{"id":"751d5286-5948-4cb1-a801-1e41ae3f81c6","arxiv_id":"2504.19771","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A systematic benchmark shows ADMM-based dual solvers outperform projected Gauss-Seidel style solvers on ill-conditioned rigid-body systems with kinematic loops, especially under large mass ratios.","lead":"This report benchmarks nine algorithms for simulating rigid mechanisms with joints, contacts, and kinematic loops, on a suite of twelve test systems from toy boxes to full Audio-Animatronics figures. It finds that an ADMM-based solver handles large mass ratios and redundant constraints far more reliably than the splitting methods used in most physics engines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Boxes-Fixed comparison may not be apples-to-apples: ADMM-NCP uses a regularized dense solve and proxy termination residuals while the splitting solvers use unregularized block iterations with unreported SOR parameters, so the mass-ratio conclusion may not be about solver class as such.","rationale":"The reader's weakest assumption concerns whether the PGS/NBGS/RaiSim proxies are representative of production solvers. I agree that this matters, but the more direct load-bearing condition for the Boxes-Fixed conclusion is the symmetry and completeness of the benchmark itself: ADMM-NCP's regularized linear solve, its proxy termination residuals, and the unreported SOR settings for the splitting solvers make it difficult to tell whether the result is about the solver family or about particular configurations. The paper is otherwise careful, with multiple independent runs, explicit performance metrics, and an honest discussion of exclusions and limitations, but the central comparison needs a fairer and more reproducible setup before the strong 'splitting solvers are unsuitable above mass ratio 1000' claim can be accepted. This does not move the verdict away from CONDITIONAL, so the reader's verdict remains unchanged.","tokens_in":57760,"tokens_out":9410,"duration_ms":104167,"concrete_test":"Re-run the Boxes-Fixed mass-ratio sweep in Fig. 18 with: (a) a documented sweep over omega0 in {0.5, 0.8, 1.0, 1.2, 1.5}, omega_min, and gamma for Alg. 5; (b) a PROX/SORProx solver as an additional splitting baseline; and (c) a wall-clock budget matched to ADMM-NCP's per-iteration cost, using the actual NCP residuals (156)-(158) as the shared stopping criterion. If any splitting configuration converges for mass ratios above 1000, or reaches ADMM-NCP-level residual accuracy within the same time budget, then the claim that splitting solvers are unsuitable for high mass ratios needs to be qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that splitting-based solvers become unsuitable at mass ratios above roughly 1000 is supported mainly by the Boxes-Fixed experiment and the mass-ratio grid in Sec. VIII-C, Figs. 18 and 39-40. For that evidence to support the claim, the solvers must be compared at equivalent effort and near their best configurations. Three related conditions are not established. First, ADMM-NCP in Alg. 6 solves each iteration with the regularized dense system D + (eta + rho) I (line 6; eta = 1e-6, rho adapted by (171)/(172)), while the splitting solvers in Alg. 5 use unregularized block projections with no equivalent global preconditioner. The SOR parameters omega0, omega_min, and gamma that control Alg. 5 convergence are not reported in Table I, and the PROX/SORProx baseline is omitted due to time considerations as stated in Sec. VI-E. Second, ADMM termination uses proxy residuals (170a-c) defined on ADMM iterates, not the actual NCP residuals (156)-(158) used for the splitting solvers; if the proxies are looser, ADMM can be declared converged where a PGS variant with the same tolerance would not. Third, the benchmark uses equal Nmax = 10^4, which is not necessarily an equal computational budget; if the dense ADMM solve costs more per iteration, fixed-iteration comparisons overstate practical robustness. These are not accusations about the results, but they are exactly the load-bearing conditions for the headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops and evaluates dual solvers for the time-stepping forward dynamics of constrained rigid multi-body systems with kinematic loops, formulated as Nonlinear Complementarity Problems (NCPs) and their convex relaxations. The authors derive the dual NCP/NSOCP from an augmented-Lagrangian/proximal perspective, describe a set of projective splitting solvers (PGS variants, NBGS, RaiSim variants) and two ADMM-based solvers (ADMM-CCP and ADMM-NCP), and benchmark them on a suite of twelve problems ranging from toy