{"id":"c0fdd4a4-56cb-441a-b3d9-dd24f6af417b","arxiv_id":"2608.09166","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"CaPTURe combines particle-based conformal scores with per-stratum Mondrian groups to construct feasible, trans-dimensional prediction regions for robot configurations in contact-rich motion planning.","lead":"The paper presents CaPTURe, an algorithm that builds one-step uncertainty regions for a robot's next configuration, with a finite-sample guarantee that the true future configuration is inside with a user-chosen probability, including in contact with obstacles. It is tested on a simulated marble maze and a tight-tolerance peg insertion task, reporting up to 30 percentage points higher task success than the best baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5 proves coverage per DTree leaf, not per contact stratum; Table 5's per-stratum undercoverage (e.g., Center Right Corner 88.6%) is not a theorem violation but shows the abstract's 'coverage in both contact cases' claim is not entailed.","rationale":"The reader's verdict is CONDITIONAL, and I agree with that verdict, but for a slightly different reason than the reader's stated weakest assumption. The reader emphasized Assumption 2 (exact stratum indexer) and Assumption 1 (exchangeability). Those are genuine limitations and are largely acknowledged in Section 6.3. The more subtle and unacknowledged issue is that Theorem 5's guarantee is conditional on the decision-tree leaf (the Mondrian group), whereas the paper's abstract and per-stratum validation claim coverage by contact stratum. These are logically different: a coarse stratum can be a strict subset of a leaf, and the Mondrian guarantee does not automatically transfer to subsets of a leaf. The empirical results in Table 5 are consistent with this gap: some per-map stratum estimates fall below 90% even though the aggregate Table 1 looks fine. Without per-leaf coverage or per-stratum sample sizes, the reader cannot tell whether the small per-stratum shortfalls are finite-sample noise or a systematic consequence of group/stratum mismatch. The theorem itself is not circular and the proof is a standard Mondrian CP argument, so I would not reject the paper; the concern is about whether the central practical claim is actually supported. A focused post-hoc analysis of leaf-conditional coverage and a tree variant that forces a stratum split would settle the question. Until then, CONDITIONAL remains appropriate, and my read does not move the verdict.","tokens_in":23256,"tokens_out":17404,"duration_ms":170082,"concrete_test":"On the same held-out evaluation transitions (or a fresh large set), group test cases by the fitted DTree leaf (Mondrian group) and by stratum index T(Y); compute empirical coverage with per-leaf sample sizes and 95% Clopper-Pearson intervals. If every leaf's lower bound is >= 0.90, Theorem 5 is confirmed and the Table 5 per-stratum values below 90% are explained by mixing strata within leaves; then rerun with a tree constrained to a top-level split on T(Y) to see whether per-stratum coverage reaches 0.90 without excessive volume increase. If any leaf's lower bound is < 0.90, the implementation deviates from Theorem 5 (e.g., discrete grid T or split leakage), and the central coverage claim is not validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central guarantee is conditional on the Mondrian group g(X,T(Y)) defined by the fitted decision-tree leaves (Theorem 5). The abstract, Section 6, and Tables 1/3 reinterpret this as coverage per contact stratum T(Y)=m. These conditionings differ: for a fixed test X, a leaf j may contain several strata, and Theorem 5 only asserts P{Y in C_hat | g=j} >= 1-alpha, not P{Y in C_hat | T(Y)=m} >= 1-alpha; coverage of a subset of a leaf can be below 1-alpha even when the leaf guarantee holds. The decision tree is fit to minimize score variance and is not constrained to split on T(Y), so leaves need not align with strata. Table 5 indeed shows per-map stratum coverage below 90% (Bottom Left Edge 89.5, Center Free Space 89.5, Center Right Corner 88.6), so the paper's claim that CaPTURe 'achieves the user-specified coverage requirement both in and out of contact' is not logically supported by Theorem 5; it rests only on aggregate per-stratum empirical numbers that appear to undercover in some maps. The theorem itself is a standard, correct application of Mondrian CP with a data-learned grouping; the gap is between the theorem's leaf-conditional statement and the paper's stratum-conditional claim. Because the paper never reports per-leaf coverage, counts, or confidence intervals, this mismatch is undiagnosed. The same gap affects both tasks and is the most load-bearing weakness in the paper's central narrative.