{"id":"c4378ebd-f560-47b3-8827-731c4b683d1a","arxiv_id":"2603.29227","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"An open-source robot mapping library estimates signed distance fields and uncertainties online by combining Bayesian Hilbert maps with Gaussian process regression.","lead":"Kernel-SDF builds a continuous, uncertainty-aware map of distances to nearby obstacles from streaming robot sensor data, in real time. It combines an occupancy-learning front end with a Gaussian-process back end, and reports better accuracy than four existing mapping pipelines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Global BHM threshold τ may not coincide with true surface ∂O across surface orientations, biasing front-end surface points and invalidating downstream SDF accuracy and uncertainty.","rationale":"The reader's weakest assumption identifies the global BHM threshold τ as the structural premise most likely to fail. This is indeed the most load-bearing concern: the front-end surface samples are the sole training data for the GP back-end, and the sign function directly determines the SDF. Because the paper itself demonstrates that log-odds vary systematically with surface orientation (Fig. 4, Section IV-A3), the τ-level set is unlikely to coincide with ∂O across all geometry. The moving-average update (Eq. 5) can at best capture the mean log-odds of hit points, not the per-surface variation; in scenes with mixed orientations, a single threshold must misclassify some surfaces. The paper provides no quantitative sign-accuracy or τ-sensitivity analysis, so the concern is unresolved. The proposed test—checking sign accuracy and log-odds distributions stratified by surface normal on a known ground-truth dataset—would directly determine whether the τ-level set is an acceptable surface proxy. If the test exposes a significant orientation-dependent bias, the central claims of superior SDF accuracy and calibrated uncertainty would need qualification. The reader's CONDITIONAL verdict is appropriate pending this test; no change in verdict is needed because the concern was already identified and the verdict accounts for it.","tokens_in":19856,"tokens_out":6797,"duration_ms":71508,"concrete_test":"Evaluate the BHM sign accuracy at ground-truth surface points on Replica room0, stratified by surface normal orientation (floor vs walls). Compute the log-odds l(x) at each ground-truth surface point using the trained BHM, and compare the per-group distributions of l(x) against the global τ. If the sign error rate (s(x) from Eq. 4 disagrees with true inside/outside) for floor points differs from that for wall points by more than 5%, or if the overall sign error rate exceeds 2%, the single global τ is not a faithful proxy for ∂O and the accuracy and uncertainty claims are compromised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Kernel-SDF produces an accurate, uncertainty-aware SDF hinges on the front-end extracting surface points that lie on the true surface ∂O. This extraction is performed at the level set l(x) = τ, where τ is a single scalar updated as a moving average of the log-odds of hit points (Eq. 5). The paper's own Fig. 4 and the text acknowledge that log-odds differ systematically between ground and walls and that τ varies with sensor viewing direction. If the BHM log-odds field is not constant on ∂O, then the τ-level set is not ∂O: for a surface patch whose log-odds exceed τ, the level set lies in free space, and for a patch whose log-odds are below τ, it lies inside the occupied region. The extracted surface points used to train the GP back-end are thus systematically biased, and the sign function (4) misclassifies points near those patches. Because the back-end GP is trained on these biased surface points and the SDF uncertainty is propagated from their locations (Eqs. 15-17), the inaccuracy propagates to SDF values, gradients, uncertainty, and mesh extraction. The paper provides no quantitative evaluation of sign accuracy or sensitivity to τ. This is a structural premise distinct from the log-GP distance computation: even if the GP regression is perfectly executed, a wrong training surface invalidates the output. The moving average only tracks the mean log-odds of observed hit points, not per-surface variation; in scenes with mixed orientations (e.g., floor and walls), a single threshold cannot simultaneously match both surfaces.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Kernel-SDF, an open-source library for real-time signed distance field (SDF) estimation from streaming depth or LiDAR data. The method couples a Bayesian Hilbert Map (BHM) front-end that learns a continuous occupancy/log-odds field, extracts surface points via marching cubes at a dynamically learned log-odds threshold τ, and estimates per-point surface uncertainty, with a log-Gaussian-process back-end that learns unsigned distances from those surface points. SDF sign is assigned by the BHM threshold, and SDF/gradient uncertainty is derived by propagating surface-point uncertainties through a softmin approximation of the distance function. Experiments compare mesh quality, SDF accuracy, gradient accuracy, timing, sensor-noise robustness, and a navigation demonstration against Voxblox, FIESTA, iSDF, and VDB-GPDF on Replica, Cow and Lady, and Newer College datasets.","tokens_in":20334,"tokens_out":6239,"duration_ms":70604,"significance":"If the claims are fully supported, this is a valuable practical contribution: a continuous, differentiable, uncertainty-aware SDF representation that runs online, with an open-source implementation and ROS interfaces. The two-stage design—BHM front-end for robust sign/surface extraction and log-GP back-end for distance regression—is well motivated, and the softmin-based uncertainty propagation is a useful alternative to the exploding variance of naive log-GP. The paper includes detailed derivations for the EM updates, gradient variance, and scaling trick, and