examples to robotic and Audio-Animatronics systems. The central empirical claim is that ADMM-NCP is the most robust and accurate solver for ill-conditioned systems, with the Boxes-Fixed experiment showing that splitting-based solvers fail to propagate constraint reactions at mass ratios above roughly 1000, while ADMM-NCP remains accurate and converges on samples where no other solver converges.","tokens_in":58070,"tokens_out":2672,"duration_ms":30211,"significance":"If the comparison is accepted, the paper makes a practically important point: the splitting-based first-order solvers used in the majority of physics engines are not reliable for closed-loop mechanisms with large mass ratios, whereas a globally coupled ADMM with spectral penalty adaptation can be. The derivation of the dual problem and the proximal framework is standard and clearly presented, and the extensive benchmark suite, performance profiles, and explicit statement of solver configurations are genuine strengths. The paper also usefully documents artifacts such as non-minimum-norm internal forces and the effects of constraint softening and stabilization. The headline claim, however, rests on the fairness of the solver comparison, and on that point the manuscript currently leaves load-bearing gaps: termination criteria differ between solver classes, SOR parameters are not reported, and equal iteration counts are not shown to be equal computational budgets.","major_comments":[{"comment":"The comparison uses equal Nmax = 10^4 for all solvers, but the per-iteration cost is not equal: ADMM-NCP (Alg. 6, line 6) solves a dense regularized system (D + (eta + rho)I) each iteration, while the splitting solvers (Alg. 5) perform block projections. The reported 'Mean Iteration Time' metric (Sec. VII-D) is collected but is not used as the budget for the performance profiles; a solver that does more work per iteration is favored by an iteration-count comparison. Please report wall-clock time (or flop-equivalent cost) for the headline experiments and show that the conclusions are unchanged under equal-time budgets.","section":"Sec. VI-E and Sec. VIII-A"},{"comment":"ADMM-NCP terminates on proxy residuals defined on the ADMM iterates (r_p = x_i - y_i, r_d = eta(x_i - x_{i-1}) + rho(y_i - y_{i-1}), r_cp = [x_i^T z_i]), whereas the splitting solvers terminate on the actual NCP residuals (156)-(158). If the proxy residuals are systematically looser than the true residuals, the robustness advantage of ADMM-NCP in Figs. 15-19 could partly reflect a more permissive stopping rule. Please report, for the same samples, the true NCP residuals (156)-(158) achieved by each solver at termination, or otherwise justify that the proxy residuals are equivalent for these problems.","section":"Sec. VI-F, Eqs. (170a-c)"},{"comment":"The SOR parameters omega0, omega_min, and gamma used in Alg. 5 are not reported, and the PROX/SORProx baseline is omitted 'purely due to time considerations' (Sec. VI-E). Since the SOR relaxation factor strongly affects convergence of PGS-type methods, the comparison may not reflect the best achievable behavior of the splitting class. Please provide the parameter values used for each solver, and ideally a sensitivity study over the relaxation parameters, before drawing the broad conclusion that splitting-based solvers are unsuitable above mass ratio 1000.","section":"Sec. VI-E and Table I"},{"comment":"The mass-ratio conclusion is drawn from a single problem family in which the two bodies are connected by a fixed joint and only ADMM-NCP converges. The paper itself notes that ADMM-NCP is 'more sensitive to the absolute scale of the heaviest body' and that its convergence depends on the spectral radius rho(D) rather than only the condition number. This is a load-bearing qualification: the claim that splitting solvers fail above mass ratio ~1000 needs to be demonstrated across several problem classes (including the Fourbar mass-ratio run, Fig. 47) with reporting of both the mass ratio and the condition/spectral data, and with the solver-class comparison made at equivalent effort and equivalent termination residuals. As written, the evidence supports the weaker statement that, on these particular problems and with these particular configurations, the splitting solvers tested did not converge while ADMM-NCP did.","section":"Sec. VIII-C, Boxes-Fixed and Fig. 18"}],"minor_comments":[{"comment":"The abstract contains 'particularly interest' and 'aggravatingly over the complete set'; both should be corrected (e.g., 'particularly interested' and 'aggregately').","section":"Abstract"},{"comment":"Several figure labels contain typos: 