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CaPTURe, a conformal prediction method for one-step robot configuration prediction in contact-rich settings. It combines probabilistic conformal prediction (particle-based scores) with Mondrian conformal prediction, where the Mondrian groups are leaves of a regression tree fitted on a random split of the calibration data, using features (state, action, future configuration stratum index). The central theoretical results are Theorem 5, which gives finite-sample leaf-conditional coverage, and Corollary 6, which gives marginal coverage by union over output strata; the construction also restricts prediction regions to the feasible C-space. The method is evaluated on two simulated tasks, a marble labyrinth and a tight-tolerance peg insertion, against four baselines, reporting per-stratum empirical coverage, region volume, and planning success rates.","tokens_in":23527,"tokens_out":12664,"duration_ms":125169,"significance":"If the guarantee is read as leaf-conditional, the theoretical core is a correct and clean application of Mondrian conformal prediction with LOCART-style data splitting: the coverage argument does not rely on fitted parameters or on the accuracy of the dynamics model, and the feasible-region construction is principled. The experimental setup is unusually thorough (on the order of 10^5 calibration and validation transitions per map, two contact-rich tasks, multiple baselines), and the reported planning improvements (up to 30% absolute success-rate gain) are practically meaningful. The weakness is that the paper's headline claim of coverage \"in and out of contact\" is per-stratum, while the theorem only guarantees coverage per DTree leaf; several per-stratum point estimates in Table 5 fall below the 90% target. The paper is therefore a solid contribution to conformal prediction for robot dynamics, but its central narrative currently overstates what is proved.","major_comments":[{"comment":"Theorem 5 establishes coverage conditional on the DTree leaf j = g(X,T(Y)), not on the contact stratum T(Y)=m, and Corollary 6 only aggregates these leaf-conditional guarantees into marginal coverage. The abstract and Section 6 nevertheless claim that CaPTURe \"achieves the user-specified coverage requirement both in and out of contact\" and use per-stratum coverage as the primary validation target. A leaf can contain several strata, so P(Y in C_hat | g = j) >= 1-alpha does not imply P(Y in C_hat | T(Y)=m) >= 1-alpha. Table 5 contains per-stratum estimates below 90% (Bottom Left Edge 89.5%, Center Free Space 89.5%, Center Right Corner 88.6%), so this is not a merely theoretical gap. Please either modify the method so the Mondrian partition refines the stratum partition (or intersect leaves with strata before computing thresholds) and prove stratum-conditional coverage, or reframe the claims and validation around leaf-conditional coverage and report the corresponding per-leaf numbers.","section":"Section 5.1, Theorem 5 and Corollary 6, versus the Abstract and Tables 1, 3, 5"},{"comment":"The empirical coverage numbers are reported only as point estimates, without confidence intervals or per-stratum (let alone per-leaf) sample counts. Since some entries fall below the nominal 90% (Table 5), the validation cannot distinguish finite-sample variability from genuine undercoverage. Please report binomial confidence intervals for every reported coverage value and give the number of test cases per stratum and per DTree leaf, because the leaf is the level at which Theorem 5 makes a claim.","section":"Section 6.1 and 6.2, Tables 1, 3, and 5"},{"comment":"The theorem treats g as a fixed grouping function, but Algorithm 1 defines g as the DTree fitted on D_bar^part_cal; Lemma 4 is stated without proof, and Theorem 5 does not explicitly state that g is fixed before the thresholds are computed from D_bar^cp_cal. Please state Theorem 5 conditionally on the DTree partition (or otherwise make the independence of g from the threshold-calibration subset explicit) and provide the proof of Lemma 4, so that the proof of Theorem 5 is formally complete.","section":"Section 5.1, Lemma 4 and