{"id":"9e134314-7419-414b-84e2-077039f02ae4","arxiv_id":"2508.12071","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"OASIS combines real-time sonar voxel carving with projected optical images to reconstruct static underwater workspaces from an eye-in-hand manipulator.","lead":"This paper describes a system that combines sonar and camera data on a robotic arm to build a 3D map of an underwater workspace in real time. It is aimed at helping underwater robots and pilots see obstacles during manipulation tasks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The evidence never tests the claimed system: Table II times only the sonar voxel update, and Table III reports single-modality errors, so 'real-time quantitatively accurate opti-acoustic fusion' is not demonstrated.","rationale":"I read the paper as claiming a practical fusion pipeline rather than a new physical model. The strongest independent support is the released code/dataset and the explicit tank experiments, and the pose-accuracy limitation is honestly stated in Section V. My concern is not internal inconsistency in the sensing physics but a mismatch between the claim and the evidence. The reader's 'weakest_assumption' on pose accuracy is legitimate but secondary: even with perfect poses, the reported numbers do not measure the fused system. I also flag an omitted-step problem in Algorithm 1: the pseudocode applies the Gocc/Gobs ratio threshold in lines 7-13 without any statements that increment those counters, so as printed it cannot accumulate occupancy across frames; the text in Section III-B Step 4 describes the missing update. This is likely a pseudocode/implementation mismatch and can be checked against the released code, but it should be corrected in revision. The central fix is an end-to-end timing/fused-accuracy benchmark. Because the reader already gave a conditional verdict, I recommend no change to that verdict.","tokens_in":9611,"tokens_out":8037,"duration_ms":91736,"concrete_test":"Run the released pipeline end-to-end on the accompanying dataset at 0.05 m voxels with per-stage timers, and measure object dimensions from the fused color mesh (tank wall, milk crate, mesh, chain) against the same ground truth. Report whether the optical fusion latency stays under 100 ms per frame and whether fused errors are at or below the best single-sensor errors from Table III. If either fails, the title claim should be narrowed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that OASIS fuses sonar and optical data in real time with quantitative accuracy. The measurements cited in support stop short of that claim. Table II reports only the acoustic voxel-update time (0.052 s/frame at 0.05 m voxels); the optical path—Open3D meshing, marching cubes, background segmentation with Rembg/ISNet, watershed masking, virtual-camera depth rendering, and pixel projection—is never timed. It is possible that the fused overlay runs materially slower than 10 Hz and is effectively a one-time texture pass, which would not support 'real-time opti-acoustic fusion.' Similarly, Table III lists acoustic-only and optical-only dimension errors; no error for the fused reconstruction is reported. The paper therefore establishes real-time sonar voxel carving plus qualitative optical texturing, but not a quantitatively accurate real-time fusion. Section V's pose-accuracy caveat is acknowledged and is not the limiting issue here.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces OASIS, an opti-acoustic fusion method for real-time 3D reconstruction in underwater manipulation workspaces. The system uses a fixed-base manipulator with an eye-in-hand imaging sonar and optical camera. Sonar frames are processed into a 3D voxel grid via a voxel-carving scheme (Algorithm 1) that extends the authors' prior work [22], with a custom preprocessing step (Algorithm 2) for deringing and intensity normalization. Optical images are then projected onto the reconstructed mesh using a virtual-camera depth render and background segmentation to provide a textured overlay. The method is validated in a tank with a 2.1 m diameter workspace, reporting sonar processing times per frame for several voxel sizes and per-object dimension errors for the acoustic and optical modalities separately. The authors also release code and a dataset.","tokens_in":9795,"tokens_out":3121,"duration_ms":35094,"significance":"If the central claims were fully substantiated, OASIS would be a practical contribution: it demonstrates that volumetric sonar reconstruction can run at interactive rates for a manipulator-mounted sensor, and the eye-in-hand sweep trajectory is a sensible way to gather