Pith. sign in

REVIEW 3 major objections 5 minor 38 references

By decoupling object targets from hand contacts, a dexterous grasp synthesizer can optimize dynamic stability while keeping penetration exactly zero.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-02 18:10 UTC pith:2L62AJGL

load-bearing objection Genuine ADMM-based decoupling that likely improves simulated grasping, but the headline gain is not fully evidenced until baselines are re-run, variance is reported, and the g=0 quality proxy is validated. the 3 major comments →

arxiv 2603.13832 v2 pith:2L62AJGL submitted 2026-03-14 cs.RO

GraspADMM: Improving Dexterous Grasp Synthesis via ADMM Optimization

Liangwang Ruan , Jiayi Chen , He Wang , Baoquan Chen This is my paper
classification cs.RO
keywords dexterous grasp synthesisADMMforce closuregrasp quality metrickinematic feasibilitydynamic stabilityrobot manipulationphysics simulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that the long-standing trade-off between dynamic stability and kinematic feasibility in dexterous grasp synthesis can be resolved by treating the target contact points on the object and the actual contact points on the hand as separate variables, coupled through an ADMM (alternating direction method of multipliers) loop. The object targets are updated to minimize a force-closure grasp quality metric, while the hand pose is adjusted in a physics simulator to reach those targets without penetration. On a benchmark of over 5,600 objects, the approach reports a nearly 15 percentage-point improvement in grasp success rate over the prior simulator-based baseline, zero penetration depth, and roughly a doubling of success rate in type-aware synthesis. If correct, this shows that active optimization of dynamic grasp metrics is compatible with exact collision resolution, and that the earlier pipeline's fixed-contact refinement was the limiting factor.

Core claim

The paper's claim is that the fixed contact points used by earlier simulator-based refinement are the bottleneck for dexterous grasp quality. By reformulating the refinement as an ADMM optimization, the authors separate the object-side contact targets from the hand-side contact locations. Alternating three steps—gradient descent on the object points projected onto the mesh, forward physics simulation of the hand driven by transposed-Jacobian virtual springs, and a dual-variable update—the pipeline actively drives the force-closure quality metric while strictly maintaining penetration-free configurations. The reported outcome is a substantially higher grasp success rate under six-wrench dynam

What carries the argument

The central object is the augmented Lagrangian of the hand-object contact constraint p_h = p_o, with separate variables p_o (target object contact points) and p_h (actual hand contact points). The penalty parameter rho balances the force-closure metric e(p_o) against the geometric matching term. The mechanism works by allowing the two subproblems to be solved with heterogeneous tools: gradient descent with surface projection for the object targets, and forward simulation with exact collision handling for the hand pose. A scaled dual variable lambda and an annealing reset keep the loop stable.

Load-bearing premise

The paper assumes that the force-closure residual e computed with a single zero gravity vector is a reliable proxy for the dynamic stability measured by the six-wrench physics simulator success test, and it provides no evidence that lowering e raises the success rate.