'Boxes-on-Plane' in Fig. 13 should be 'Box-on-Plane', and 'PGS-CPP' in Fig. 14 should be 'PGS-CCP'.","section":"Fig. 13 and Fig. 14 captions"},{"comment":"The sentence 'An final curated run' should read 'A final curated run'.","section":"Sec. VIII-C, Fourbar"},{"comment":"References to 'Baumgarter' should be 'Baumgarte' to match the standard spelling in the literature.","section":"Sec. V-C"},{"comment":"The paper does not state whether the implementation, benchmark definitions, or raw data will be released; given that the experimental comparison is the core contribution, a code/data availability statement would substantially improve reproducibility.","section":"Sec. VIII-A"}],"recommendation":"major_revision","confidential_remarks":"The paper is a well-organized technical report with a standard derivation and a very extensive benchmark. My main reservation is that the headline claim about the superiority of ADMM-NCP over splitting-based solvers is only as strong as the fairness of the comparison, and the manuscript currently does not establish equal computational effort or equal termination criteria across solver classes. I would not reject because the issues are fixable within the paper's scope: report wall-clock and per-iteration costs, report SOR parameters, and re-evaluate with common termination residuals. I also note that the paper omits its own PROX/SORProx baseline and excludes some solvers from some experiments due to 'erroneous behavior' (Sec. VIII-C); while this is not disqualifying, the exclusions should be documented more prominently and the missing baseline noted in the conclusions. The journal may also want to consider whether a purely empirical benchmark without code/data release meets its reproducibility standards."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a solid empirical benchmark of dual solvers for constrained rigid-body dynamics with kinematic loops. The genuinely new piece is the systematic demonstration that splitting-based solvers (PGS, NBGS, RaiSim variants) stop converging around mass ratios of ~1000 on loop-closed systems, while an ADMM-style NCP solver keeps going. That's a useful, actionable finding for anyone building or choosing a physics engine. The survey and the maximal-coordinate adaptation of ADMM-NCP to include joints and limits are also original and well presented. The math sections are careful and standard.\n\nBut the headline comparison is not fully apples-to-apples. The stress-test note is right: ADMM-NCP uses a regularized dense solve and proxy termination residuals (170a-c), while the splitting solvers use unregularized block projections with SOR parameters that are never reported. Same Nmax does not mean same computational effort, and a shared tolerance on different residual definitions doesn't mean the same convergence criterion. The paper is transparent about this, but it leaves the central claim weaker than the prose suggests. On top of that, some solvers are excluded after initial testing (ADMM-CCP in Boxes-Fixed; PGS-NCP and ADMM-CCP in Fourbar), which introduces selection bias. And no code or data is released, so the benchmark cannot be rerun or inspected. These are fixable, but they are load-bearing for the claim.\n\nI should also note the paper handles the missing PROX/SORProx baseline honestly by saying it was omitted for time; that is fine as a limitation but means the comparison misses a theoretically founded splitting solver.\n\nOverall: this is a serious piece of work, worth engaging. The empirical result on mass-ratio thresholds is likely to be true in broad strokes, but the paper overclaims when it recommends ADMM-style solvers for closed-loop mechanisms based on iteration counts alone. The right fix is to release code and data, report the SOR parameters, and add at least a wall-clock or equal-time comparison for the Boxes-Fixed mass-ratio grid.\n\nWho this is for: people working on contact-rich simulation, especially those building or tuning physics engines for robots or animated characters. It deserves a serious referee and will likely be useful after revision.","headline":"A substantial, honest benchmark with a suggestive mass-ratio finding, but the ADMM-vs-splitting comparison has methodological loose ends and no released artifacts.","tokens_in":58603,"tokens_out":4351,"would_cite":false,"duration_ms":40897,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["70E55","70F40","90C33"],"pacs":[],"model":"deepseek-v4-flash","headline":"Proximal ADMM is the most robust dual solver for constrained rigid-body simulation with kinematic loops.","keywords":["constrained rigid-body dynamics","kinematic loops","nonlinear