multi-view data with minimal arm motion. The release of code, dataset, and visualization tools is commendable and will benefit the community. However, the evidence as presented supports only 'real-time sonar voxel carving with qualitative optical texturing,' not 'real-time quantitatively accurate opti-acoustic fusion.' The missing timing and fused-accuracy measurements are the decisive gap between what is claimed and what is shown.","major_comments":[{"comment":"The real-time claim is based solely on the acoustic voxel-update time: Table II reports 0.052 s per sonar frame at 0.05 m resolution (18.6 FPS). The optical fusion pipeline described in Section III-C—Open3D meshing, marching cubes smoothing, Rembg/ISNet background removal, watershed depth masking, virtual-camera depth rendering, and pixel projection—is never timed. Consequently, the abstract's statement that OASIS achieves 'real-time 3D reconstruction' and Table I's entry 'OASIS Real-time (18 Hz)' are not supported for the fused system; 18 Hz is the sonar-only update rate. The authors should either measure and report the end-to-end per-frame latency of the full pipeline, or revise the claims to specify that real-time applies only to the acoustic mapping stage and that optical fusion is an on-demand overlay step.","section":"Section IV, Table II"},{"comment":"The quantitative evaluation does not test the fused output. Table III reports acoustic-only dimension errors (tank, mesh full width) and optical-only errors (milk crate, mesh 10 cells, chain), but no dimension error is given for the opti-acoustic fused reconstruction. Thus the contribution claim of a 'quantitatively accurate 3D reconstruction' enabled by fusion is not directly demonstrated. The authors should provide a quantitative metric for the fused result—for example, measured dimensions or a point-to-mesh distance of the final textured model against ground truth—or explicitly state that quantitative accuracy is only claimed for the individual modalities.","section":"Section IV, Table III"},{"comment":"The conclusion states that OASIS 'integrates voxel carving and Gaussian splatting techniques for 3D reconstruction.' The method section and Algorithm 1 use voxel carving only; Gaussian splatting is discussed only in the related work and is not part of the proposed pipeline. This is a factual mischaracterization of the method and should be corrected.","section":"Section VI, Conclusion"}],"minor_comments":[{"comment":"The phrase 'real-time 3D reconstruction unstructured underwater workspaces' is missing a preposition; it should read 'real-time 3D reconstruction in unstructured underwater workspaces.'","section":"Abstract"},{"comment":"The occupancy threshold t_r is said to be 'determined empirically based on the data's false negative rate,' but no details are given about how it was chosen or how sensitive the reported accuracy and real-time performance are to its value. A brief sensitivity statement would strengthen the reproducibility of the method.","section":"Section III-B, Step 4"},{"comment":"The phrase 'rendered alongside the meshed voxel grid' is vague. It would be clearer to state explicitly whether the optical pixels are texture-mapped onto the mesh surface, stored as a colored point cloud, or composited as an overlay in a separate layer.","section":"Section III-C"},{"comment":"The entry 'OASIS Real-time (18 Hz)' under 'Optimization Time' is misleading because the 18 Hz figure in Table II covers only the sonar voxel update. Add a footnote or change the entry to indicate that the rate is for the acoustic component only.","section":"Table I"},{"comment":"The text refers to side-by-side reconstruction results and to objects (a,b), (c,g), etc., but does not describe the layout of Figure 6 or point to specific subfigures in the evaluation. Adding a sentence that guides the reader through the figure would improve clarity.","section":"Section IV, Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an integration of existing sonar voxel-carving ideas with a practical trajectory and a public benchmark, and the authors are transparent about the fixed-base pose assumption and static-workspace limitation in Section V. The two main gaps—the absence of end-to-end timing for the optical fusion path and the absence of a quantitative fused-reconstruction error—are directly fixable with additional experiments and are not fundamental flaws. If the authors supply those measurements or appropriately scope their claims, the paper could become acceptable. I saw no citation or novelty-disclosure concerns; the self-citation to [22] is appropriate background."