Pith. sign in

REVIEW 2 major objections 5 minor 30 references

MANGO-Grasp proposes a single cross-embodiment grasp pipeline using Mahalanobis fields over surface-aligned Gaussian plates, reporting 97.59% seen-hand and 84.17% zero-shot unseen-hand simulation success plus 86% real-world success.

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 ↗

MANGO-Grasp uses geometry-oriented 3D Gaussians and Mahalanobis fields to achieve strong cross-embodiment dexterous grasping, with zero-shot transfer to an unseen hand at 84% simulation and 86% real-world success.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Strong representation and zero-shot results, but the seen-hand gains may owe an untested debt to a training-derived contact prior. the 2 major comments →

arxiv 2608.02014 v1 pith:ODWBLV7B submitted 2026-08-03 cs.RO cs.AI

MANGO-Grasp: Mahalanobis Fields over Geometry-Oriented 3D Gaussians for Cross-Embodiment Dexterous Grasping

Heng Zhang , Kevin Yuchen Ma , Mike Zheng Shou , Weisi Lin , Yan Wu This is my paper
classification cs.RO cs.AI
keywords cross-embodiment dexterous graspingMahalanobis fields3D Gaussian primitivesgrasp synthesiszero-shot transfermorpho-kinematic descriptorsrobot hand keypointscontact modeling
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 reading

The paper is trying to establish that cross-embodiment dexterous grasping can be unified by learning anisotropic interaction fields rather than per-hand models. It represents objects as surface-aligned plate-like 3D Gaussian primitives and hands as keypoints with morphology- and kinematics-aware descriptors, then defines the keypoint–primitive interaction as a Mahalanobis distance that is sensitive to normal displacement but tolerant of tangential sliding. The same learned fields are used as training targets and as optimization guidance, so one shared formulation produces grasps for all hands. If true, this means a new robotic hand can be added without retraining or hand-specific tuning; the paper's simulation and real-world numbers support that claim.

Core claim

The central discovery is that an anisotropic, surface-aware interaction representation—Mahalanobis fields over hand keypoints and geometry-oriented 3D Gaussian primitives—can serve simultaneously as a supervised training target and as a differentiable objective for grasp realization, yielding a single cross-embodiment grasp synthesis pipeline. The object is converted into a fixed budget of 256 surface-aligned Gaussian plates with outward normals, adaptively densified by geometric complexity; the hand is represented by 256 surface keypoints with descriptors pretrained to preserve keypoint identity while predicting configuration-dependent motion. The Mahalanobis distance evaluated in each prim

What carries the argument

The key machinery is the Mahalanobis field M_ij = sqrt((R_j^T(p_i - μ_j))^T diag(σ_j^{-2})(R_j^T(p_i - μ_j))) between hand keypoint p_i and object primitive g_j = (μ_j, R_j, σ_j, n_j). Because the Gaussian primitives are refined into thin plates aligned with the local surface, with the short axis along the outward normal, this distance makes the interaction sharply sensitive to motion along the normal and only mildly sensitive to tangential motion. It functions both as the training target for the interaction-field generator and as the guidance term in grasp realization, so the same representation couples prediction and optimization. Supporting components are the geometry-driven densification

Load-bearing premise

The contact-aware keypoint weights computed from training grasps of the three seen hands are used in the training loss and seen-hand realization, then simply set to 1 for the unseen hand; the central zero-shot and seen-hand claims assume this does not underestimate or overestimate transfer.

What would settle it

Retrain the model with all keypoint weights set to 1 (no contact-aware weighting) and measure seen-hand and zero-shot success on the same benchmarks; if the unweighted model matches or surpasses the weighted one, the paper's claimed role for contact-aware weights is not supported. Alternatively, compute the predicted vs ground-truth Mahalanobis fields on a holdout hand; if the fields are accurate but grasp success is low, the realization optimizer rather than the interaction representation is the bottleneck.

