{"id":"c488ea8c-47bd-4269-967b-bf47b3d6c56f","arxiv_id":"2506.00589","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The authors introduce constrained SVGD frameworks and demonstrate collision-free planning, constrained inverse kinematics, and pose estimation with table placement constraints.","lead":"This paper proposes two frameworks, called Q and p, for adding constraints to Stein Variational Gradient Descent (SVGD), a particle-based method for approximating probability distributions. They show the frameworks on robot motion planning, arm placement, and object pose estimation, producing diverse solutions that respect constraints.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Exact feasibility is asserted but never proven; the finite-penalty augmented Lagrangian updates in Alg. 1/2 only reach h=0 in an unproved limit, and the paper's own SE(3) experiment shows the required convergence can fail.","rationale":"The paper's contribution is a framework, not a theorem paper, so absence of a proof alone need not be fatal. What makes the missing convergence result load-bearing is the specific wording of the central claim: 'exactly avoid collisions', 'exact table placement constraints', and 'without violating constraints'. For inequality constraints one could stop when all particles are feasible, but for equality constraints the augmented Lagrangian and quadratic penalty terms in Alg. 1/2 do not produce h=0 at any finite multiplier; exactness is a limit statement. SVGD itself is not guaranteed to find the global optimum of the penalized objective, and the paper's own Section V-B report that the p method failed to converge on exactly such an equality problem is direct manuscript evidence that the convergence assumption is violated in one of the three headline applications. The Q method still shows good practical behavior in its successful cases, and the EMD numbers in Section IV-A are a useful sanity check, but they do not yet establish the advertised exactness or a reliable constrained-SVGD convergence theory. The sign error in Eq. 17 reinforces that the written core update is not self-consistent, though it is secondary to the exactness gap. For these reasons, I agree with the reader's weakest-assumption analysis and with the CONDITIONAL verdict; a concrete equality-constraint convergence test would determine whether the concern lands.","tokens_in":13086,"tokens_out":5371,"duration_ms":56575,"concrete_test":"Run Algorithm 1 on the simple equality-constrained 2D problem from Section IV-A (f = x1 + x2, h = 2 - x1^2 - x2^2 = 0) with the augmented Lagrangian, recording max_i |h(q_i)| and the EMD to the rejection-sampling ground truth after every outer iteration, using the same 'until converged' stopping rule and a generous iteration budget. If max violation cannot be driven below a pre-specified tolerance (e.g., 1e-8) for all particles, or EMD stops decreasing before reaching the reported Q-method value, the advertised exact feasibility claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract promises that sampled plans 'exactly avoid collisions' and poses satisfy 'exact table placement constraints'. For both Alg. 1 and Alg. 2, the only mechanism moving particles onto the feasible set is the soft-constraint term L_theta (augmented Lagrangian or quadratic penalty) plus a projection that applies only to box bounds. At any finite outer iteration and finite penalty multiplier, an equality constraint h=0 is not enforced; particles satisfy h=0 only in the limit as the multiplier and penalty tend to infinity, and only if the inner SVGD loop converges to the global optimum of the penalized objective. No theorem, error bound, or tolerance analysis is given: the paper simply runs 'until converged' and reports total gradient steps. This assumption is load-bearing because the Q method's definition in Eq. 16 requires each particle to be exactly feasible, and the p method's target in Eq. 14 has zero mass outside the feasible set. The manuscript itself supplies evidence that the convergence assumption is not safe: Section V-B reports that on the SE(3) equality-constrained problem, the p method 'failed to converge in 25 times that many steps' with 50 particles. The displayed core update, Eq. 17, additionally has a sign error (it uses +alpha k grad f where the SVGD formula Eq. 8 requires -alpha k grad f), so the written algorithm is not even the one claimed; without released code, the discrepancy cannot be resolved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces two ways to incorporate arbitrary equality and inequality constraints into Stein variational gradient descent (SVGD). The 'Q method' restricts the variational family to particle configurations whose particles are feasible and adds a soft-constraint gradient (augmented Lagrangian, quadratic penalty, or log barrier) to the SVGD update, with projection used only for box bounds. The 'p method' redefines the target as a soft-constrained density proportional to e^{-alpha f(x)} times products of constraint terms and runs ordinary SVGD on that density, updating the constraint parameters in an outer loop. The authors illustrate both methods on a 2D problem and evaluate them on 2D trajectory