{"id":"7afae5da-a282-4de9-8f30-77abb6b2a9fa","arxiv_id":"2412.20612","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Kernel SHAP can attribute ICP registration uncertainty to sensor noise, initial pose error, and partial overlap, with sensor noise usually the most important source.","lead":"This paper applies an existing AI explanation technique, SHAP, to explain why a point cloud registration algorithm is uncertain, attributing the uncertainty to sensor noise, bad starting poses, and partial overlap. On eight real-world sequences, sensor noise is usually the dominant contributor, which could help robots and operators diagnose and respond to perception failures.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The omission of the SHAP bias term assumes zero baseline uncertainty, but the reference input has nonzero initial-pose scale and finite-sample KL, so attributions may be biased.","rationale":"The reader's weakest_assumption correctly identifies the zero-baseline assumption as load-bearing. My independent reading confirms this and finds an internal inconsistency that strengthens it: Section 5 states the bias phi0 is omitted because the expected value for the unperturbed case is 0, but Algorithm 1 defines reference values r = {0, 1, 0}, meaning the 'unperturbed' input has initial-pose scale s = 1 and therefore nonzero initial-pose uncertainty. The pose uncertainty target is the KL divergence between two 100-sample distributions; with finite samples, even two draws from the same input distribution give positive KL. Thus f(r) is not zero, and forcing phi0 = 0 distorts all SHAP values. The central claim that the method 'reasonably explains uncertainty sources' depends on the SHAP ordering in Table 1, which is exactly what this bias could change. The proposed test—computing f(r) across seeds and re-running the regression with phi0 free—settles the concern. If f(r) is negligible relative to the reported SHAP values, the paper's conclusion survives; if not, the attributions are unreliable. I agree with the reader's CONDITIONAL verdict: the idea is promising, but this specific assumption must be validated before the explanation can be trusted. The paper also lacks error bars and repeated-seed analysis, which the reader noted; my primary concern is the zero-baseline issue, which should be addressed first.","tokens_in":8478,"tokens_out":5071,"duration_ms":47494,"concrete_test":"Quantify the empirical baseline uncertainty f(r) for the reference input (no sensor noise, initial-pose scale s = 1, no partial-overlap removal) by computing, across multiple random seeds, the KL divergence between two independent 100-sample ICP pose distributions. If the mean of f(r) is not negligibly small relative to the median SHAP values in Table 1 (e.g., more than 10% of the sensor-noise SHAP value), then re-run Algorithm 1 without omitting phi0, i.e., allow phi0 as a free parameter in the weighted linear regression. If the resulting median SHAP ordering changes, the zero-baseline assumption is load-bearing and the reported attributions are unreliable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim hinges on the additive model g(z') = phi_sn*z'_sn + phi_ip*z'_ip + phi_po*z'_po with the bias phi0 omitted because 'the expected value for the unperturbed case is 0' (Section 5). However, in the experiments the reference input is r = {0, 1, 0} (Algorithm 1): no sensor noise, initial-pose scale s = 1, no overlap removal. The 'unperturbed case' is therefore not a zero-uncertainty input; it is the lowest-perturbation input, and the pose uncertainty f(r) is the KL divergence between two 100-sample distributions of ICP pose estimates, both drawn with s = 1. Even if no perturbation were applied, two finite samples from the same input distribution yield different sample covariances, so the estimated KL is positive and does not vanish at 100 samples. Because phi0 is forced to zero, the weighted linear regression in Algorithm 1 absorbs this baseline into the three feature coefficients, biasing every SHAP value. In particular, the ordering in Table 1 (sensor noise > partial overlap > initial pose uncertainty) may partly reflect the baseline offset rather than true marginal contributions. The paper also contains an internal inconsistency: Section 5 says all z' = 0 is the unperturbed case with zero expected value, while Algorithm 1 sets r = {0, 1, 0}, a nonzero-uncertainty reference; the reference-values section is left as 'Section ??', an omitted detail that obscures this tension.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the interpretability of uncertainty estimates produced by a probabilistic ICP pipeline. The authors identify three sources of pose uncertainty (sensor noise, initial pose uncertainty, partial overlap) and propose to attribute an estimated pose-uncertainty scalar to these sources via kernel SHAP. The uncertainty scalar is computed as the KL divergence between two 100-sample distributions of ICP pose estimates obtained under a 'pseudo-true' and a perturbed setting. The experiments on the Challenging datasets include SHAP summary, waterfall, and dependence plots for one sequence pair, and a median SHAP table across eight sequences. The authors conclude that the explanation method can reasonably identify which uncertainty source dominates.","tokens_in":8796,"tokens_out":6525,"duration_ms":61264,"significance":"If the result were fully validated, this would be a useful step toward human-interpretable failure analysis for point cloud registration, as operators could see whether