which matters because the headline differences (0.43 vs 1.01 etc.) are small; (4) Test Set 1 is also the validation set, so reporting it as the flagship in-distribution number is double-dipping, even if common.\n\nMinor: the abstract says 0.42 cm and the body says 0.43 cm. Not a big deal but sloppy.\n\nThis deserves a serious referee. The contribution is real, the experiments are well-designed for the questions asked, and the missing pieces are fixable. I'd send it to review with a request for revisions, not desk-reject it.","headline":"The paper's few-centimeter localization claim is credible, but the headline 0.43 cm in-distribution number rests on unvalidated mesh-intersection labels and should not be taken at face value.","tokens_in":11288,"tokens_out":2955,"would_cite":false,"duration_ms":28803,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Six microphones on a robot arm localize a bump to under half a centimeter by listening to vibrations in the arm itself.","keywords":["contact localization","acoustic tactile sensing","microphone array","GCC-PHAT","haptic mapping","agricultural robotics","transformer sensor fusion"],"falsifier":"Strike a set of physically marked, known positions on the stationary tube while tracking a tool or target with an optical motion-capture system, then compare SonicBoom's predictions with the optical positions. If the two ground-truth sources disagree by about the claimed 0.43 cm error, the mesh-averaging label pipeline cannot be treated as accurate enough to certify sub-centimeter localization.","tokens_in":10285,"feed_emoji":"🎤","tokens_out":7415,"duration_ms":77754,"temperature":0.7,"pith_summary":"The paper tries to establish that a robot can localize where it bumps into a rigid object by listening to the vibration traveling through its own end-effector link. It builds a PVC tube lined with six contact microphones, collects 18,000 real collision sounds with an automated robot, and trains a transformer to map audio and motion into a contact point on the tube's surface. The motivating use case is agriculture, where foliage hides branches and trellis wires from cameras, and a robot that can feel where it bumped could navigate and map these occluded structures. If the claims hold, the same hardware trick could turn ordinary robot links into large-area touch sensors at low cost.","feed_headline":"Robot arm localizes bumps to 0.43 cm with six microphones","feed_subtitle":"Acoustic sensing maps collision points on a PVC end-effector, so robots can feel through occluded foliage.","key_machinery":"The central object is SonicBoom's sensing geometry: six piezoelectric contact microphones arranged in two rings of three on a split PVC tube, which turns the tube into an instrumented surface parameterized by cylinder coordinates. The argument is carried by relative features: per-channel mel spectrograms encode local intensity patterns, while GCC-PHAT (a phase-normalized cross-correlation between each microphone pair) encodes arrival-time differences without requiring gain calibration; robot proprioception supplies a motion-direction prior. A multi-modal transformer with modality-specific encoders fuses these into a single $(z, \\theta)$ prediction, and the azimuth loss is computed in Cartesian form to avoid angle-wrapping discontinuities.","core_discovery":"The central claim is that relative acoustic features between microphones let a supervised model localize single-point impulsive contacts on a robot's PVC end-effector tube in cylindrical coordinates (height $z$ and azimuth $\\theta$). Using six channels of mel spectrogram and 15 pairwise GCC-PHAT vectors fused with robot proprioception in a multi-sensory transformer, the model reports mean Euclidean errors of 0.43 cm on in-distribution data and remains under 2.3 cm when tested on novel objects, exploratory motions, and human strikes, including stationary-robot zero-shot transfer. The paper further claims that phase information, not magnitude alone, is what generalizes to out-of-distribution contacts.","pith_inferences":["The same hardware should extend to longer links by adding more microphone rings, as the paper notes, which suggests the per-ring triangulation scheme scales linearly with tube length.","Because phase features are what generalize, an audio-only variant could be mounted on existing rigid links without robot proprioception, such as tools or grippers, at the cost of the larger errors seen in the stationary test.","Applying temporal filtering or a probabilistic tracker over consecutive predictions could lift the single-impact assumption and