optimization with collision constraints, a robot-arm inverse-kinematics/placement problem with SE(3) equality constraints, and Stein-ICP pose estimation with table constraints. They report that the Q method converges faster and that the p method can fail on low-dimensional equality constraints (Section V-B), and they claim that the resulting samples satisfy constraints exactly.","tokens_in":13488,"tokens_out":9456,"duration_ms":90840,"significance":"The proposed unification is potentially useful: it places the augmented-Lagrangian SVGD method of [31] and CVSTO [19] in a common family and gives practitioners a menu of soft-constraint mechanisms. The empirical comparison of Q versus p on three robotic tasks is informative, particularly the observation that sharing constraint gradients through the kernel (the p method) degrades on equality-constrained manifolds; this is a falsifiable, reproducible finding in principle. I did not find any fitted-constant circularity: the method is evaluated directly against the intended constrained distribution. However, the contribution is currently undermined by the sign error in Eq. (17) and by the absence of any guarantee or tolerance analysis for exact feasibility, so the advertised 'exactly avoid collisions' claim is not supported by the presented analysis.","major_comments":[{"comment":"The displayed steepest-descent term has the opposite sign from Eq. (8). For p(x) = e^{-alpha f(x)}, we have grad log p = -alpha grad f, so the SVGD functional gradient is phi* = (1/n) sum_j [ -alpha k grad f + grad k ], not (1/n) sum_j [ alpha k grad f - grad k ]. As written, Eq. (17) updates particles in the direction of increasing f and replaces the repulsive term with an attractive one. This is a load-bearing error because Eq. (17) is the paper's only explicit instantiation of the Q method for equality constraints; please correct it and state whether the experiments use the corrected sign.","section":"IV-A, Eq. (17)"},{"comment":"The definition of the Q method in Eq. (16) requires every particle to be exactly feasible, and the abstract promises collision-free and placement-constrained samples. Algorithm 1 nevertheless enforces only the box constraints by projection (lines 3 and 6); the equality and inequality constraints enter only through the gradient of the soft-constraint term L_theta. For finite penalty and multiplier values, the resulting particles are not exactly feasible, and no theorem or error bound is given for the outer loop. This gap is load-bearing because the paper's two main frameworks are characterized as producing distributions 'without violating constraints.' The authors should either prove a convergence or tolerance statement for Algorithm 1 or explicitly soften the exactness claims.","section":"IV-A, Algorithm 1 and Eq. (16)"},{"comment":"For the p method, Eq. (18) defines p_hat proportional to e^{-alpha f(x)} times product_i e^{-mu h_i(x)^2}; for any finite mu this density assigns positive probability to infeasible points and therefore does not equal the target p in Eq. (14), whose support is exactly the feasible set. The outer-loop multiplier updates are intended to close this gap, but the manuscript provides no convergence argument for the sequence of soft targets, and the min-pooling rule for shared multipliers is introduced without analysis. Since Algorithm 2 is claimed to approximate p under constraints, the limit of the soft-target sequence and the effect of the pooling choice should be analyzed, or the claims should be restricted to a soft-constraint approximation.","section":"IV-B, Eq. (18) and Algorithm 2"},{"comment":"The paper reports that, on the SE(3) equality-constrained inverse-kinematics problem, the p method with 50 particles failed to converge in 25 times the number of steps needed by the Q method, and that with 10 particles it converged only after 100,000 iterations. This is not merely a performance difference: it shows that the general claim of learning distributions 'without violating constraints' does not hold for a core problem class studied in the paper. The manuscript should either characterize when each method is guaranteed to work or remove the universal statement and present the p method as an approximate, task-dependent alternative.","section":"V-B"}],"minor_comments":[{"comment":"As typeset, the relaxed log-barrier branches do not match at g = -delta; the quadratic branch should be continuous with the log branch, for example 1/2 ((g+2 delta)/delta)^2 - 1/2 - log(delta). Please fix the typo.","section":"III-B, Eq. (13)"},{"comment":"The backtracking line search minimizes sum_j (-log p(z_j) + L_theta(z_j)) and omits the SVGD repulsive term. The authors acknowledge this, but the adjacent claim in Section IV-B that diversity is still improved during optimization should be supported empirically, since a line search that ignores repulsion can select step sizes that reduce the repulsion-driven spread.","section":"IV-A and IV-B"},{"comment":"The figure reports total gradient steps over 50 trials, but no dispersion measure or statistical test is provided; please add medians and interquartile ranges or individual trial points so the reader can assess the claimed speed