sensor noise, initial pose error, or partial overlap drives the reported pose uncertainty. The paper applies kernel SHAP, a principled additive attribution method, in a domain where it has not been used before, and the experimental setup covers a realistic multi-sequence dataset. The authors also make the standard choice to explain only three of the five known uncertainty sources and state this limitation explicitly. However, the central validation is qualitative, and one load-bearing technical assumption (the zero-bias baseline) is not met in the implementation; the significance is thus conditional on correcting that assumption and providing a more quantitative evaluation of explanation quality.","major_comments":[{"comment":"The claim that the bias phi0 can be omitted because the expected value for the unperturbed case is 0 is contradicted by the experimental setup. Algorithm 1 sets the reference values to r = {0, 1, 0}, so the reference input has initial-pose scale s = 1, not zero; moreover, the pseudo-true distribution in Section 6 is itself sampled with s = 1. Since the pose uncertainty is the KL divergence between two 100-sample distributions, f(r) is generally positive even for unperturbed inputs, because two finite samples from the same input distribution yield different sample covariances. The intercept-free weighted linear regression then forces this nonzero baseline into the three feature coefficients, biasing every SHAP value. This issue directly affects the attribution ordering reported in Table 1 and must be resolved before the central claim can be assessed.","section":"Section 5, Algorithm 1, Section 6"},{"comment":"The definition of the model output f is under-specified. The text says the uncertainty is obtained by 'comparing the covariance matrices of these two normal distributions via KL divergence', but it does not provide the KL formula, the coordinate parameterization of the pose distributions (e.g., R^6 via se(3) or a manifold-valued Gaussian), nor the procedure for converting the 100 sampled poses into the two normal distributions. It also does not state how singular or near-singular covariance matrices are handled. Because kernel SHAP explains this scalar f, the paper must give a complete, reproducible mathematical definition of f.","section":"Section 6, Pose Uncertainty Estimation"},{"comment":"The evidence for the central claim ('can reasonably explain the uncertainty sources') is qualitative. The summary, waterfall, and dependence plots are interpreted subjectively, negative SHAP values in Figure 2 are mentioned but not analyzed, and Table 1 only reports medians. No quantitative criterion (e.g., faithfulness of the additive model, agreement with controlled single-source perturbation responses, or comparison against a baseline attribution method such as LIME or simple sensitivity analysis) is provided. Without such a criterion, the conclusion that the method 'can reasonably explain' the uncertainty sources is not established beyond a descriptive level.","section":"Sections 6.1 and 6.2"}],"minor_comments":[{"comment":"Algorithm 1 contains the placeholder 'see details in Section ??'; this cross-reference should be completed.","section":"Algorithm 1"},{"comment":"'More details are discuss in our experiments later' contains a grammatical error ('discuss' should be 'discussed').","section":"Section 3.3"},{"comment":"'the uncertainty of ICP uncertainty comes mainly' is a typo; the duplicated word should be removed.","section":"Section 3.2"},{"comment":"The selection of point clouds 6 and 7 from the Apartment sequence is not motivated; a brief explanation of how these pairs are chosen would improve reproducibility.","section":"Section 6.1"},{"comment":"The mapping function h_x is only described qualitatively; a precise definition of h_x for each feature (e.g., how z' = 0 maps to the reference value and z' = 1 maps to the perturbed value) would reduce ambiguity.","section":"Section 5 and Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The central flaw (bias omission vs. nonzero baseline) is methodologically serious and currently prevents the main conclusion from being accepted. The lack of quantitative validation and the under-specified KL definition further weaken the paper. However, the flaws are fixable within the manuscript's scope: the authors could include the intercept in the SHAP model, redefine the reference to a genuinely zero-uncertainty baseline (or estimate the baseline offset), and add a quantitative evaluation. In its current form the paper is more suitable as a workshop or short paper; after a major revision it could be a reasonable journal contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is applying kernel SHAP to explain why a probabilistic ICP method reports high pose uncertainty, with a perturbation scheme that maps sensor noise, initial-pose scale, and partial overlap into binary features. That mapping is clean and the idea is worth publishing as a proof of concept. The paper is also honest about what it does not do: it explicitly says causality is unsettled and leaves under-constrained cases and intrinsic randomness out.