{"id":"9424aef7-9c2d-47bd-b3f3-c9178600b9e8","arxiv_id":"2502.09824","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"PUGS reweights grasp confidence by occupancy uncertainty from multi-view stereo, improving simulated grasping on partial and noisy underwater reconstructions.","lead":"This paper proposes PUGS, a method that estimates how uncertain a robot's 3D reconstruction is in different regions, and uses that uncertainty to pick more reliable grasp points for underwater manipulation. In simulations, reweighting grasps by this uncertainty improves success on partial and noisy reconstructions, though real-world tests are only qualitative.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed benefit rests on an untested reweighting heuristic: no ν-sweep, no simpler baseline, and clear degradation on complete scenes, so the uncertainty signal is not shown to drive the improvement.","rationale":"The reader's weakest-assumption analysis identifies Eq. (13) and the fixed ν=5 as the unexplained core of the method; I agree that this is the most load-bearing concern. The paper's entire value proposition is that perceptual uncertainty, propagated through the FOF and SVGP, provides a better grasp-ranking signal than raw TSGrasp confidence. But the fusion rule is a heuristic with a fixed exponent, shown in Eq. (13), and the paper gives no theoretical justification for dividing by variance raised to ν=5. Without a ν-sweep or a comparison to simpler uncertainty surrogates, the improvements on partial and noisy reconstructions cannot be attributed to the uncertainty representation itself. The complete-reconstruction results strengthen this concern: PUGS degrades several metrics compared to TSGrasp, which is not merely a neutral effect of uncertainty being less useful but an indication that the reweighting can be harmful. A secondary technical weakness is Eq. (3), which propagates pose uncertainty by simply adding the translation covariance to the measurement covariance, dropping the rotational covariance and the R Σ_pc R^T term; this means the pose-uncertainty channel is not actually modeled as claimed. However, even if Eq. (3) were corrected, the central claim would still rest on the unsupported heuristic in Eq. (13), so the reweighting rule is the single most load-bearing issue. The proposed concrete test is straightforward and would settle whether the improvement is due to the propagated uncertainty or to any generic penalty on low-observation regions. Given this, the reader's CONDITIONAL verdict stands: the paper is a plausible pipeline but needs the ablation, hyperparameter sensitivity, and more trials before the central claim can be accepted.","tokens_in":11474,"tokens_out":4196,"duration_ms":47970,"concrete_test":"Run a controlled ablation on the same simulated partial and noisy-partial scenes: (a) sweep ν over {0, 1, 2, 5, 10}; (b) replace σ²_occ with (i) inverse local point density and (ii) a random permutation of the same σ²_occ values; (c) run at least 20 trials per condition and report the three Table I metrics. If any simple alternative matches or beats PUGS, or if performance flips sharply with ν or with the random seed, then the propagated uncertainty itself is not what drives the improvement, and the central claim requires revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (13) is the only mechanism by which perceptual uncertainty changes grasp selection: c̃_G = GS(p_G) / (σ²_occ)^ν, with ν fixed to 5. The paper provides no derivation for this functional form, no sensitivity analysis for ν, and no comparison to simpler penalties such as local point density or distance to the reconstruction boundary. The central claim — that propagating measurement and pose uncertainty improves grasp robustness — therefore depends entirely on an unexplained hyperparameter heuristic. The degradation on complete reconstructions in Table I makes this concern concrete: PUGS lowers kettlebell gripper-closed success from 100% to 60% and goal-reached from 100% to 60%, and lowers coffee-mug goal-reached from 60% to 20%. The authors call this expected because uncertainty is less impactful on complete reconstructions, but systematic degradation is not the same as being less impactful; it shows the reweighting can actively harm grasp ranking when the uncertainty signal is not informative. Since the reported partial/noisy improvements are at most three successful trials out of five per condition, the observed advantage could easily be an artifact of the ν=5 heuristic re-ranking a few borderline grasps rather than evidence that the propagated uncertainty is a robust selection signal. Until Eq. (13) is shown to be more effective than plausible alternative penalties, and robust to its own hyperparameter, the central claim is not empirically established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PUGS, a pipeline that estimates occupancy uncertainty from multi-view stereo depth and camera-pose uncertainty by propagating measurement and pose covariances into a fused occupancy field (FOF) and then training a stochastic variational Gaussian process (SVGP) to produce predictive occupancy variance. That