What would settle it

Compare, on a random set of optimized grasps, the computed quality metric e against the binary outcome of the six-wrench simulator test; if the correlation is near zero or negative, the proxy assumption fails and the reported success-rate gains cannot be attributed to the ADMM objective. Alternatively, disabling contact sliding (rho approaching infinity) should reproduce the baseline's success rate; if it does not, some other part of the pipeline is responsible.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A grasp synthesizer can now simultaneously maximize force-closure stability and maintain zero penetration, removing the post-hoc filtering step in prior pipelines.
  • The decoupling allows local sliding of contact points, which increases the effective reachability of the hand and yields a higher proportion of successful grasp attempts.
  • The same ADMM formulation applies to any differentiable grasp quality metric, suggesting it can be swapped for task-oriented metrics.
  • Under low friction, the method still produces stable multi-fingered grasps, pointing toward physical robustness in slippery real-world conditions.
  • The gain in success rate relative to the fixed-contact baseline confirms that contact-point rigidity, not sampling or initialization, was the key limiting factor.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The g=0 simplification in the force-closure QP is left unvalidated against the six-wrench test; if lower e does not predict better dynamic success, the reported gains may instead come from the extra hand mobility (contact sliding) rather than from optimizing the stated stability metric.
  • The ADMM decoupling idea could transfer to other high-DoF contact-rich problems, such as foot placement for legged locomotion or in-hand manipulation, where one part of the contact is geometry-driven and the other is dynamics-driven.
  • Because the pipeline is agnostic to the metric, it could incorporate learned stability predictors or wrench-space objectives, turning the optimizer into a policy that produces grasps tailored to a downstream task.
  • The reported speed of roughly 19 samples per second suggests that with a GPU-native simulator, the pipeline could become a real-time grasp generator for online perception-to-grasp systems, though the paper itself does not demonstrate this.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents GraspADMM, a dexterous grasp synthesis framework that replaces Dexonomy's fixed-contact simulator-based refinement with an ADMM optimization. The method decouples target contact points on the object surface from actual hand contact locations. One subproblem updates object target points by gradient descent on a force-closure grasp quality metric; the other updates the hand pose and contacts via MuJoCo forward simulation with transposed-Jacobian virtual springs, thereby enforcing exact collision-free kinematics; a dual variable enforces consistency. Experiments on 5,697 DexGraspNet objects report a 74.6% vs. 60.5% grasp success rate over Dexonomy for type-unaware synthesis while maintaining zero penetration depth, and large gains in type-aware synthesis and low-friction robustness. The central claim is that the ADMM decoupling actively improves dynamic stability while preserving kinematic feasibility, unlike Dexonomy's post-hoc filtering.

Significance. If the reported results hold, GraspADMM is a practically valuable contribution to dexterous grasp synthesis. The conceptual advance is meaningful: replacing a fixed-contact kinematic refinement with an alternating scheme that optimizes a force-closure metric while using exact collision simulation is a clean way to avoid both penetration-prone analytic contact models and unstable post-hoc filtering. The strengths include the scale of the benchmark (5,697 objects with multiple sizes), the ablations over the key penalty parameter rho, and the low-friction stress test showing consistent gains in a regime where prior methods degrade severely. The evaluation is grounded in a standard MuJoCo six-wrench success criterion rather than only in the optimized metric. However, confidence in the quantitative claims is reduced by two issues: the optimized grasp-quality proxy (Eq. 6 with g=0) is not validated against the six-wrench test, and the main type-unaware baseline numbers are quoted from Dexonomy's paper rather than re-run in the same environment. These issues are fixable with additional experiments and do not invalidate the method's potential.