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"OASIS does something the cited literature doesn't: it takes the authors' earlier min-max sonar voxel carving, makes it fast enough to keep up with the sonar, and overlays optical texture using an eye-in-hand sweep that moves the arm minimally. That combination is new. The paper also does a few things well: the preprocessing for deringing sonar is clearly explained, the 0.052 s/frame acoustic update at 5 cm voxels is a concrete, reproducible timing, and the code/data are released. The self-citation to [22] is legitimate background, not a red flag.\n\nThe soft spots are exactly what the reader's report and stress-test flag. The headline claims real-time opti-acoustic fusion, but Table II only times the acoustic voxel update. The optical side—Open3D meshing, marching cubes, Rembg segmentation, watershed masking, and pixel projection—is never timed, so we don't know whether the full loop runs at 10 Hz or whether the optical pass is a one-time texture drag. Similarly, Table III reports acoustic errors and optical errors separately, but never an error for the fused reconstruction. That means the paper demonstrates real-time sonar carving plus qualitative optical texturing, not a quantitatively accurate real-time fusion. These are fixable: an end-to-end timing budget and a fused-versus-single-sensor accuracy comparison would close the gap.\n\nA smaller issue: the conclusion says the method integrates \"Gaussian splatting techniques,\" but no part of the method uses 3DGS. That looks like a leftover line from related-work drafting and should be corrected. The pose-accuracy limitation is properly acknowledged in Section V, and for the stated small-static-workspace scope it is a fair limitation, not a hidden flaw.\n\nIs the central idea sound? Yes. The components are established, the integration is sensible, and the experiments, while small (a 2.1 m tank, four objects), are the right first test. The paper would benefit from a serious referee; I would send it out rather than desk reject. My recommendation: conditional acceptance after the authors add an end-to-end timing analysis, report a fused reconstruction error, and fix the Gaussian splatting sentence. I'd cite this if I were working on underwater manipulation perception, and it's a reasonable reading-group pick for robotics folks interested in sensor fusion.","headline":"OASIS is a genuinely useful integration of real-time sonar voxel carving with optical texturing for underwater manipulation, but the paper's central claim outruns its evidence: the fused pipeline is never timed as a whole and Table III reports no fused accuracy.","tokens_in":10316,"tokens_out":1220,"would_cite":true,"duration_ms":15288,"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":"OASIS combines a wrist-mounted sonar and camera to reconstruct underwater workspaces in real time, reporting 18.6 FPS sonar processing and centimetre-level object accuracy in tank tests.","keywords":["opti-acoustic fusion","voxel carving","underwater 3D reconstruction","real-time perception","eye-in-hand sensing","sonar preprocessing","underwater manipulation","occupancy grid"],"falsifier":"Run the same sweep trajectory twice with a rigid target fixed in the tank and the arm returned to its nominal home pose between runs; if the reconstructed target position shifts by more than a few centimetres, or drifts relative to a laser-measured reference, the fixed-kinematics pose assumption is falsified. A sharper test is to apply a known sagging load to the wrist and check whether the acoustic grid and the optical overlay separate by more than the 5 cm cube size.","tokens_in":9405,"feed_emoji":"🤿","tokens_out":10071,"duration_ms":94594,"temperature":0.7,"pith_summary":"OASIS sets out to show that an underwater manipulator can build a usable 3D picture of its workspace while it works, rather than after offline computation. The method fuses a wrist-mounted multibeam imaging sonar with an optical camera: the sonar carves occupied space into a grid of 5 cm cubes from intersecting views, and close-up optical images are then projected onto the meshed grid to add texture and object identity. In tank experiments the paper reports per-frame sonar processing of 0.052 seconds (18.6 FPS) at 0.05 m resolution, and reconstructed object dimensions within 1.2 to 10.6 cm of ground-truth measurements. The practical point is that real-time opti-acoustic fusion could let a pilot or an autonomous controller see which regions are safe to touch and which objects are present, without long optimization times or bright artificial lighting.","feed_headline":"Sonar-camera fusion maps underwater