complementarity problem","second-order cone programming","alternating direction method of multipliers","projected Gauss-Seidel","frictional contact simulation","ill-conditioned forward dynamics"],"falsifier":"Re-run the identical curated Boxes-Fixed sample sequence using the actual production implementations of the splitting-based solvers (their native projectors, relaxation schedules, and termination criteria) instead of the in-house proxies; if any of them converges to ADMM-NCP-level residuals on a substantial fraction of the high-mass-ratio samples, the paper's central claim is falsified.","tokens_in":57513,"feed_emoji":"⚙️","tokens_out":9908,"duration_ms":93078,"temperature":0.7,"pith_summary":"This paper claims that for simulating constrained rigid-body systems with kinematic loops, large mass ratios, and frictional contacts, an ADMM-based solver for the dual forward-dynamics problem is substantially more robust and accurate than the projective splitting solvers used in most physics engines. The authors formulate the time-stepping problem in maximal coordinates as a nonlinear complementarity problem, add constraint stabilization and softening, and compare eight solver variants on a benchmark suite ranging from toy problems to full-scale closed-loop mechanisms. In the Boxes-Fixed experiment with a mass ratio of $10^4$, ADMM-NCP was the best solver on 100% of the samples across all metrics and was the only solver able to converge on some samples, while the splitting solvers underestimated joint reactions and let constraints drift. If the claim is correct, physics engines for closed-loop mechanisms and high inertial disparity should move to ADMM-style dual solvers rather than projected Gauss-Seidel variants.","feed_headline":"ADMM solver beats splitting methods on stiff closed-loop dynamics","feed_subtitle":"In benchmarks, only ADMM converged on the hardest high-mass-ratio cases; splitting solvers drifted.","key_machinery":"The load-bearing object is the dual forward-dynamics NCP: find $\\lambda$ in the composite cone $K$ such that the augmented constraint velocity $\\hat v(\\lambda)$ lies in $K^{*}$ and is complementary to $\\lambda$. The problem data is the Delassus matrix $D = J^{T} M^{-1} J$ and the free velocity $v_f$; the nonlinearity comes from the De Saxcé correction $\\Gamma(v^+(\\lambda))$, which enforces maximal dissipation and is fixed iteratively to obtain a convex second-order cone program. ADMM-NCP solves that convexified problem with proximal ADMM, whose key device is an extra proximal term $\\eta \\|x - x^{-}\\|^2$ added to the Delassus quadratic, making the regularized system strictly convex without biasing the solution; it then projects globally onto $K$ and updates the dual variable. The splitting solvers instead apply block-wise projected SOR with local projectors, either simple Euclidean cone projections or per-contact QCP solvers.","core_discovery":"On the paper's own terms, the central discovery is empirical: a proximal ADMM applied to the NSOCP transcription of the dual forward dynamics is the most robust and accurate solver among those compared, especially when the Delassus matrix is ill-conditioned by hyperstatic kinematic loops and inertial disparity. The evidence is the Boxes-Fixed mass-ratio grid, where splitting-based methods degrade once the mass ratio exceeds roughly 1000, while ADMM-NCP keeps converging except at very high absolute mass scales tied to the spectral radius of $D$. ADMM's global simultaneous projection avoids the ordering bias of per-contact iteration, yields minimum-norm contact reactions with negligible internal coupling, and can still render usable forces even when it does not converge within the iteration budget.","pith_inferences":["Editorial inference: if the empirical ranking holds up, the engineering bottleneck shifts from crafting per-contact local projectors to building fast linear solvers and preconditioners for the regularized Delassus matrix, because ADMM turns each iteration into one dense linear solve plus a global projection.","Editorial inference: the same dual ADMM should extend naturally to compliant, deformable, or viscoelastic contact models and to batched GPU simulation, where the global Delassus solve can be amortized across many problem instances.","Editorial inference: the paper's mass-ratio grid suggests a minimal robustness benchmark — a two-body fixed-joint drag scenario sweeping the mass ratio from 10 to $10^6$ — that could be adopted as a standardized stress test for any new forward-dynamics solver.","Editorial inference: a natural testable follow-up is whether the ADMM advantage persists at larger time-steps or