major comments (3)
  1. [Sec. III-B, Eq. (6); Sec. IV-A] The optimization objective e in Eq. 6 is computed from a force-closure QP with g=0, following [36]. The evaluation success criterion is different: a grasp must resist six external wrenches in MuJoCo. The paper states that the g=0 simplification is 'without compromising grasp quality' but provides no evidence that a lower e is predictive of six-wrench survival. This is load-bearing because e is exactly the term being minimized in Eq. 10; the headline GSR gain (74.6 vs. 60.5) could in principle come from kinematic side-effects of the decoupled update (e.g., additional contact links or sliding contacts) rather than from genuine stability optimization. I request a direct validation: e.g., report the correlation between the optimized e and MuJoCo success on a subset, or replace the g=0 objective with the sum of residuals over the six test wrenches and compare GSR. Without this, the mechanism
  2. [Sec. IV-B, Table II] The type-unaware baseline numbers for Dexonomy are not re-run in the current pipeline but 'directly come from the paper of Dexonomy.' For the paper's central quantitative claim, this is problematic: the comparison is between numbers produced by the authors' own pipeline (with potentially different code, hardware, or MuJoCo version) and numbers reported in a prior paper. The concern is amplified by Sec. IV-C, where the Dexonomy numbers 'are slightly higher than those reported in the original paper due to minor implementation differences and object randomization'—this indicates that implementation details materially affect results. Please re-run Dexonomy under the identical evaluation harness, at least on a representative subset, and report full-table results with error bars or confidence intervals. Currently no variance information is given for any GSR number.
  3. [Sec. III-C, Algorithm 1] The dual-variable reset threshold (||lambda_i|| > 2e-2) and the maximum iteration count K are ad hoc and their values are not specified in the text or the algorithm. The annealing strategy appears to be essential for the method's behavior (Fig. 6), but no sensitivity study is provided for the threshold or K. Since the overall pipeline is not guaranteed to converge in this nonconvex, nonsmooth setting, the practical behavior may be substantially determined by these heuristics. Please report the chosen K and threshold, and provide an ablation or at least a convergence diagnostic (e.g., the primal residual ||p_h - p_o|| over iterations).
minor comments (5)
  1. [Sec. II-B / Eq. (8)] The augmented Lagrangian expression L = e + rho*lambda^T(p_h - p_o) + rho/2 ||p_h - p_o||^2 appears to mix unscaled and scaled dual forms. Since the update in Eq. (13) uses the scaled dual convention, the notation should be clarified to avoid confusion.
  2. [Sec. III-C / Algorithm 1] The step size alpha in line 6 and the maximum iterations K are not defined in the text or the algorithm listing. Please state their values and how they were chosen.
  3. [Table II] The table formatting for the Dexonomy row appears to run entries together ('0.210.034.2'); this should be corrected.
  4. [Abstract / Sec. IV-C] The abstract claims 'roughly a 100% relative improvement' in type-aware synthesis. From Table III, the relative gains are about 91%, 83%, and 58% for Power, Intermediate, and Precision, respectively. 'Roughly 100%' overstates the average; please either report the exact average or soften the wording.
  5. [Sec. IV-F] The time comparison is useful, but the 'successful grasps per second' argument would be stronger with an explicit computation using the measured GSR and throughput numbers.

Circularity Check

1 steps flagged

Minor circularity: key hyperparameter ρ is tuned on the same GSR benchmark used for the headline claim; the ADMM derivation itself is not circular.

specific steps
  1. fitted input called prediction [Sec. IV-E (Ablation Study), Table IV; final row 'Ours (ρ=10^3)']
    "Setting a moderate ρ, like 10^2 and 10^3, achieves overall high GSR and OSR, but ρ=10^4 is too large, causing performance drops. ... Ours (ρ=10^3): GSR 74.6"

    The penalty parameter ρ is selected by directly comparing GSR and OSR values on the same benchmark (5,697 DexGraspNet assets scaled across six sizes) that is later used to report the headline type-unaware GSR of 74.6% (Table II). Thus the reported success rate is not an independent prediction of the method's performance, but the maximum over a ρ sweep on the evaluation set. The ablation table doubles as model-selection: the chosen ρ (10^3) was selected because it yields the highest GSR, so the 'nearly 15% absolute improvement' is partially self-fulfilling. However, the improvement is robust across moderate ρ (ρ=10^2 gives 73.5), so the effect is not entirely forced.

full rationale

The core ADMM formulation (Eqs. 7–13) is a mathematical optimization method that does not, by construction, reduce to the evaluation metric. The objective e (Eq. 6) is a standard force-closure residual, computed for g=0 following [36], and is different from the six-wrench MuJoCo success test used in evaluation. No equation equates e with grasp success rate (GSR); the claimed improvements are empirical measurements on an external benchmark (DexGraspNet assets) at the chosen hyperparameter ρ=10^3. The only concrete circularity-like step is the tuning of ρ on the same GSR and OSR metrics used to report the headline result (Sec. IV-E, Table IV). This is a statistical selection effect rather than a definitional reduction, and it is mitigated by the fact that the method outperforms the baseline over a range of ρ values (e.g., ρ=10^2 gives GSR 73.5 vs. Dexonomy's 60.5). Self-citations to Dexonomy and BODex are present but not load-bearing: the baseline comparison uses published numbers (or identical initialization in type-aware experiments), and no uniqueness theorem is imported. The assumed validity of the g=0 proxy is an unvalidated approximation, but that is a correctness risk, not a circularity. Overall, the derivation chain is self-contained apart from the minor hyperparameter-selection issue, so the score is low.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 0 invented entities