variance is fused with the FOF through cubature integration and used to reweight the grasp confidences output by the TSGrasp baseline via Eq. (13), c_obs = GS(p_G) / (sigma_occ^2)^nu, with nu fixed to 5. The method is evaluated in Gazebo simulation on a kettlebell and a coffee mug under partial, noisy, and complete reconstructions, and qualitatively in a real-world test tank with a Bravo 7 manipulator and Trisect stereo sensor. The paper claims that accounting for perceptual uncertainty makes grasp selection robust to partial and noisy underwater measurements.","tokens_in":11899,"tokens_out":3908,"duration_ms":38696,"significance":"If the central claim is established, PUGS would be a useful contribution: it converts perceptual uncertainty that is typically discarded into a grasp-ranking signal, and it does so in an underwater setting where real evaluations are scarce. The paper has clear strengths: it explicitly models two uncertainty sources (depth measurement noise and pose covariance), it provides a full pipeline from stereo images to grasp selection, it includes simulation and real-world test-tank experiments, and it reports the degradations on complete reconstructions rather than hiding them. However, the current evidence is not strong enough to support the central claim. The evaluation relies on only five simulation trials per condition with no error bars, the real-world results are qualitative, and the key reweighting rule in Eq. (13) is an unexplained heuristic with a fixed exponent. The paper's own Table I shows systematic degradation on complete reconstructions, which indicates that the uncertainty signal can actively harm grasp ranking when it is not informative.","major_comments":[{"comment":"The entire proposed benefit is mediated by Eq. (13), which reweights grasp confidence by dividing by the occupancy variance raised to a fixed exponent nu=5. The paper provides no derivation for this functional form, no sensitivity analysis over nu, and no comparison to simpler penalties such as local point density or distance to the reconstruction boundary. Table I makes the risk concrete: on complete reconstructions, PUGS reduces kettlebell gripper-closed success from 100% to 60% and goal-reached from 100% to 60%, and coffee-mug goal-reached from 60% to 20%. The authors call this 'expected,' but systematic degradation is not the same as being less impactful; it shows that the reweighting can actively harm grasp ranking when the uncertainty signal is uninformative. Unless Eq. (13) is shown to be more effective than plausible alternatives and robust to nu, the central claim that perceptual uncertainty improves grasp selection is not empirically established.","section":"III.E, Eq. (13), and IV.A"},{"comment":"The simulation evaluation uses N=5 trials per condition with no error bars or confidence intervals, and the reported PUGS improvements over TSGrasp are at most three successful trials out of five per condition. The protocol described in Section IV.B compounds this: 'we transform the object's pose to make the grasp collision-free and kinematically feasible. If neither of these is achievable, the experiment is repeated.' This means the reported success rates are conditioned on a post-hoc selection of feasible runs, which can systematically favor whichever method happens to produce feasible grasps more often. To assess the central claim, the paper should report unmodified trial counts (or clearly separate the feasibility filter from grasp success), increase the number of trials, and provide statistical measures such as confidence intervals or a paired test.","section":"IV.B and Table I"},{"comment":"The predictive occupancy uncertainty is defined as sigma2_pred = sigma2_Z / (1 + |mu_Z|), with no derivation or validation. The text notes that SVGP variance is dictated by spatial frequency rather than the regressed function value, and Eq. (8) is introduced purely as a rescaling to account for occupancy density. Since this rescaled variance flows directly into the cubature integration and then into Eq. (13), the arbitrariness of Eq. (8) compounds the arbitrariness of the grasp reweighting. The paper should justify this form empirically (e.g., by comparing different rescaling choices or by showing that the resulting variance correlates with measured grasp failures) or replace it with a more principled uncertainty definition.","section":"III.C, Eq. (8)"},{"comment":"The real-world evaluation is entirely qualitative. The paper reports selected grasp poses overlaid on reconstructions and asserts that PUGS 'consistently leads the gripper pose to areas with more observations while retaining the grasping area's geometric feasibility,' but there is no quantitative measure of grasp success, no repeated trials, and no error analysis for the real-world setting. Given that the simulation results are themselves limited, a quantitative real-world evaluation (even with a small number of attempts) is necessary to support the claim that the method works outside simulation.","section":"V.B"}],"minor_comments":[{"comment":"The conclusion states that PUGS requires 'no hand-tuning,' but Section IV.A fixes the