workspaces in real time","feed_subtitle":"Voxel carving runs at 18.6 frames per second and keeps object errors within centimetres for manipulator tasks.","key_machinery":"The load-bearing mechanism is a precomputed voxel template combined with ratio voting. The template holds the minimum set of grid cubes needed to represent the sonar's 130-degree horizontal and 20-degree vertical fields of view, so each incoming ping is projected into the template instead of casting every beam and range bin into the world frame, which is what brings per-frame cost down to tens of milliseconds. Each template cube is marked occupied or empty after a statistical deringing step, then transformed into the world frame using the manipulator's forward kinematics. Two count grids accumulate total observations and occupied observations, and a cube survives as occupied only when the ratio $G_{\\mathrm{occ}}/G_{\\mathrm{obs}}$ exceeds an empirical threshold $t_r$; this voting across intersecting views is what suppresses the elevation ambiguity. For the optical layer, the voxel grid is converted to a mesh, a virtual camera renders a depth image from each optical frame, background segmentation is applied to both optical and depth images, and the segmented pixels are back-projected onto the mesh using the same kinematic pose estimate.","core_discovery":"The central discovery is that volumetric sonar carving, which resolves a sonar beam's 20-degree elevation ambiguity by counting how often each cube is observed versus marked occupied across intersecting views, can be made fast enough for real-time use and can then serve as the geometric scaffold for optical texturing. The tank validation reconstructs both submerged objects and the tank wall despite occlusion and reverberation, and dimension measurements from the reconstruction agree with ground truth to within 1.2 cm for individual mesh cells and within roughly 5 to 11 cm for larger objects such as the tank, milk crate, and chain. The paper argues these errors are small enough for collision-safe manipulation, while the optical overlay adds the semantic information the sonar alone cannot provide. The contribution is presented as an extension of prior volumetric sonar reconstruction, with the novelty lying in the real-time pipeline, the deringing and normalization preprocessing, and the eye-in-hand sweep trajectory that gathers diverse views with minimal arm motion.","pith_inferences":["Processing time grows cubically with resolution in this scheme, so the advertised real-time rate is tied to the 5 cm grid; a sparse or hierarchical voxel representation is a natural way to push finer resolutions into the real-time range, though the paper does not test one.","The paper lists SLAM and dynamic object tracking as future work; if pose registration could come from the sensors rather than fixed-base kinematics, the same voxel-template pipeline would transfer to free-floating vehicle-manipulator systems.","The optical overlay inherits the failure modes of the background-segmentation step, so objects unlike those in the segmentation model's training data could yield a correct acoustic map while the projected texture mislabels or misplaces them.","The accuracy figures come from one tank with a small set of objects; a testable extension is to measure reconstruction error with targets at varying ranges and turbidity levels to see whether the same occupancy threshold $t_r$ holds outside the tested conditions."],"forward_implications":["The paper reports that at 0.05 m voxel resolution each sonar frame is processed in 0.052 s (18.6 FPS), which is faster than the sonar's 10 FPS capture rate, so the map can keep up with the sensor during the sweep.","The roughly 90-second sweep trajectory produces a reconstruction sufficient to guide the camera arm to close stand-off distances, converting the sonar-only map into a texture-rich scene for a human operator.","Because the method imposes no assumptions on scene geometry or vehicle motion, reconstruction quality should improve with view diversity, and the eye-in-hand arm supplies that diversity from a short baseline.","The reported dimension errors, 5.1 to 6.6 cm for acoustic-only measurements and 1.2 to 10.6 cm for optical-overlay measurements, are small relative to the 5 cm cube size and support the paper's claim that the grid is usable for real-time intervention.","Decimating sonar pixels increases false positives but not false negatives, so the low-resolution occupancy grid preserves the free-space information needed to avoid collisions."],"supporting_citations":[{"why":"Supplies the min-max volumetric sonar reconstruction and voxel-template