Theorem 5, with Section 5.2"}],"minor_comments":[{"comment":"The text reads \"kNN = L/27\", which contradicts the following sentences and footnote 7 that indicate kNN = L/2 (8 for L=16, 4 for L=8); please correct the typo.","section":"Section 6"},{"comment":"Equation (8) defines g on X times M, but Theorem 5 and surrounding text write g(X,T(Y)); please define the composed map explicitly to avoid notational ambiguity.","section":"Section 5.1"},{"comment":"Total test cases per map are given, but not per stratum; please add per-stratum counts so the reader can interpret the point estimates.","section":"Table 5 caption"},{"comment":"The dependence of the feasibility claim on Assumption 2 (a correct, known stratum indexer T) is stated only in the limitations; because the abstract emphasizes feasible regions, this assumption should be flagged prominently in Section 3.","section":"Section 6.3"},{"comment":"Theorem 3 is cited rather than proved; a one-sentence sketch of the finite-sample quantile argument would make the review section self-contained.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The deciding issue is the gap between the leaf-conditional theorem and the stratum-conditional reporting. This is fixable either by changing the grouping to enforce stratum refinement or by re-scoping the abstract and validation to leaf-conditional coverage; the underlying Mondrian argument is sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The theoretical core is sound: Theorem 5 is a correct Mondrian conformal guarantee conditioned on the DTree leaf, and the union-over-strata construction in Corollary 6 correctly yields marginal coverage. What is genuinely new is the combination of particle-based PCP scores with a data-learned grouping over (state, action, future stratum), plus the feasible-region constraint. That combination is worth having, and the proof is standard but properly applied—the 70/30 split with a separately fit partition tree preserves exchangeability, and the kNN score is symmetric.\n\nThe experiments are extensive: hundreds of thousands of one-step transitions across six maze sections and a peg-in-hole task, with sensible Lie-algebra distances and multiple ablations. The limitation section is unusually honest about single-step guarantees and hyperparameter sensitivity.\n\nThe load-bearing gap is the leaf/stratum mismatch. Theorem 5 conditions on g(X,T(Y))=j, a DTree leaf. The tree is fit to minimize score variance and is not constrained to split on T(Y), so a leaf can contain several strata. The abstract and Section 6 reinterpret the guarantee as coverage per contact stratum, which does not follow. Table 5 shows the symptom: Center Right Corner 88.6%, Bottom Left Edge 89.5%, and Center Free Space 89.5%—all point estimates below the 90% nominal level, with no error bars. The claim that CaPTURe \"achieves the user-specified coverage requirement both in and out of contact\" is therefore supported only by aggregate numbers, not by the theorem. This is fixable: report per-leaf coverage, check whether each reported stratum is a union of leaves, or constrain the DTree to split along strata.\n\nReproducibility is also incomplete—no code, data, or DTree hyperparameters are provided. Assumption 2 (known exact stratum indexer) is stated explicitly, and Section 6.3 flags geometric uncertainty, but the exchangeability assumption's failure under deployment distribution shift is not named as a risk.\n\nThis paper is for robotics researchers doing conformal prediction with particle-based models. The core method is useful and the proof is correct. It deserves a serious referee, but the gap between the leaf-conditional guarantee and the stratum-conditional claims must be closed before publication. Send it to review with a request to fix that mismatch and provide code and data.","headline":"Correct Mondrian CP theorem and a useful particle-based region construction, but the paper's per-stratum coverage claims outrun what Theorem 5 actually proves.","tokens_in":24136,"tokens_out":1788,"would_cite":true,"duration_ms":18904,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CaPTURe proves per-contact-stratum coverage for one-step next-configuration prediction regions built from particle-based dynamics, confined to the feasible configuration space.","keywords":["CaPTURe","conformal prediction","stratified configuration spaces","contact-rich manipulation","uncertainty quantification","particle-based dynamics","Mondrian