advantage.","section":"V-A, Fig. 3"},{"comment":"The statement that 'the rotation about y and z are exactly zero for all particles' is inconsistent with the stated constraint that the object z-axis is vertical and with Fig. 5, which shows a wide distribution about z; the sentence should say x and y, or the constraint should be restated.","section":"V-C"},{"comment":"The title and introduction promise applications to system identification, but Section V contains no system identification experiment; either add one or adjust the title and abstract to match the presented experiments.","section":"Title and Section V"}],"recommendation":"major_revision","confidential_remarks":"The paper is clearly written and the problem is relevant. My main concerns are the sign error in Eq. (17) and the gap between the 'exact feasibility' claims and the penalty/Lagrangian machinery; both are fixable in revision, so I do not recommend rejection. I would ask the authors to release code or at least provide the corrected update, because without it one cannot tell whether the reported experiments use the correct SVGD sign. The overlap with [19] and [31] should also be sharpened in the revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTwo things to know. First, this paper delivers a useful conceptual unification: it shows that several existing constrained SVGD methods—Heiden et al.'s augmented Lagrangian, Power and Berenson's CVSTO—fit under one umbrella (the Q method), and it introduces a genuinely new variant (the p method) that puts soft constraint terms inside the target distribution and updates their weights. Second, the paper has a sign error in its core update, Eq. 17, which as written would push particles uphill, and it promises \"exactly avoid collisions\" without proving the convergence that would justify that word.\n\nThe Q method framing is the strongest contribution. The authors are honest about where their methods work and where they don't: on the equality-constrained SE(3) problem, the p method fails to converge with 50 particles, and they say so plainly. That negative result is useful. Their empirical observation that the Q method converges faster on equality constraints, and their explanation for why, is the kind of comparison practitioners actually want.\n\nThe soft spots are real but not disqualifying. The sign error in Eq. 17 contradicts the derivation in Eq. 8 and the surrounding text; it is very likely a typo, but it needs to be fixed before anyone can trust the paper. More substantively, the exact-feasibility claim is not backed by any theorem or tolerance analysis. Both algorithms rely on soft penalties with increasing multipliers, and the paper's own SE(3) experiment demonstrates that convergence can fail. The authors do not compare against the closest baselines (CVSTO, orthogonal-space SVGD), and no code is released, so the empirical claims are hard to verify or benchmark.\n\nWho should read this? Robotics researchers using SVGD for planning or perception who want a clear taxonomy of constrained variants and a practical speed comparison. It is not a theory paper, and it should not be judged as one. I would send it to a peer-reviewed venue with a request to fix the equation, add the missing baselines, and release code. With those changes it could be a solid contribution. As it stands, it is a conditional accept.\n\nRecommendation: engage with it, but require the sign fix and a direct comparison to CVSTO before acceptance.","headline":"A useful unification and a genuinely new p-method, but a sign error in Eq. 17 and unproven exact-feasibility claims make this a conditional accept.","tokens_in":13961,"tokens_out":3442,"would_cite":false,"duration_ms":30683,"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":"This paper proposes two methods, the Q method and the p method, that make Stein variational gradient descent approximate distributions whose samples all respect arbitrary constraints.","keywords":["Stein variational gradient descent","constrained variational inference","robot motion planning","inverse kinematics","point cloud pose estimation","augmented Lagrangian","penalty methods","particle-based inference"],"falsifier":"A concrete check is to take an equality-constrained problem whose constrained distribution is known in closed form (for instance, a Gaussian restricted to a circle), run the Q method with a fixed finite inner-loop budget, and measure the maximum constraint violation and the Wasserstein distance to the true constrained distribution as the penalty multiplier grows; the central claim fails if the violation or the distance does not decrease to zero. A second check would test the p method's equality-constraint failure on a one-particle system, where kernel averaging is absent; if it still fails, the difficulty is in the shared Lagrange multipliers rather than the kernel.","tokens_in":12868,"feed_emoji":"🤖","tokens_out":8382,"duration_ms":71084,"temperature":0.7,"pith_summary":"Stein variational gradient descent (SVGD) is a particle-based approximate inference method that is useful in robotics precisely because it returns many diverse solutions, but it has no built-in way to enforce constraints. The paper argues that constraints