depth noise variance to sigma2_uv = 0.001 m and the grasp weight exponent to nu = 5, and Section V.B describes tuning the outlier removal threshold to sigma_thresh = 0.1 for the leftmost real-world result. Please clarify which parameters are considered hand-tuned and which are fixed by the problem.","section":"Section VI vs. IV.A and V.B"},{"comment":"The depth noise variance is described as 'static across frames and fixed as sigma2_uv = 0.001 m,' but the noisy partial reconstruction in simulation uses sigma2_uv = 0.01 m. It would help to state explicitly how sigma2_uv is set in the real-world experiments and whether it comes from the depth model or is chosen by the authors.","section":"IV.A and IV.B"},{"comment":"There is a typo in the text following Eq. (7): 'variance of the occuancy density' should be 'variance of the occupancy density.'","section":"III.C, Eq. (7)"},{"comment":"The introduction contains a grammatical error: 'the uncertainty of the occupied regions can be a useful for improving existing grasp selection methods' should be 'can be useful for improving' or 'can be a useful signal for improving.'","section":"Section I"},{"comment":"The paper states 'Code will be made available at https://onurbagoren.github.io/PUGS/' but the repository is not linked at the time of review. If the code is not yet public, the statement should be revised to indicate the planned release date or an alternative availability mechanism.","section":"Abstract and Section VI"}],"recommendation":"major_revision","confidential_remarks":"The paper appears aimed at a robotics venue (ICRA/IROS style), where N=5 simulation trials and qualitative real-world demonstrations are often tolerated, but for a journal the standards are higher. The most serious issue is that the method's only mechanism for changing grasp selection, Eq. (13), is an unvalidated heuristic that demonstrably harms performance on complete reconstructions. The authors should be pushed to either provide a principled derivation for the reweighting rule or to show empirically that it outperforms simpler alternatives and is insensitive to nu. The admission in Section IV.B that experiments are repeated when no feasible grasp is found is a red flag for selective reporting and should be addressed transparently in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague —\n\nThe genuinely new piece here is the specific pipeline: take depth and pose covariances from an underwater stereo/SLAM setup, build a fused occupancy field as a GMM of backprojected points, train an SVGP on that field to get predictive occupancy variance, then reweight TSGrasp's grasp confidence by that variance (Eq. 13). The individual ingredients are not new — SVGP terrain modeling, Bayesian fusion, cubature integration all come from prior work — but the combination, and the target domain, is a legitimate contribution. The paper is clearly written and the authors are straight about what TSGrasp does and what they add.\n\nThe soft spots are real. The evaluation is thin: N=5 per condition, no error bars, and a protocol that repeats experiments when no feasible grasp is found and transforms the object pose to make grasps feasible. That alone would make me want more evidence before trusting the headline numbers. The bigger issue is Eq. (13). The weighting by σ^ν with ν=5 fixed is a heuristic, with no derivation, no sensitivity sweep, and no comparison to simpler penalties like point density or distance to reconstruction boundary. The complete-reconstruction results show PUGS actively hurting performance — kettlebell gripper-closed drops 100%→60%, coffee mug goal-reached 60%→20% — and the paper's explanation that uncertainty is \"less impactful\" does not address the fact that the reweighting is systematically degrading grasp ranking when it shouldn't. The stress-test note is right that until Eq. (13) is shown to beat plausible alternatives and to be robust to ν, the central claim is not empirically established.\n\nI wouldn't call it a takedown. The uncertainty-propagating machinery is coherent, the real-world test tank is a genuinely useful setting, and the qualitative results are suggestive. But the current evidence supports only a modest claim: PUGS can shift grasps toward well-observed regions in some failure cases. The stronger claim — that propagating measurement and pose uncertainty makes grasp selection robust to partial/noisy reconstruction — needs more runs, quantitative real-world metrics, and a serious ablation of the reweighting rule.