track sliding contacts, a natural next step the paper lists as future work."],"forward_implications":["A robot arm fitted with this microphone array can localize a single impact to under half a centimeter on the training distribution, so the vibration signal itself is spatially informative.","The error grows to about 1.0 cm on unseen rod and branch-like geometries, suggesting the learned mapping generalizes beyond the specific training objects rather than memorizing them.","In an occluded mock canopy, haptic mapping reconstructs branch positions with roughly 2 cm error, which is precise enough to support obstacle avoidance when vision is blocked.","With the robot stationary, audio alone reaches about 2.2 cm error, so the method does not depend on knowing the robot's motion to locate a contact."],"supporting_citations":[{"why":"Defines the physical difficulty the paper targets: vibration through non-uniform structures causes scattering and mode conversions, motivating learning over analytical models.","marker":"[13]"},{"why":"Supplies the triangulation principle behind the two three-microphone rings used to parameterize the tube surface.","marker":"[25]"},{"why":"Supports treating tree branches as cylinders, which justifies the single-point, impulsive-contact assumption.","marker":"[26]"},{"why":"Provides precedent for using the full GCC-PHAT vector rather than extracted time delays, the phase feature SonicBoom adopts.","marker":"[27]"},{"why":"Contributes the multi-sensory transformer fusion design that SonicBoom adapts to combine audio and proprioceptive encoders.","marker":"[28]"},{"why":"Demonstrates learning robot manipulation from in-the-wild audio-visual data, supporting the transformer-based fusion of acoustic and motion modalities.","marker":"[29]"}],"fun_headline_variants":["Robot arm 'hears' collisions to 0.43 cm accuracy","Acoustic sensing lets robots feel through foliage","Sound-based contact localization: 0.43 cm error","Hear your bumps: microphones map robot arm contacts","Feel occlusion? Robot arm uses sound to sense touch"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracy assumes the automatically generated ground-truth contact labels—averaged mesh-intersection points projected onto the tube—are more accurate than the predictions, but they are never checked against an independent measurement.","fun_headline_variants_meta":{"raw":{"variants":["Robot arm 'hears' collisions to 0.43 cm accuracy","Acoustic sensing lets robots feel through foliage","Sound-based contact localization: 0.43 cm error","Hear your bumps: microphones map robot arm contacts","Feel occlusion? Robot arm uses sound to sense touch"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000369,"raw_usage":{"total_tokens":1940,"prompt_tokens":865,"completion_tokens":1075,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":995}},"tokens_in":481,"tokens_out":1075,"duration_ms":10838,"temperature":1.0,"reasoning_tokens":995,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:38:16.645981+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Strike a set of physically marked, known positions on the stationary tube while tracking a tool or target with an optical motion-capture system, then compare SonicBoom's predictions with the optical positions. If the two ground-truth sources disagree by about the claimed 0.43 cm error, the mesh-averaging label pipeline cannot be treated as accurate enough to certify sub-centimeter localization.","supporting_citations":[{"cited_title":"A biomimetic elastomeric robot skin using electrical impedance and acoustic tomogra- phy for tactile sensing,","cited_arxiv_id":null,"evidence_quote":"Defines the physical difficulty the paper targets: vibration through non-uniform structures causes scattering and mode conversions, motivating learning over analytical models."},{"cited_title":"A survey of sound source local- ization and detection methods and their applications,","cited_arxiv_id":null,"evidence_quote":"Supplies the triangulation principle behind the two three-microphone rings used to parameterize the tube surface."},{"cited_title":"Towards robotic tree manipulation: Leveraging graph representations,","cited_arxiv_id":null,"evidence_quote":"Supports treating tree branches as cylinders, which justifies the single-point, impulsive-contact assumption."},{"cited_title":"Deep neural networks for multiple speaker detection and localization,","cited_arxiv_id":null,"evidence_quote":"Provides precedent for using the full GCC-PHAT vector rather than extracted time delays, the phase feature SonicBoom adopts."}],"review_version":1}