The central claim borrows the sampling/template prior from Dexonomy and depends on a simulator-based hand solver and a proxy grasp-quality metric. The genuinely new ADMM split is a real algorithmic contribution, but its success hinges on several unspecified parameters and on the unvalidated g=0 quality proxy.

free parameters (8)
  • rho (ADMM penalty stiffness) = 1e3
    Balances the grasp quality metric against the contact-matching penalty; selected via ablation on the same benchmark that reports the headline result (Table IV).
  • lambda_min = not reported
    Minimum total normal force in the QP constraint (Eq. 5), preventing trivial zero-force solutions; value not reported.
  • mu (friction coefficient in QP metric) = not reported
    Friction coefficient used in the objective's Coulomb cone (Eq. 1); the evaluation uses 0.6 tangential friction, but the objective's mu is not explicitly stated.
  • dual reset threshold = 2e-2
    Threshold on ||lambda_i|| in Algorithm 1 line 15-16, chosen by hand to prevent divergence.
  • k_f (virtual spring stiffness) = not reported
    Spring stiffness in Eq. 12 for the transposed Jacobian control; not reported.
  • alpha (learning rate for po update) = not reported
    Learning rate for the one-step gradient descent in Algorithm 1 line 6; not reported.
  • K (max ADMM iterations) = not reported
    Maximum iteration count in Algorithm 1; not reported.
  • friction-cone linearization = not reported
    Number of faces in the polyhedral pyramid approximating the friction cone, used to convert the QP to linear constraints; not reported.
axioms (5)
  • domain assumption ADMM alternating updates converge to a useful solution for this nonconvex, non-smooth problem.
    The paper states there is no strict convergence guarantee and relies on dual-variable annealing (Sec. III-C, Eq. 13 and Algorithm 1) to keep the algorithm stable; the whole pipeline depends on this heuristic convergence.
  • domain assumption MuJoCo forward simulation with transposed Jacobian spring forces solves the hand subproblem and preserves non-penetration.
    Sec. III-C Step 2 uses MuJoCo to step the hand under virtual spring forces and assumes this yields an interpenetration-free hand satisfying q in C; the evaluation then measures PD=0.0.
  • domain assumption The force-closure QP with polyhedral friction cone and g=0 is a valid proxy for dynamic stability under the six-wrench MuJoCo evaluation.
    The optimization objective e in Eq. 6 is used to drive the grasps, while success is measured by whether the grasp resists six external forces in MuJoCo; no systematic calibration of this proxy is provided.
  • domain assumption Projecting po onto the object surface with barycentric interpolation preserves differentiability and contact topology.
    Sec. III-C Step 1 computes the gradient of the surface normal n_o with respect to p_o via barycentric interpolation on local mesh triangles, assuming this parameterization is smooth enough for one-step gradient descent.
  • domain assumption Dexonomy templates and initialization constitute a diverse, valid prior.
    The method inherits template annotation and global alignment from Dexonomy [13]; the ablation in Table IV shows a large drop without initialization, so the final result depends on this external prior.

pith-pipeline@v1.3.0-alltime-deepseek · 11141 in / 12201 out tokens · 123717 ms · 2026-08-02T18:10:55.071137+00:00 · methodology