the experimental suite covers multiple datasets and baselines. However, the current evidence does not yet support the strongest claims, particularly 'calibrated uncertainty', the global-threshold surface extraction, and 'consistently superior' SDF accuracy.","major_comments":[{"comment":"SDF sign accuracy is not evaluated. In Sec. V-B, the authors deliberately compare absolute distances to 'isolate distance accuracy from occupancy misclassification'. That is reasonable for distance evaluation, but the actual SDF is signed: sign errors can make free space appear occupied or vice versa, which is critical for planning and collision checking. The paper does not report signed SDF accuracy, sign confusion matrices, or sign error rates on any dataset, despite the sign being produced by the same τ-level-set mechanism questioned above. A signed evaluation would directly test the front-end's role in the central claim.","section":"V-B and Appendix F"}],"minor_comments":[{"comment":"The open-source link is 'available after acceptance' and Appendix F says 'For other parameters, please refer to our released code'. Key hyperparameters—α in Eq. (15), β in Eq. (6), the priority-queue weights η1, η2, γ, and the caps—are not listed in the paper. Even with a code link, the parameter settings should be fully documented for reproducibility.","section":"Footnote 1, Appendix F"},{"comment":"The surface-point uncertainty derivation assumes δx aligns with the gradient direction. This first-order approximation may be poor near high-curvature surface features or where the log-odds field is locally flat; no validation of this assumption is provided.","section":"IV-A4, Eq. (6)"},{"comment":"The 'real-time' claim is supported by average per-frame times around 0.11–0.19 s on the tested datasets. This is roughly 5–9 Hz; the paper should state the target sensor frame rate and whether the system keeps up on the actual data streams used in the navigation demo.","section":"Table IV"},{"comment":"For Voxblox and FIESTA, query points where prediction fails are excluded from SDF metrics. This is a reasonable practical choice, but it should be reported how many points were excluded for each baseline and dataset, since systematic exclusion could affect the comparison.","section":"VII-F.3"},{"comment":"Minor typos and notation inconsistencies exist (e.g., 'Voxel-based' in the introduction, the acronym expansions, and '∇xi gk' in Eq. (17) without dimensional indices). A careful proofread is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"This is a serious systems paper with a plausible architecture, but the current submission overstates its support for calibrated uncertainty and consistent accuracy. The global-τ surface extraction issue is the most structurally important concern and can be tested with sign-accuracy and sensitivity experiments. I would encourage the editor to require the code link to be functional and the missing hyperparameters documented before any final acceptance decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine systems contribution that should go to peer review, but the headline claims—'superior accuracy' and especially 'calibrated uncertainty'—are stronger than the evidence. The core design, a BHM front-end feeding a log-GP back-end with softmin uncertainty propagation, is a new combination of published pieces, and the engineering is serious: octree partitioning, weight sync, priority-queue updates, kernel scaling to prevent underflow. The Replica mesh metrics are consistently good, and the gradient MAE numbers beat the baselines by a wide margin.\n\nThe soft spots are real but not fatal. The global threshold τ is the load-bearing variable: the paper itself notes that log-odds differ between ground and walls, and the moving average in Eq. (5) is a single scalar that may not coincide with the true surface across all orientations. The stress-test worry is legitimate—if the τ-level set is biased, the training surface and therefore all downstream SDF values are biased. But this is exactly the kind of thing a sensitivity analysis would answer, and the authors don't provide one. That is the first thing I would ask a referee to probe.\n\nThe other gaps: uncertainty calibration is shown qualitatively in Fig. 8 only, no quantitative error-vs-variance metric; the closest GP baselines (specifically Log-GPIS) are cited but not compared; and the open-source library—a stated main contribution—is not actually released. Hyperparameters are tuned per dataset, which weakens the generality claim but is common in this literature.\n\nThe math is coherent. The softmin variance propagation is self-referential (it carries the front-end's uncertainty scalar β through), but that's a modeling choice, not a contradiction. The paper deserves serious refereeing. If the τ sensitivity holds up and the code ships, it will be a useful reference for anyone building online SDF maps for planning. I'd send it out with a request for calibration numbers and a τ-robustness study.","headline":"A solid, well-engineered SDF mapping system with a coherent new combo; the uncertainty-calibration claims outrun the evidence and the global τ threshold needs a sensitivity analysis.","tokens_in":20770,"tokens_out":2559,"would_cite":true,"duration_ms":28130,"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":"Kernel regression alone can build real-time signed distance fields with calibrated uncertainty for robot mapping.","keywords":["signed distance function","kernel regression","Gaussian process","occupancy mapping","uncertainty quantification","real-time mapping","robot navigation","LiDAR and depth sensors"],"falsifier":"In a scene with a floor and a wall, bin near-surface query points by the angle between the surface normal and the sensor ray and measure sign error per bin; if sign errors concentrate in specific angle bins, the single scalar threshold is not faithful and the front-end surface is