\n\nFor me: this deserves a real referee, but I'd want major revision before publication. Not something I'd cite yet. I'd bring it to a reading group as an example of how a plausible uncertainty pipeline can outrun its evaluation.","headline":"A coherent uncertainty-aware grasp selection pipeline, but the central reweighting heuristic is untested and the evaluation is too thin to support the robustness claim.","tokens_in":12339,"tokens_out":1567,"would_cite":false,"duration_ms":15188,"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":"Reweighting grasp confidence by occupancy uncertainty makes autonomous underwater grasping succeed on partial and noisy 3D reconstructions.","keywords":["perceptual uncertainty","occupancy uncertainty","grasp selection","underwater manipulation","multi-view reconstruction","stochastic variational Gaussian process","stereo depth noise","pose uncertainty"],"falsifier":"Run the partial-kettlebell simulation with a sweep of the weighting exponent $\\nu$ in Eq. (13) over values such as 0, 1, 3, 5, 7, and 10, and record grasp traversal success; if the reported 60% success at $\\nu=5$ is not part of a broad plateau, or if $\\nu=0$ reaches the same success, then the advantage is a tuned artifact rather than a general property of uncertainty reweighting. The complete-reconstruction result already provides a check: at $\\nu=5$ on a complete kettlebell, PUGS drops from the baseline's 100% to 60% traversal success.","tokens_in":11305,"feed_emoji":"🦾","tokens_out":9979,"duration_ms":84543,"temperature":0.7,"pith_summary":"PUGS is built around a single idea: when a robot reconstructs an object from noisy stereo images and drifting pose estimates, not all points on the reconstructed surface deserve equal trust, so grasp selection should be driven by a map of how sure the reconstruction is. The paper constructs that map by propagating depth-measurement noise and camera-pose covariance into a fused occupancy field, fitting a stochastic variational Gaussian process over it to get predictive variance, and fusing both uncertainty sources into an occupancy variance per point. It then reweights the grasp confidences from an existing grasp model (TSGrasp) by dividing them by occupancy variance raised to a fixed exponent. In simulation on partial and noisy partial reconstructions, the paper reports that this reweighting turns failed grasps into successful ones, while on complete reconstructions the baseline still performs better on several metrics. The intended payoff is practical: underwater manipulators can pick more reliable grasp locations without new hardware, using uncertainty that already exists in the mapping pipeline.","feed_headline":"Uncertainty reweighting turns failed underwater grasps into successes","feed_subtitle":"PUGS divides grasp confidence by occupancy variance, steering robot hands to well-observed parts of partial 3D scenes.","key_machinery":"The central object is the occupancy uncertainty map $\\sigma_{\\mathrm{occ}}^2$, a per-point variance over whether a location is occupied, assembled by fusing observational and predictive uncertainties. Observational uncertainty flows from stereo depth noise and SLAM pose covariance through the backprojection Jacobian into a fused occupancy field (FOF), a Gaussian mixture over backprojected points with weighted Bayesian fusion of nearby Gaussians. Predictive uncertainty comes from a stochastic variational Gaussian process regressed on the FOF, with variance rescaled by occupancy density via $\\sigma_{\\mathrm{pred}}^2 = \\sigma_Z^2/(1+|\\mu_Z|)$. Cubature integration over $\\sigma$ points derived from the positional covariance combines the two, and the final scalar variance is used as the denominator in Eq. (13) with exponent $\\nu=5$. All of the reported behavior flows from this single reweighting identity.","core_discovery":"The paper's central claim is that a grasp's confidence should be penalized where the 3D reconstruction is uncertain, via the rule $\\tilde{c}_G = GS(p_G)/(\\sigma_{\\mathrm{occ}}^2)^\\nu$ with $\\nu=5$, where $GS$ is a pretrained grasp-selection network and $\\sigma_{\\mathrm{occ}}^2$ is the occupancy variance at the grasp points. That occupancy variance is the product of two modeled sources: observational uncertainty, obtained by backprojecting stereo depth with variance $\\sigma_{uv}^2$ and camera pose with covariance $\\Sigma_C^i$ into a fused occupancy field of Gaussians, and predictive uncertainty, obtained by training a stochastic variational Gaussian process on that field and scaling its variance by occupancy density. The two are merged through cubature integration over the positional covariance. The paper reports that this reweighted score selects grasp poses away from the incomplete edges of partial reconstructions, converting 0% grasp success for the baseline into 40--80% success depending on object and metric, while on complete reconstructions TSGrasp retains the advantage on most metrics.","pith_inferences":["A testable extension is to treat the exponent $\\nu$ in Eq. (13) as a learned or per-object parameter; the reported numbers use one fixed value, and a sweep would reveal whether the benefit is a plateau or a tuned peak.","The occupancy-uncertainty field is also a map of where observations are missing, so the same representation could drive next-best-view planning or active perception rather than only grasp reranking.","Because the real-world evaluation is qualitative and limited to one object, a quantitative tank trial with multiple objects and repeated trials would show how much of the simulation benefit survives real stereo noise and pose drift.","On complete reconstructions the uncertainty weighting hurts, so a principled variant might gate the reweighting by a global completeness measure, combining the