0 comments
read the original abstract

Synthesizing high-quality dexterous grasps is a fundamental challenge in robot manipulation, requiring adherence to diversity, kinematic feasibility (valid hand-object contact without penetration), and dynamic stability (secure multi-contact forces). The recent framework Dexonomy successfully ensures broad grasp diversity through dense sampling and improves kinematic feasibility via a simulator-based refinement method that excels at resolving exact collisions. However, its reliance on fixed contact points restricts the hand's reachability and prevents the optimization of grasp metrics for dynamic stability. Conversely, purely gradient-based optimizers can maximize dynamic stability but rely on simplified contact approximations that inevitably cause physical penetrations. To bridge this gap, we propose GraspADMM, a novel grasp synthesis framework that preserves sampling-based diversity while improving kinematic feasibility and dynamic stability. By formulating the refinement stage using the Alternating Direction Method of Multipliers (ADMM), we decouple the target contact points on the object from the actual contact locations on the hand. This decomposition allows the pipeline to alternate between updating the target object points to directly maximize dynamic grasp metrics, and adjusting the hand pose to physically reach these targets while strictly respecting collision boundaries. Extensive experiments demonstrate that GraspADMM significantly outperforms state-of-the-art baselines, achieving a nearly 15\% absolute improvement in grasp success rate for type-unaware synthesis and roughly a 100\% relative improvement in type-aware synthesis. Furthermore, our approach maintains robust, physically plausible grasp generation even under extreme low-friction conditions.

Figures

Figures reproduced from arXiv: 2603.13832 by Baoquan Chen, He Wang, Jiayi Chen, Liangwang Ruan.

Figure 1
Figure 1. Figure 1: Overview. From the same initialization, our Gras￾pADMM framework generates robust grasps by optimizing contact points on the hand (red) and the object (yellow), while Dexonomy [13] simply fixes these points. a promising direction. Frameworks like GraspIt! [4] used sampling to handle these non-differentiable metrics, but were highly inefficient for high-DoF hands. Recent works [5]– [12] improved efficiency … view at source ↗
Figure 2
Figure 2. Figure 2: ADMM Optimization Pipeline. (a) Update target object contact points p o via gradient descent to maximize dynamic stability. (b) Update hand pose q and points p h via forward-simulated transposed Jacobian control to satisfy kinematic feasibility. (c) Update the dual variable λ. Finally, the dynamic grasp quality metric e is defined as the residual wrench magnitude: e = ∥ Xm i=1 J T o,ifi − g∥ 2 , (6) where … view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of Optimized Grasps. Comparison using the Allegro hand (blue) and Shadow hand (black). Given identical initializations, our method synthesizes more physically stable grasps than Dexonomy. succeeds only if it resists six external forces in MuJoCo. The success criteria for the object pose are 5cm and 15◦ . • OSR (%): Object Success Rate, the percentage of objects for which at least one successf… view at source ↗
Figure 4
Figure 4. Figure 4: Performance on the Hard Benchmark. Grasp success rate and object success rate across varying tangential friction coefficients. Our method consistently outperforms Dexonomy [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effect of Dual Variable Annealing. Green: target ob￾ject contacts p o . Red: dual-shifted targets p o −λ. Annealing bounds the accumulation of λ, preventing the shifted targets from diverging and ensuring stable optimization. V. CONCLUSION & LIMITATION In this work, we presented GraspADMM, a novel dex￾terous grasp synthesis framework that improves the kine￾matic feasibility and dynamic stability. By formul… view at source ↗
Figure 5
Figure 5. Figure 5: Robustness under Extreme Low Friction (µ = 0.1). While Dexonomy struggles to find more than a single stable grasp out of 100 attempts for the shown objects, Gras￾pADMM reliably generates a diverse set of stable grasps. pipeline, whereas the other baselines rely primarily on GPU optimization. On a single NVIDIA RTX 4090 GPU, DexGraspNet [8], FRoGGeR [10], and SpringGrasp [11] average fewer than 3 samples pe… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 2 canonical work pages

  1. [1]

    Graspnet-1billion: A large- scale benchmark for general object grasping,

    H.-S. Fang, C. Wang, M. Gou, and C. Lu, “Graspnet-1billion: A large- scale benchmark for general object grasping,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11 444–11 453

    2020

  2. [2]