can be added generically in one of two places: inside the variational family, by allowing only feasible particle positions (the Q method), or inside the target distribution, by reshaping it so infeasible points are penalized (the p method). Both recipes reuse standard constrained-optimization machinery—augmented Lagrangian, quadratic penalty, and log-barrier terms—and both are demonstrated on robot motion planning, inverse kinematics with exact placement constraints, and point-cloud object pose estimation. The result that matters for robotics is that particles can be kept feasible while preserving the diversity SVGD is valued for, with the Q method converging faster and handling equality constraints that the p method struggles with.","feed_headline":"Two wrappers make Stein variational descent respect constraints","feed_subtitle":"Q method constrains particles; p method reshapes the target; both give collision-free robot distributions.","key_machinery":"The central object is Stein variational gradient descent, which represents an approximate distribution as a set of particles and updates each particle with a functional gradient combining a kernel-weighted pull toward high target density and a repulsive term for diversity. The machinery introduced here is the constraint wrapper: in the Q method, a soft-constraint cost $L_\\theta$ from the constrained-optimization literature contributes its negative gradient to each particle's SVGD update, box constraints are enforced by projection, and the constraint parameters $\\theta$ are updated in an outer loop once the inner SVGD loop converges; in the p method, the same $L_\\theta$ is folded into the target density as $\\hat p(x)\\propto e^{-\\alpha f(x)}e^{-L_\\theta(x)}$, so constraint information flows between particles through the kernel. The outer-inner loop structure, where an unconstrained SVGD solve alternates with an update of penalty weights or Lagrange multipliers, is what lets the soft constraints grow strong enough to approximate hard feasibility.","core_discovery":"The paper's central claim is that constrained optimization principles can be applied to SVGD in two general ways without changing SVGD's particle-based core. The Q method restricts the search to particle configurations in which every particle satisfies the constraints, then solves the constrained KL-minimization by alternating SVGD updates with soft-constraint gradient terms and a projection step for box constraints; the p method instead builds a sequence of smoothed target distributions of the form $\\hat p(x) \\propto e^{-\\alpha f(x)}\\prod_i e^{-L_{\\theta,i}(x)}$ and runs ordinary SVGD against the smoothed target, updating shared constraint parameters between inner loops. The same soft-constraint formulations (augmented Lagrangian, quadratic penalty, relaxed log barrier) serve both methods. In experiments, both approaches produce collision-free trajectory distributions, arm joint-angle distributions satisfying exact table-placement equality constraints, and table-constrained object pose distributions from point clouds, and the Q method reaches feasibility in substantially fewer gradient steps. The paper reports that the Q method preserves the underlying distribution while the p method's shared constraint gradients lose their way when the feasible set is a lower-dimensional manifold.","pith_inferences":["A natural extension the authors do not test is applying the same Q-method wrapper to other particle-based samplers (e.g., stochastic-gradient Langevin dynamics), since the recipe only assumes a gradient-based particle update and a soft-constraint cost.","The p method's failure mode suggests a diagnostic: when the constrained feasible set has codimension greater than zero, kernel-averaged soft-constraint gradients dilute the direction toward feasibility; one could predict this failure from the kernel's gradient norm restricted to the constraint surface.","If a certified guarantee is desired, the missing piece is an error bound between the constrained target $p$ of Eq. 14 and the particles after a finite number of outer iterations; without it, 'exact' feasibility rests on the empirical convergence of the inner SVGD loop.","A testable design choice is replacing box projection with a manifold projection or a kernel defined on the feasible set, which could let the Q method handle arbitrary equality constraints without relying on growing penalty weights."],"forward_implications":["Existing constrained-SVGD algorithms, including the augmented-Lagrangian form of [31], fit inside the Q method, so they can be viewed as instances of a single recipe rather than separate inventions.","The Q method can turn a deterministic constrained trajectory optimizer into a generator of diverse, collision-free plans, since independent runs of the optimizer tend to produce duplicate trajectories.","The p method can under-approximate target distributions more faithfully than Q on inequality-constrained problems, as measured by Earth mover distance on the 2D benchmark, but at the cost of more gradient steps.","On equality constraints whose feasible sets are lower-dimensional manifolds, the Q method remains practical while the p method's shared