with higher-order integrators, since the current study fixes semi-implicit Euler and the accuracy gap may narrow or widen with integration order."],"forward_implications":["Closed-loop mechanisms with passive joints and mass ratios above roughly 1000 should be simulated with an ADMM-based dual solver rather than projected Gauss-Seidel to avoid constraint drift and force underestimation.","ADMM-NCP produces minimum-norm contact reactions with virtually no internal coupling, so contact force distributions on resting and sliding bodies are physically plausible without per-contact ordering heuristics.","Constraint stabilization and softening improve constraint satisfaction for all solvers but do not overturn the relative ranking; the NCP's complementarity structure can be preserved under both augmentations.","ADMM-NCP convergence depends more on the spectral radius of the Delassus matrix than on its condition number, so preconditioning the Delassus system is the natural next lever for making the method faster.","Early stopping based on objective-function improvement drastically cuts iteration counts but costs multiple orders of magnitude in accuracy, so it is not a free performance win."],"supporting_citations":[{"why":"Supplies the proximal ADMM-NCP algorithm that is the paper's winning solver, including the proximal regularization term.","marker":"[32]"},{"why":"Provides the ADMM-CCP baseline and the contact-model comparison methodology the evaluation builds on.","marker":"[38]"},{"why":"Supplies the PMDP local QCP projector and the double-sided contact stabilization bias used in the implementation.","marker":"[56]"},{"why":"Supplies the bisection-search local QCP projector used by the nonlinear-block splitting solver variants.","marker":"[40]"},{"why":"Introduces the bi-potential method and the De Saxcé correction that defines the nonlinear complementarity structure being solved.","marker":"[30]"},{"why":"Introduces the NSOCP formulation of the dual problem and the successive-approximation approach for the De Saxcé term.","marker":"[31]"},{"why":"Supplies the constraint-softening regularization and objective-based early-stopping criteria adapted for the comparison.","marker":"[12]"},{"why":"Provides the benchmark methodology, residual metrics, and performance-profile analysis used to compare solvers.","marker":"[7]"}],"fun_headline_variants":["ADMM wins on stiff closed-loop contact dynamics","Proximal ADMM beats splitting in hyperstatic loop benchmarks","ADMM robust where splitting solvers drift on high mass ratios","Benchmarks show ADMM superior for hard-contact loops","ADMM converges where splitting methods fail in stiff loops"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes the re-implemented splitting solvers faithfully represent the production physics-engine solvers they stand in for; if their local contact projections, relaxation schedules, or termination checks are unrepresentative, the measured superiority of ADMM-NCP could be an artifact of the proxies rather than a property of the methods.","fun_headline_variants_meta":{"raw":{"variants":["ADMM wins on stiff closed-loop contact dynamics","Proximal ADMM beats splitting in hyperstatic loop benchmarks","ADMM robust where splitting solvers drift on high mass ratios","Benchmarks show ADMM superior for hard-contact loops","ADMM converges where splitting methods fail in stiff loops"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1363,"prompt_tokens":963,"completion_tokens":400,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":321}},"tokens_in":579,"tokens_out":400,"duration_ms":3769,"temperature":1.0,"reasoning_tokens":321,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:43:37.249741+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the identical curated Boxes-Fixed sample sequence using the actual production implementations of the splitting-based solvers (their native projectors, relaxation schedules, and termination criteria) instead of the in-house proxies; if any of them converges to ADMM-NCP-level residuals on a substantial fraction of the high-mass-ratio samples, the paper's central claim is falsified.","supporting_citations":[{"cited_title":"Contact models in robotics: A comparative analysis,","cited_arxiv_id":null,"evidence_quote":"Provides the ADMM-CCP baseline and the contact-model comparison methodology the evaluation builds on."},{"cited_title":"The maximum dissipation principle in rigid-body dynamics with inelastic impacts,","cited_arxiv_id":null,"evidence_quote":"Supplies the PMDP local QCP projector and the double-sided contact stabilization bias used in the implementation."}],"review_version":1}