conformal prediction","one-step coverage guarantees"],"falsifier":"Run a one-step validation set with a deliberately wrong stratum indexer, for example reducing the wall-inflation radius by 2 mm after calibration so edge and corner labels shift; if any group's empirical coverage falls below $1-\\alpha$ by more than the finite-sample quantile slack, or if any predicted region contains an obstacle cell, the exact-$T$ premise fails.","tokens_in":22965,"feed_emoji":"🤖","tokens_out":12753,"duration_ms":119486,"temperature":0.7,"pith_summary":"This paper sets out to give planners a prediction region for the robot's next configuration that is guaranteed, with user-set probability $1-\\alpha$, to contain the true next configuration separately for each contact regime---free space, sliding contact, corner contact---rather than only on average. The proposed method, CaPTURe, combines Mondrian conformal prediction with particle-based dynamics models and the geometry of stratified configuration spaces. If the proof is right, the regions are both contact-aware and feasible: they adapt their size to the local uncertainty regime and are built by intersecting each candidate contact stratum with a conformal threshold, so they never include configurations that intersect obstacles. Numerical validation on a marble labyrinth and a tight-tolerance peg-insertion task shows the method meeting per-stratum coverage in every regime, while baselines undercover free space and overcover contact strata, and it improves task success by up to 30 percentage points over the best baseline.","feed_headline":"CaPTURe hits target coverage in every contact mode","feed_subtitle":"Particle-based sets keep future robot poses covered in free space, sliding, and corners—no obstacle poses.","key_machinery":"The load-bearing mechanism is Mondrian conformal prediction adapted to a finite stratified configuration space $C_{\\mathrm{feas}}=\\bigsqcup_m S_m$ with a known stratum indexer $T$: the group assignment $g(X,T(Y))$ sends each transition to a leaf of a regression tree fitted on augmented inputs $(X,T(Y))$ with the particle-score $R_i$ as target, and each leaf gets its own split-conformal threshold $\\hat{q}_j$. At inference the region for a candidate future stratum $m$ is built by thresholding the $k$-th-nearest-particle distance inside $S_m$, and the per-stratum regions are unioned. This one construction produces conditional coverage per contact regime, keeps all candidate configurations in feasible strata, and makes the set size adapt to local state- and action-dependent uncertainty.","core_discovery":"The central claim is Theorem 5: given exchangeable calibration data, a fixed group map $g(X, T(Y))$ over the prediction-time information and the future stratum index, and any symmetric nonconformity score $r$, the region $$\\hat{C}(X_{n+1}) = \\{y\\in C_{\\mathrm{feas}} : r(\\hat{f}(X_{n+1}), y) \\le \\hat{q}_{g(X_{n+1},T(y))}\\}$$ satisfies $\\Pr(Y_{n+1}\\in \\hat{C}(X_{n+1}) \\mid g(X_{n+1},T(Y_{n+1}))=j)\\ge 1-\\alpha$ for every group $j$ with positive mass. The nonconformity score is the distance from the true next configuration to its $k$-th nearest particle among $L$ samples from an approximate dynamics model, so the region is distribution-free and can represent multimodal, trans-dimensional post-contact uncertainty. Since the strata are disjoint and cover $C_{\\mathrm{feas}}$, unioning the per-stratum regions over candidate $m$ preserves marginal coverage and yields $\\hat{C}\\cap C_{\\mathrm{obs}}=\\emptyset$. The paper's experiments confirm this per-stratum coverage empirically at the 90% level in free-space, edge, and corner regimes, where non-stratum baselines undercover free space or overcover contacts. The discovery is that conformal calibration can be made contact-aware by treating the configuration space's stratified structure as part of the group assignment, so coverage and feasibility come from the same construction.","pith_inferences":["Beyond the paper: the exact-$T$ premise is the main limitation in real systems; extending CaPTURe to calibrate over an uncertain or learned stratum indexer, or to inflate strata near boundaries, would be the natural next step.","Beyond the paper: since Theorem 5 is single-step, a testable extension is to measure cumulative closed-loop coverage over $H$-step MPC rollouts and to study whether per-step conditional calibration composes along the executed trajectory.","Beyond the paper: the decision-tree