concept that OASIS extends for real-time occupancy updates.","marker":"[22]"},{"why":"Defines the offline opti-acoustic neural rendering state of the art whose speed and accuracy OASIS is contrasted against.","marker":"[10]"},{"why":"Provides the neural implicit surface method whose long optimization time motivates the real-time approach.","marker":"[28]"},{"why":"Supplies the 3D Gaussian splatting baseline that achieves real-time rendering but has not been applied to opti-acoustic data.","marker":"[29]"},{"why":"Supplies the real-time Gaussian splatting SLAM system with depth input, the closest real-time baseline in the comparison table.","marker":"[32]"},{"why":"Provides the marching cubes algorithm used to convert the voxel occupancy grid into a mesh for optical overlay.","marker":"[34]"},{"why":"Supplies the dichotomous-image-segmentation model used to segment the optical image for the overlay.","marker":"[37]"},{"why":"Supplies the watershed algorithm used to derive the background mask from the rendered depth image.","marker":"[38]"},{"why":"Offers the alternative acoustic-lensing method for reducing elevation ambiguity that the paper sets aside to preserve field of view.","marker":"[33]"}],"fun_headline_variants":["Real-time opti-acoustic 3D mapping for underwater robots","Underwater 3D reconstruction at 18.6 frames per second","Sonar carving plus optical texture: real-time underwater mapping","Live underwater 3D from sonar and camera fusion","Subcentimeter underwater mapping with real-time sonar fusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes the manipulator's joint-angle sensors and forward kinematics give the true sonar and camera poses, so any undetected wrist deflection, mounting flex, or base motion will misalign the voxel grid and the optical overlay; the paper acknowledges this by stating that the method relies on accurate pose information from a fixed-base manipulator.","fun_headline_variants_meta":{"raw":{"variants":["Real-time opti-acoustic 3D mapping for underwater robots","Underwater 3D reconstruction at 18.6 frames per second","Sonar carving plus optical texture: real-time underwater mapping","Live underwater 3D from sonar and camera fusion","Subcentimeter underwater mapping with real-time sonar fusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000738,"raw_usage":{"total_tokens":3274,"prompt_tokens":903,"completion_tokens":2371,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":2285}},"tokens_in":519,"tokens_out":2371,"duration_ms":18648,"temperature":1.0,"reasoning_tokens":2285,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:25:15.035119+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same sweep trajectory twice with a rigid target fixed in the tank and the arm returned to its nominal home pose between runs; if the reconstructed target position shifts by more than a few centimetres, or drifts relative to a laser-measured reference, the fixed-kinematics pose assumption is falsified. A sharper test is to apply a known sagging load to the wrist and check whether the acoustic grid and the optical overlay separate by more than the 5 cm cube size.","supporting_citations":[{"cited_title":"Sonar-aided ma- nipulation in low-visibility conditions by novice users,","cited_arxiv_id":null,"evidence_quote":"Supplies the min-max volumetric sonar reconstruction and voxel-template concept that OASIS extends for real-time occupancy updates."},{"cited_title":"Aoneus: A neural rendering framework for acoustic-optical sensor fusion,","cited_arxiv_id":null,"evidence_quote":"Defines the offline opti-acoustic neural rendering state of the art whose speed and accuracy OASIS is contrasted against."},{"cited_title":"Rtg-slam: Real-time 3d reconstruction at scale using gaussian splatting,","cited_arxiv_id":null,"evidence_quote":"Supplies the real-time Gaussian splatting SLAM system with depth input, the closest real-time baseline in the comparison table."},{"cited_title":"Highly accurate dichotomous image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the dichotomous-image-segmentation model used to segment the optical image for the overlay."},{"cited_title":"Watersheds in digital spaces: An efficient algorithm based on immersion simulations,","cited_arxiv_id":null,"evidence_quote":"Supplies the watershed algorithm used to derive the background mask from the rendered depth image."},{"cited_title":"A method of generating 3d views of aquatic plants with didson,","cited_arxiv_id":null,"evidence_quote":"Offers the alternative acoustic-lensing method for reducing elevation ambiguity that the paper sets aside to preserve field of view."}],"review_version":2}