    Anygrasp: Robust and efficient grasp perception in spatial and temporal domains,

    H.-S. Fang, C. Wang, H. Fang, M. Gou, J. Liu, H. Yan, W. Liu, Y . Xie, and C. Lu, “Anygrasp: Robust and efficient grasp perception in spatial and temporal domains,”IEEE Transactions on Robotics, vol. 39, no. 5, pp. 3929–3945, 2023

    2023

  3. [3]

    Planning optimal grasps

    C. Ferrari, J. F. Cannyet al., “Planning optimal grasps.” inICRA, vol. 3, no. 4, 1992, p. 6

    1992

  4. [4]

    Graspit! a versatile simulator for robotic grasping,

    A. T. Miller and P. K. Allen, “Graspit! a versatile simulator for robotic grasping,”IEEE Robotics & Automation Magazine, vol. 11, no. 4, pp. 110–122, 2004

    2004

  5. [5]

    Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator,

    T. Liu, Z. Liu, Z. Jiao, Y . Zhu, and S.-C. Zhu, “Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator,”IEEE Robotics and Automation Letters, vol. 7, no. 1, pp. 470–477, 2021

    2021

  6. [6]

    Grasp’d: Differentiable contact-rich grasp synthesis for multi-fingered hands,

    D. Turpin, L. Wang, E. Heiden, Y .-C. Chen, M. Macklin, S. Tsogkas, S. Dickinson, and A. Garg, “Grasp’d: Differentiable contact-rich grasp synthesis for multi-fingered hands,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 201–221

    2022

  7. [7]

    Fast-grasp’d: Dexterous multi-finger grasp generation through differentiable simulation,

    D. Turpin, T. Zhong, S. Zhang, G. Zhu, E. Heiden, M. Macklin, S. Tsogkas, S. Dickinson, and A. Garg, “Fast-grasp’d: Dexterous multi-finger grasp generation through differentiable simulation,” in 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 8082–8089

    2023

  8. [8]

    Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation,

    R. Wang, J. Zhang, J. Chen, Y . Xu, P. Li, T. Liu, and H. Wang, “Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 11 359–11 366

    2023

  9. [9]

    Task-oriented dexterous grasp synthesis via differentiable grasp wrench boundary estimator,

    J. Chen, Y . Chen, J. Zhang, and H. Wang, “Task-oriented dexterous grasp synthesis via differentiable grasp wrench boundary estimator,” arXiv preprint arXiv:2309.13586, 2023

    Pith/arXiv arXiv 2023

  10. [10]

    Frogger: Fast robust grasp generation via the min-weight metric,

    A. H. Li, P. Culbertson, J. W. Burdick, and A. D. Ames, “Frogger: Fast robust grasp generation via the min-weight metric,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 6809–6816

    2023

  11. [11]

    Springgrasp: An optimization pipeline for robust and compliant dexterous pre-grasp synthesis,

    S. Chen, J. Bohg, and C. K. Liu, “Springgrasp: An optimization pipeline for robust and compliant dexterous pre-grasp synthesis,”arXiv preprint arXiv:2404.13532, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Bodex: Scalable and efficient robotic dexterous grasp synthesis using bilevel optimization,

    J. Chen, Y . Ke, and H. Wang, “Bodex: Scalable and efficient robotic dexterous grasp synthesis using bilevel optimization,”arXiv preprint arXiv:2412.16490, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Dexonomy: Synthesizing all dexterous grasp types in a grasp taxonomy,

    J. Chen, Y . Ke, L. Peng, and H. Wang, “Dexonomy: Synthesizing all dexterous grasp types in a grasp taxonomy,”Robotics: Science and Systems, 2025

    2025

  14. [14]

    Grasp quality measures: review and performance,

    M. A. Roa and R. Su ´arez, “Grasp quality measures: review and performance,”Autonomous robots, vol. 38, pp. 65–88, 2015

    2015

  15. [15]

    Dex-net 1.0: A cloud-based network of 3d objects for robust grasp planning using a multi-armed bandit model with correlated rewards,