partitioner can be replaced by any exchangeability-preserving partitioner of $(X,T(Y))$; denser data or domain-structured groups should tighten regions without breaking coverage.","Beyond the paper: the comparison between $k_{\\mathrm{NN}}=1$ and $k_{\\mathrm{NN}}=L/2$ suggests that smoothing the particle-density score affects planning efficiency; one could test whether larger particle budgets with smoothed scores yield smaller feasible regions at equal coverage."],"forward_implications":["Planners can use prediction regions that are trustworthy per contact regime, not just on average, because each group carries its own finite-sample $1-\\alpha$ coverage guarantee.","Because every region lies inside $C_{\\mathrm{feas}}$, uncertainty-aware planning will not be made overly conservative by obstacle-intersecting hypothetical configurations.","Any approximate dynamics model that can emit particles can be calibrated this way; no Gaussian or other parametric uncertainty representation is required.","The paper's simulations indicate that achieving per-stratum coverage improves downstream task outcomes, with up to a 30 percentage-point absolute gain in peg-insertion success over the best baseline."],"supporting_citations":[{"why":"Establishes Theorem 2 that minimum-norm distance to particle samples preserves conformal coverage, the base of CaPTURe's nonconformity score.","marker":"(Wang et al., 2023)"},{"why":"Introduces Mondrian conformal prediction and per-group coverage, the mechanism CaPTURe adapts to strata.","marker":"(Vovk et al., 2003)"},{"why":"Provides the conditional exchangeability lemma and general Mondrian coverage framework used in the proof of Theorem 5.","marker":"(Angelopoulos et al., 2024)"},{"why":"Supplies the LOCART regression-tree procedure that CaPTURe uses to partition state-action-stratum space into adaptive groups.","marker":"(Cabezas et al., 2025)"},{"why":"Supplies the k-th nearest-neighbor distance as a local particle-density score used in Equation (9).","marker":"(Loftsgaarden and Quesenberry, 1965)"},{"why":"Introduces split/inductive conformal prediction, the per-group quantile construction CaPTURe applies inside each leaf.","marker":"(Papadopoulos et al., 2002)"},{"why":"Provides the local conformal calibration baseline LUCCa and the state-action-adaptive grouping idea CaPTURe extends with stratum labels.","marker":"(Marques and Berenson, 2026a)"},{"why":"Supplies the formal definition of a finite stratified space used to define $C_{\\mathrm{feas}}$ and the stratum indexer $T$.","marker":"(Tran, 2020)"}],"fun_headline_variants":["CaPTURe: particle-based sets with provable contact coverage","Contact-aware calibration lifts robot task success by 30%","Guaranteed coverage in contact and free space via particles","Stratified conformal regions for safe autonomous interaction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee rests on calibration data having the same distribution as test data and on the contact-regime labeler being exactly right; the paper concedes in Section 6.3 that geometric uncertainty could put infeasible configurations in the regions.","fun_headline_variants_meta":{"raw":{"variants":["CaPTURe: particle-based sets with provable contact coverage","Contact-aware calibration lifts robot task success by 30%","Guaranteed coverage in contact and free space via particles","Stratified conformal regions for safe autonomous interaction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1802,"prompt_tokens":1166,"completion_tokens":636,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":782,"completion_tokens_details":{"reasoning_tokens":568}},"tokens_in":782,"tokens_out":636,"duration_ms":6772,"temperature":1.0,"reasoning_tokens":568,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:15:13.732561+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a one-step validation set with a deliberately wrong stratum indexer, for example reducing the wall-inflation radius by 2 mm after calibration so edge and corner labels shift; if any group's empirical coverage falls below $1-\\alpha$ by more than the finite-sample quantile slack, or if any predicted region contains an obstacle cell, the exact-$T$ premise fails.","supporting_citations":[{"cited_title":"Machine Learning: ECML 2002 , series =","cited_arxiv_id":null,"evidence_quote":"Introduces split/inductive conformal prediction, the per-group quantile construction CaPTURe applies inside each leaf."}],"review_version":1}