biased.","tokens_in":19779,"feed_emoji":"🗺️","tokens_out":6502,"duration_ms":60271,"temperature":0.7,"pith_summary":"The paper claims that a signed distance field (SDF) with per-point uncertainty can be learned online from streaming depth or LiDAR data using only kernel regression, without neural network training or fixed-resolution voxels. The method splits the problem into a front-end that learns a continuous occupancy log-odds field and extracts surface points at the occupancy boundary, and a back-end that fits Gaussian process regressors to those surface points to produce distance, gradient, and variance estimates. Why this matters: safe robot planning needs both distance to obstacles and a sense of confidence in that distance, and current voxel, neural, and Gaussian process approaches each sacrifice one of accuracy, speed, scalability, or uncertainty. The paper reports that its uncertainty estimates track actual SDF errors, and that an autonomous ground robot can use the map to plan a bubble-cover trajectory in real time.","feed_headline":"Kernel regression maps rooms into distance fields at 150 ms per frame","feed_subtitle":"Accurate distance, gradients, and uncertainty estimates stream from depth or LiDAR at online speed.","key_machinery":"The pipeline is carried by two linked regression stages: the front-end is a continuous occupancy log-odds field learned by Bayesian kernel regression with sparse RBF features, from which marching cubes extracts surface points and surface normals at the learned occupancy threshold; the back-end is a set of octree-local Gaussian process regressors trained on those surface points with log-transformed labels, using the short-time heat-kernel identity to recover unsigned distance. The bridge between the stages is a softmin reinterpretation of the log-GP posterior mean, which enables RBF kernels via a scaling factor and propagates surface-point position variances through a first-order Taylor expan","core_discovery":"The central claim is that an SDF can be decomposed into sign and unsigned distance, d(x)=s(x)u(x), with the sign supplied by a Bayesian kernel-regression occupancy field and the unsigned distance supplied by Gaussian process regression in log space on surface points extracted from that field. The paper shows that the log-GP posterior mean behaves like a softmin over the surface points, which justifies using a computationally cheaper RBF kernel and, crucially, lets surface-point position uncertainties propagate through a first-order expansion into uncertainties on SDF values and gradients. Across synthetic and real-world indoor and outdoor scenes, this two-stage design reports lower SDF and g","pith_inferences":["The softmin variance estimate depends on the geometry of nearby surface points; in regions where the nearest surface changes abruptly or the surface-point cloud is sparse, the reported variance may understate true error. This follows from the first-order propagation and is not a claim the paper tests.","The single moving-average log-odds threshold is an engineering simplification: the paper itself observes that log-odds differ between floor and wall surfaces, so a spatially varying threshold or per-surface calibration could further reduce sign errors.","A natural testable extension is to feed the SDF variance back into a pose-graph or SLAM backend as measurement weights, since the library already exposes per-query uncertainty."],"forward_implications":["A robot can maintain a differentiable SDF map with gradients from depth or LiDAR streaming at online speed, without offline training.","Motion planners can scale safety margins using the predicted distance uncertainty, as the paper demonstrates with a safe bubble-cover path.","The octree partitioning and priority-queue updates make the approach applicable to large environments while focusing computation on regions actively queried.","Because the representation is continuous, the map supports gradient-based optimization for trajectory and manipulation planning.","If the claimed error-uncertainty consistency holds, the variance output can be used to gate decisions in risk-aware navigation rather than treated as an uncalibrated heuristic."],"fun_headline_variants":["Kernel-SDF: real-time distance fields with calibrated uncertainty","Sign from occupancy, distance from log-GP softmin","150 ms per frame: kernel SDF with calibrated uncertainty","Open-source kernel SDF: fast, accurate, uncertainty-aware","Kernel regression SDF: sign and distance with uncertainty"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that one moving-average log-odds threshold separates occupied from free space at the true surface for all octants and viewing directions; if that threshold drifts from the true boundary on some surface, the extracted training surface is wrong and both the sign and the distance regressions inherit the error.","fun_headline_variants_meta":{"raw":{"variants":["Kernel-SDF: real-time distance fields with calibrated uncertainty","Sign from occupancy, distance from log-GP softmin","150 ms per frame: kernel SDF with calibrated uncertainty","Open-source kernel SDF: fast, accurate, uncertainty-aware","Kernel regression SDF: sign and distance with uncertainty"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000714,"raw_usage":{"total_tokens":3034,"prompt_tokens":714,"completion_tokens":2320,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":2237}},"tokens_in":458,"tokens_out":2320,"duration_ms":16053,"temperature":1.0,"reasoning_tokens":2237,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T17:03:48.487205+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a scene with a floor and a wall, bin near-surface query points by the angle between the surface normal and the sensor ray and measure sign error per bin; if sign errors concentrate in specific angle bins, the single scalar threshold is not faithful and the front-end surface is biased.","supporting_citations":[],"review_version":1}