    J. Mahler, F. T. Pokorny, B. Hou, M. Roderick, M. Laskey, M. Aubry, K. Kohlhoff, T. Kr ¨oger, J. Kuffner, and K. Goldberg, “Dex-net 1.0: A cloud-based network of 3d objects for robust grasp planning using a multi-armed bandit model with correlated rewards,” in2016 IEEE international conference on robotics and automation (ICRA). IEEE, 2016, pp. 1957–1964

    2016

  16. [16]

    Dex- net 3.0: Computing robust vacuum suction grasp targets in point clouds using a new analytic model and deep learning,

    J. Mahler, M. Matl, X. Liu, A. Li, D. Gealy, and K. Goldberg, “Dex- net 3.0: Computing robust vacuum suction grasp targets in point clouds using a new analytic model and deep learning,” in2018 IEEE International Conference on robotics and automation (ICRA). IEEE, 2018, pp. 5620–5627

    2018

  17. [17]

    Suctionnet-1billion: A large- scale benchmark for suction grasping,

    H. Cao, H.-S. Fang, W. Liu, and C. Lu, “Suctionnet-1billion: A large- scale benchmark for suction grasping,”IEEE Robotics and Automation Letters, vol. 6, no. 4, pp. 8718–8725, 2021

    2021

  18. [18]

    Dexterous grasping via eigengrasps: A low-dimensional approach to a high-complexity prob- lem,

    M. Ciocarlie, C. Goldfeder, and P. Allen, “Dexterous grasping via eigengrasps: A low-dimensional approach to a high-complexity prob- lem,” inRobotics: Science and systems manipulation workshop- sensing and adapting to the real world, 2007

    2007

  19. [19]

    Oakink: A large-scale knowledge repository for understanding hand-object interaction,

    L. Yang, K. Li, X. Zhan, F. Wu, A. Xu, L. Liu, and C. Lu, “Oakink: A large-scale knowledge repository for understanding hand-object interaction,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 20 953–20 962

    2022

  20. [20]

    Learning human-like functional grasping for multi-finger hands from few demonstrations,

    W. Wei, P. Wang, S. Wang, Y . Luo, W. Li, D. Li, Y . Huang, and H. Duan, “Learning human-like functional grasping for multi-finger hands from few demonstrations,”IEEE Transactions on Robotics, 2024

    2024

  21. [21]

    Cross-category functional grasp tansfer,

    R. Wu, T. Zhu, X. Lin, and Y . Sun, “Cross-category functional grasp tansfer,”arXiv preprint arXiv:2405.08310, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Boyd and L

    S. Boyd and L. Vandenberghe,Convex Optimization. Cambridge University Press, 2004

    2004

  23. [23]

    Fast simulation of mass-spring systems,

    T. Liu, A. W. Bargteil, J. F. O’Brien, and L. Kavan, “Fast simulation of mass-spring systems,”ACM Trans. Graph., vol. 32, no. 6, Nov

  24. [24]

    Admm⊇projective dynamics: fast simulation of general constitutive models,

    R. Narain, M. Overby, and G. E. Brown, “Admm⊇projective dynamics: fast simulation of general constitutive models,” inProceed- ings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation, ser. SCA ’16. Goslar, DEU: Eurographics Association, 2016, p. 21–28

    2016

  25. [25]

    Progressive parameterizations,

    L. Liu, C. Ye, R. Ni, and X.-M. Fu, “Progressive parameterizations,” ACM Transactions on Graphics(SIGGRAPH), vol. 37, no. 4, 2018

    2018

  26. [26]

    Simple and scalable frictional contacts for thin nodal objects,

    G. Daviet, “Simple and scalable frictional contacts for thin nodal objects,”ACM Trans. Graph., vol. 39, no. 4, 8 2020. [Online]. Available: https://doi.org/10.1145/3386569.3392439

    arXiv 2020

  27. [27]

    Interactive hair simulation on the gpu using admm,

    ——, “Interactive hair simulation on the gpu using admm,” inACM SIGGRAPH 2023 Conference Proceedings, ser. SIGGRAPH ’23. New York, NY , USA: Association for Computing Machinery, 2023. [Online]. Available: https://doi.org/10.1145/3588432.3591551

    arXiv 2023

  28. [28]