\n\nThe soft spots are real but mostly fixable. The biggest one is the omitted bias term. Section 5 says phi_0 is zero because the unperturbed case has expected uncertainty zero, but Algorithm 1 sets the reference to {0, 1, 0} — meaning initial-pose scale 1, not zero. The pseudo-true distribution is also sampled with s = 1, so the \"unperturbed\" input already carries uncertainty, and with only 100 samples the empirical KL divergence between two draws from the same distribution will be positive. Forcing the intercept to zero makes the weighted linear regression absorb that baseline into the three feature coefficients, so the median SHAP ordering in Table 1 may partly reflect the baseline offset rather than true marginal contributions. The dangling 'Section ??' for the reference values is a symptom of this under-specification.\n\nThe KL-based uncertainty estimate is also under-described: no formula, no regularization for near-singular covariances, no details on how the two sampled distributions are compared. That makes it hard to judge whether the numbers in Table 1 are meaningful. The evaluation is qualitative throughout — summary, waterfall, dependence plots, plus a median table — with no error bars, no repeated-seed analysis, and no comparison to a simpler baseline like LIME or even random attribution. That said, the qualitative evidence is consistent with the intended story: sensor noise dominates, interactions appear in dependence plots, and the method reproduces the expected ordering in seven of eight sequences.\n\nAll of this says: promising direction, but the central claim \"can reasonably explain\" is not yet backed by quantitative validation. The zero-baseline issue is load-bearing and should be fixed before the attributions are trusted.\n\nWho gets value: robotics researchers working on uncertainty-aware perception who want a human-readable diagnostic tool. It deserves a serious referee, but with the expectation of major revision. My recommendation: send it to review, and tell the authors to include the intercept, specify the KL computation, and add repeated-seed error bars or a baseline attribution method.","headline":"A first application of kernel SHAP to ICP uncertainty that is plausible as a proof of concept but carries an unexamined zero-baseline assumption that may shift every attribution.","tokens_in":9299,"tokens_out":1475,"would_cite":false,"duration_ms":16050,"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 SHAP can attribute probabilistic ICP's pose uncertainty to sensor noise, initial pose error, or partial overlap, and in seven of eight real-world sequences sensor noise is the dominant source.","keywords":["point cloud registration","Iterative Closest Point (ICP)","uncertainty estimation","explainable AI","SHAP","kernel SHAP","sensor noise","pose uncertainty"],"falsifier":"Compute the KL divergence between two independent 100-sample sets of ICP pose estimates drawn from the same unperturbed inputs, with identical sensor noise, initial pose scale, and overlap ratio. If the resulting values are not negligible compared with the SHAP attributions, the zero-baseline assumption fails and the reported SHAP values would shift.","tokens_in":8257,"feed_emoji":"🤖","tokens_out":8219,"duration_ms":67491,"temperature":0.7,"pith_summary":"This paper argues that kernel SHAP can explain why a probabilistic ICP point cloud registration algorithm reports high pose uncertainty, by attributing that uncertainty to three named sources: sensor noise, initial pose uncertainty, and partial overlap. The authors model each source as a perturbation of the ICP input, run ICP under the eight possible perturbation coalitions, and fit a weighted linear model whose coefficients are SHAP values. Across eight real-world sequences, the attributions are mostly nonnegative, and sensor noise emerges as the dominant contributor in seven of eight cases. The paper concludes that the explanations are reasonable and human interpretable, a step toward robots that know when and why they failed.","feed_headline":"Kernel SHAP names the biggest uncertainty source in ICP alignment","feed_subtitle":"Attributing pose uncertainty to three error sources gives robots a readable failure reason.","key_machinery":"The central object is the Shapley kernel applied to three binary feature indicators that encode whether sensor noise, initial pose uncertainty, or partial overlap is present. Each coalition is mapped to a concrete perturbation magnitude via the reference value $\\{0,1,0\\}$, ICP is run to produce 100 pose samples, and the pose uncertainty is quantified as the KL divergence between the perturbed and pseudo-true 100-sample distributions. The weighted linear regression with the Shapley kernel produces the unique additive attribution satisfying local accuracy, missingness, and consistency, and its closed-form solution yields the SHAP values $\\phi_{\\text{sn}}$, $\\phi_{\\text{ip}}$, $\\phi_{\\text{po}}$ for the three uncertainty sources.","core_discovery":"The paper's central claim is that uncertainty estimates in ICP can be explained by additive feature attribution over the perturbation magnitudes of three uncertainty sources, using kernel SHAP. Given an instance with sensor noise $\\sigma$, initial pose scale $s$, and overlap removal $\\lambda$, the method samples 100 ICP pose estimates for each of the $2^3$ coalitions, computes the KL divergence between the perturbed pose distribution and a pseudo-true unperturbed distribution as the uncertainty output, and solves a weighted linear regression with the Shapley kernel to obtain attributions $\\phi_{\\text{sn}}$, $\\phi_{\\text{ip}}$, $\\phi_{\\text{po}}$. The experiments show that these attributions are mostly positive, increase with feature values, and exhibit interpretable interaction effects, such as high initial pose uncertainty overshadowing sensor noise. The authors state that the method reasonably explains the uncertainty sources, while cautioning