baseline's complete-scene performance with PUGS's partial-scene gains."],"forward_implications":["Underwater robots can get better grasps on partially reconstructed objects without additional sensors, because the uncertainty signal already present in stereo depth and SLAM covariance is enough to steer the gripper away from unreliable geometry.","The reweighting is agnostic to the grasp network's internals, so any grasp model that outputs per-grasp confidences can be wrapped by PUGS instead of being retrained.","PUGS reuses the reconstructed point cloud and adds no new hardware, so the only added cost is the post-processing uncertainty computation, which the paper notes currently takes minutes for real-world scenes.","Because the same occupancy variance map hurts performance on complete reconstructions, the benefit is specific to partial or noisy inputs; a deployable system would gain from activating the weighting only when the reconstruction is known to be incomplete."],"supporting_citations":[{"why":"TSGrasp is the baseline grasp-selection network whose per-grasp confidences PUGS divides by occupancy variance; it supplies the scores being reweighted and the comparison in every experiment.","marker":"[19]"},{"why":"RTAB-Map provides the synchronized camera poses and keyframe covariance estimates that PUGS treats as pose uncertainty and propagates into the occupancy field.","marker":"[30]"},{"why":"RAFT-Stereo computes the depth images from the stereo pairs; the depth variance assumption enters the backprojection covariance in Eq. (1).","marker":"[32]"},{"why":"Torroba et al. supply the stochastic variational Gaussian process mapping and the cubature-integration fusion pattern that PUGS adapts for predictive occupancy uncertainty.","marker":"[28]"},{"why":"Hensman et al. provide the SVGP formulation used to regress the fused occupancy field and return predictive mean and variance.","marker":"[35]"},{"why":"Barfoot's weighted Bayesian fusion is the mechanism used to fuse nearby Gaussians in the fused occupancy field before cubature integration.","marker":"[34]"},{"why":"Rosinol et al. show depth-estimation uncertainty as a weighting factor for volumetric fusion, the closest prior that weights measurements by uncertainty rather than treating them equally.","marker":"[15]"}],"fun_headline_variants":["Uncertainty reweighting lifts underwater grasp success up to 80%","PUGS steers grasps to certain parts of partial 3D scenes","Underwater grasping: divide confidence by uncertainty to succeed","Robots that doubt their 3D vision grasp better underwater"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything hinges on the unexamined choice that dividing a grasp score by the fifth power of occupancy variance ranks grasps better than the raw score; the paper fixes $\\nu=5$ without deriving it, and on complete reconstructions the same rule makes PUGS perform worse than the baseline, so if that exponent is miscalibrated the claimed benefit has no principled foundation.","fun_headline_variants_meta":{"raw":{"variants":["Uncertainty reweighting lifts underwater grasp success up to 80%","PUGS steers grasps to certain parts of partial 3D scenes","Underwater grasping: divide confidence by uncertainty to succeed","Robots that doubt their 3D vision grasp better underwater"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000833,"raw_usage":{"total_tokens":3611,"prompt_tokens":897,"completion_tokens":2714,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":2638}},"tokens_in":513,"tokens_out":2714,"duration_ms":18360,"temperature":1.0,"reasoning_tokens":2638,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T20:21:20.255679+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the partial-kettlebell simulation with a sweep of the weighting exponent $\\nu$ in Eq. (13) over values such as 0, 1, 3, 5, 7, and 10, and record grasp traversal success; if the reported 60% success at $\\nu=5$ is not part of a broad plateau, or if $\\nu=0$ reaches the same success, then the advantage is a tuned artifact rather than a general property of uncertainty reweighting. The complete-reconstruction result already provides a check: at $\\nu=5$ on a complete kettlebell, PUGS drops from the baseline's 100% to 60% traversal success.","supporting_citations":[{"cited_title":"Real-Time Generative Grasping with Spatio-temporal Sparse Convolution,","cited_arxiv_id":null,"evidence_quote":"TSGrasp is the baseline grasp-selection network whose per-grasp confidences PUGS divides by occupancy variance; it supplies the scores being reweighted and the comparison in every experiment."},{"cited_title":"Scalable Variational Gaussian Process Classification,","cited_arxiv_id":null,"evidence_quote":"Hensman et al. provide the SVGP formulation used to regress the fused occupancy field and return predictive mean and variance."},{"cited_title":"Probabilistic volumetric fusion for dense monocular slam,","cited_arxiv_id":null,"evidence_quote":"Rosinol et al. show depth-estimation uncertainty as a weighting factor for volumetric fusion, the closest prior that weights measurements by uncertainty rather than treating them equally."}],"review_version":1}