    Sparse iterative closest point,

    S. Bouaziz, A. Tagliasacchi, and M. Pauly, “Sparse iterative closest point,” inProceedings of the Eleventh Eurographics/ACMSIGGRAPH Symposium on Geometry Processing, ser. SGP ’13. Goslar, DEU: Eurographics Association, 2013, p. 113–123. [Online]. Available: https://doi.org/10.1111/cgf.12178

    work page doi:10.1111/cgf.12178 2013

  29. [29]

    Sparse localized deformation components,

    T. Neumann, K. Varanasi, S. Wenger, M. Wacker, M. Magnor, and C. Theobalt, “Sparse localized deformation components,”ACM Trans. Graph., vol. 32, no. 6, Nov. 2013. [Online]. Available: https://doi.org/10.1145/2508363.2508417

    arXiv 2013

  30. [30]

    Fast and accurate matrix completion via truncated nuclear norm regularization,

    Y . Hu, D. Zhang, J. Ye, X. Li, and X. He, “Fast and accurate matrix completion via truncated nuclear norm regularization,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 35, no. 9, pp. 2117–2130, 2013

    2013

  31. [31]

    Accelerating admm for efficient simulation and optimization,

    J. Zhang, Y . Peng, W. Ouyang, and B. Deng, “Accelerating admm for efficient simulation and optimization,”ACM Trans. Graph., vol. 38, no. 6, Nov. 2019. [Online]. Available: https: //doi.org/10.1145/3355089.3356491

    arXiv 2019

  32. [32]

    A stochastic alternating direction method of multipliers for non-smooth and non-convex optimization,

    F. Bian, J. Liang, and X. Zhang, “A stochastic alternating direction method of multipliers for non-smooth and non-convex optimization,” Inverse Problems, vol. 37, no. 7, p. 075009, jul 2021. [Online]. Available: https://doi.org/10.1088/1361-6420/ac0966

    work page doi:10.1088/1361-6420/ac0966 2021

  33. [33]

    Global convergence of admm in nonconvex nonsmooth optimization,

    Y . Wang, W. Yin, and J. Zeng, “Global convergence of admm in nonconvex nonsmooth optimization,” 2018. [Online]. Available: https://arxiv.org/abs/1511.06324

    Pith/arXiv arXiv 2018

  34. [34]

    The grasp taxonomy of human grasp types,

    T. Feix, J. Romero, H.-B. Schmiedmayer, A. M. Dollar, and D. Kragic, “The grasp taxonomy of human grasp types,”IEEE Transactions on human-machine systems, vol. 46, no. 1, pp. 66–77, 2015

    2015

  35. [35]

    Clarabel: An interior-point solver for conic programs with quadratic objectives,

    P. J. Goulart and Y . Chen, “Clarabel: An interior-point solver for conic programs with quadratic objectives,”arXiv preprint arXiv:2405.12762, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Learning diverse and physically feasi- ble dexterous grasps with generative model and bilevel optimization,

    A. Wu, M. Guo, and C. K. Liu, “Learning diverse and physically feasi- ble dexterous grasps with generative model and bilevel optimization,” arXiv preprint arXiv:2207.00195, 2022

    Pith/arXiv arXiv 2022

  37. [37]

    Objaverse: A universe of annotated 3d objects,

    M. Deitke, D. Schwenk, J. Salvador, L. Weihs, O. Michel, E. Vander- Bilt, L. Schmidt, K. Ehsani, A. Kembhavi, and A. Farhadi, “Objaverse: A universe of annotated 3d objects,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 13 142–13 153

    2023

  38. [38]

    NVIDIA Corporation

    (2025) Newton physics engine documentation. NVIDIA Corporation. Joint development with Google DeepMind and Disney Research; contributed to the Linux Foundation. [Online]. Available: https: //developer.nvidia.com/newton-physics

    2025

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.