constraint gradients can fail to converge within reasonable budgets.","Both methods can be used as anytime algorithms: with many particles initialized, the number of feasible particles grows with compute, giving an anytime trade-off between diversity and feasibility."],"supporting_citations":[{"why":"Defines the SVGD particle update equation that both methods wrap with constraint terms.","marker":"[38]"},{"why":"Supplies the kernelized Stein discrepancy whose closed-form functional gradient becomes the SVGD update.","marker":"[39]"},{"why":"Earlier augmented-Lagrangian SVGD that the Q method subsumes as one of its constraint formulations.","marker":"[31]"},{"why":"Recent constrained Stein trajectory optimization method that the paper contrasts with its own constraint mechanisms.","marker":"[19]"},{"why":"Soft-constraint trajectory distributions that the p method resembles but extends by updating the constraint weights.","marker":"[17]"},{"why":"Standard reference for penalty, barrier, and augmented-Lagrangian algorithms that provide the soft-constraint terms $L_\\theta$.","marker":"[40]"},{"why":"Provides the augmented-Lagrangian cost for inequalities used in the experiments.","marker":"[41]"},{"why":"Source of the backtracking line search and projection used for step-size selection and box constraints in the Q method.","marker":"[44]"},{"why":"Stein ICP, the point-cloud pose estimation approach that the constrained state-estimation experiment builds on.","marker":"[22]"}],"fun_headline_variants":["SVGD with constraints: two routes to collision-free motion","Q and p methods: constraining SVGD for robot tasks","Constrained SVGD: two general frameworks for robot perception and planning","Two SVGD wrappers keep robot plans on the constraint manifold"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that running the inner SVGD loop to convergence and updating the penalty or Lagrange multipliers in the outer loop drives the particles to the exact feasible distribution, since no theorem or error bound is given for that convergence; the abstract's promise of exact collision avoidance and exact placement constraints depends on it.","fun_headline_variants_meta":{"raw":{"variants":["SVGD with constraints: two routes to collision-free motion","Q and p methods: constraining SVGD for robot tasks","Constrained SVGD: two general frameworks for robot perception and planning","Two SVGD wrappers keep robot plans on the constraint manifold"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001549,"raw_usage":{"total_tokens":6164,"prompt_tokens":889,"completion_tokens":5275,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":5203}},"tokens_in":505,"tokens_out":5275,"duration_ms":31652,"temperature":1.0,"reasoning_tokens":5203,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:01:57.017129+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check is to take an equality-constrained problem whose constrained distribution is known in closed form (for instance, a Gaussian restricted to a circle), run the Q method with a fixed finite inner-loop budget, and measure the maximum constraint violation and the Wasserstein distance to the true constrained distribution as the penalty multiplier grows; the central claim fails if the violation or the distance does not decrease to zero. A second check would test the p method's equality-constraint failure on a one-particle system, where kernel averaging is absent; if it still fails, the difficulty is in the shared Lagrange multipliers rather than the kernel.","supporting_citations":[{"cited_title":"Stein variational gradient descent: A general purpose bayesian inference algorithm,","cited_arxiv_id":null,"evidence_quote":"Defines the SVGD particle update equation that both methods wrap with constraint terms."},{"cited_title":"A kernelized stein discrepancy for goodness-of-fit tests and model evaluation,","cited_arxiv_id":null,"evidence_quote":"Supplies the kernelized Stein discrepancy whose closed-form functional gradient becomes the SVGD update."},{"cited_title":"Probabilistic Inference of Simulation Parameters via Parallel Differentiable Simulation","cited_arxiv_id":"2109.08815","evidence_quote":"Earlier augmented-Lagrangian SVGD that the Q method subsumes as one of its constraint formulations."},{"cited_title":"Constrained stein variational trajectory optimization,","cited_arxiv_id":null,"evidence_quote":"Recent constrained Stein trajectory optimization method that the paper contrasts with its own constraint mechanisms."},{"cited_title":"Stein variational model predictive control,","cited_arxiv_id":null,"evidence_quote":"Soft-constraint trajectory distributions that the p method resembles but extends by updating the constraint weights."},{"cited_title":"A novel augmented lagrangian approach for inequalities and convergent any-time non-central updates,","cited_arxiv_id":null,"evidence_quote":"Provides the augmented-Lagrangian cost for inequalities used in the experiments."},{"cited_title":"Projected stein variational gradient descent,","cited_arxiv_id":null,"evidence_quote":"Source of the backtracking line search and projection used for step-size selection and box constraints in the Q method."}],"review_version":1}