that the explanations assess correlation, not causality.","pith_inferences":["The same attribution pipeline could be applied to learning-based pose estimators with predictive variance, as long as a sampling-based uncertainty output is available; the perturbation-to-coalition mapping would carry over directly.","A natural testable extension would be a closed-loop experiment where the robot removes the top-attributed source and measures whether the pose uncertainty drops; if it does not, the attribution is not causally faithful.","The dominance of sensor noise in the median SHAP values may partly reflect the choice of perturbation ranges, so the ranking should be read as relative to the specified perturbation regime, not as an absolute property of the scenes.","Since the paper leaves causality unresolved, SHAP values here are diagnostic summaries; converting them into causal claims would require interventional data or a structured causal model of the registration pipeline."],"forward_implications":["A robot using any uncertainty-aware ICP variant can, at run time, receive a ranked list of the uncertainty sources it should mitigate, with sensor noise at the top in most tested environments.","Because kernel SHAP is model-agnostic, the same pipeline can be attached to other ICP algorithms and to additional uncertainty sources such as under-constrained geometry without changing the explanation formalism.","In teleoperation, an operator told that sensor noise contributed most to this alignment uncertainty can choose a corrective action such as closing range, re-sensing, or denoising, rather than restarting blind.","The interaction effects shown in the dependence plots suggest that fixing the dominant source may be most effective: when initial pose uncertainty is already high, reducing sensor noise yields less benefit."],"supporting_citations":[{"why":"Supplies kernel SHAP, the explanation method used for all attributions.","marker":"[10]"},{"why":"Establishes the closed-form ICP covariance that underlies the uncertainty-source taxonomy.","marker":"[5]"},{"why":"Derives the concentrated-Gaussian perturbation model used to realize initial pose uncertainty.","marker":"[7]"},{"why":"Shows an uncertainty-aware ICP variant that the explanation method can attach to.","marker":"[6]"},{"why":"Characterizes depth-sensor noise and gives the mean bias value guiding the noise range.","marker":"[18]"},{"why":"Provides the overlap-ratio calculation used to perturb partial overlap.","marker":"[19]"},{"why":"Defines concentrated Gaussian distributions on SE(3) used for sampling perturbed initial poses.","marker":"[21]"},{"why":"Supplies the eight real-world sequences used to evaluate the explanations.","marker":"[22]"}],"fun_headline_variants":["Kernel SHAP names culprits behind ICP uncertainty","Attributing ICP uncertainty to noise, occlusion, and ambiguity","Which ICP error source drives uncertainty? Kernel SHAP answers","Explaining uncertainty in point cloud registration via SHAP","SHAP pinpoints why ICP alignment is uncertain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire attribution is anchored to the assumption that the unperturbed input has exactly zero pose uncertainty, so the SHAP bias term $\\phi_0$ is omitted; with finite 100-sample distributions, a second independent sample of the same unperturbed input will have slightly different covariance and hence nonzero KL divergence.","fun_headline_variants_meta":{"raw":{"variants":["Kernel SHAP names culprits behind ICP uncertainty","Attributing ICP uncertainty to noise, occlusion, and ambiguity","Which ICP error source drives uncertainty? Kernel SHAP answers","Explaining uncertainty in point cloud registration via SHAP","SHAP pinpoints why ICP alignment is uncertain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000374,"raw_usage":{"total_tokens":1946,"prompt_tokens":842,"completion_tokens":1104,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":1026}},"tokens_in":458,"tokens_out":1104,"duration_ms":10476,"temperature":1.0,"reasoning_tokens":1026,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:16:00.207596+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the KL divergence between two independent 100-sample sets of ICP pose estimates drawn from the same unperturbed inputs, with identical sensor noise, initial pose scale, and overlap ratio. If the resulting values are not negligible compared with the SHAP attributions, the zero-baseline assumption fails and the reported SHAP values would shift.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the closed-form ICP covariance that underlies the uncertainty-source taxonomy."},{"cited_title":"Brossard, S","cited_arxiv_id":null,"evidence_quote":"Derives the concentrated-Gaussian perturbation model used to realize initial pose uncertainty."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows an uncertainty-aware ICP variant that the explanation method can attach to."},{"cited_title":"Pomerleau, A","cited_arxiv_id":null,"evidence_quote":"Characterizes depth-sensor noise and gives the mean bias value guiding the noise range."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the overlap-ratio calculation used to perturb partial overlap."},{"cited_title":"Bourmaud, R","cited_arxiv_id":null,"evidence_quote":"Defines concentrated Gaussian distributions on SE(3) used for sampling perturbed initial poses."},{"cited_title":"Pomerleau, M","cited_arxiv_id":null,"evidence_quote":"Supplies the eight real-world sequences used to evaluate the explanations."}],"review_version":1}