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

If this is right

  • A new multi-fingered hand can be added to the system without retraining or per-hand hyperparameter tuning; only forward kinematics, joint limits, and keypoints are needed.
  • Object geometry is represented by a fixed-budget set of surface-aligned anisotropic primitives, so grasp quality no longer depends on uniform point sampling density.
  • The directional contact model should generalize across objects of varying scale and geometry, since the performance degradation between the two benchmarks is smaller than for prior approaches.
  • The same interaction-field formulation can be used for both training and inference-time optimization, eliminating the need for a separate grasp-realization network.
  • Zero-shot transfer to an unseen hand morphology is possible in simulation and on physical hardware, reaching 86% real-world success without fine-tuning.

Where Pith is reading between the lines

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

  • An obvious next test is to drop the contact-aware keypoint weights entirely during training and realization; comparing that variant against the paper's results would isolate how much of the gain comes from the weights themselves rather than the Mahalanobis-field representation.
  • The Mahalanobis field over Gaussian primitives might extend to contact-force prediction or in-hand manipulation, where the same normal/tangential anisotropy governs sliding and rolling constraints.
  • Because the paper's real-world evaluation uses open-loop execution and mesh-based object models, closed-loop execution and mesh-free primitive construction are natural extensions that could improve the reported 86% and remove the dependence on mesh availability.
  • The empirical choice of 256 primitives and 256 keypoints may interact with object size and hand complexity; a scale-adaptive budget could be tested by abating the budget across small and large objects.
Share X Bluesky LinkedIn Reddit HN

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

2 major / 5 minor

Summary. The paper proposes MANGO-Grasp, a cross-embodiment dexterous grasping method. Objects are represented as a fixed budget of geometry-oriented 3D Gaussian primitives with outward normals and anisotropic scales; robot hands are represented by surface keypoints encoded by a pretrained morpho-kinematic encoder. Interaction is modeled as a pairwise Mahalanobis distance between keypoints and primitives, predicted by a cross-attention network with a CVAE. Grasps are realized by optimizing the hand configuration to match the predicted fields, with penetration and self-collision penalties using one weight setting. Experiments on CMAP and MultiGripperGrasp report seen-hand success rates up to 97.59% and 89.47%, zero-shot transfer to an unseen SharpaWave hand at 84.17% and 81.47%, and 86% real-world success. Ablations compare kinematic-awareness pretraining, Mahalanobis vs Euclidean fields, Gaussians vs point clouds, and Gaussians vs mesh-fitted primitives.

Significance. The paper offers a coherent and well-motivated representation: anisotropic, surface-aligned primitives with Mahalanobis interaction fields are a natural fit for contact modeling, and the morpho-kinematic pretraining is a sensible extension of DRO. Strengths include the external simulator and real-robot evaluation, the three-run standard deviations, and the explicit component ablations. If the reported gains are robust, the method would be a meaningful advance. However, the unablated contact prior and missing construction hyperparameters currently prevent the reader from attributing the seen-hand gains to the proposed representation.

major comments (2)
  1. [Sec. III-D/III-E/III-F, Eq. (13), Eq. (17), Eq. (20), Table II] The contact-aware keypoint weights w_e^i (Eq. 13), computed from near-contact statistics over the training grasps of the three seen hands, are used in three places: the kinematic-awareness loss (Eq. 14), the interaction-field regression loss (Eq. 17), and the realization guidance (Eq. 20). No ablation sets w_e^i to uniform. Since DRO and TRO do not use such a hand-specific contact prior, the reported seen-hand margins (up to 8.24 pp, Table I) may be due to this privileged training-data statistic rather than to the proposed representation. Please retrain with w_e^i ≡ 1 for all seen hands and report the resulting Table I/II numbers. In addition, the statement of "one shared optimization/hyperparameter setting" is not literally true while Eq. (20) contains an embodiment-dependent weight that is reset to 1 only for the unseen hand (Sec. IV-C).
  2. [Sec. III-C3, Eq. (7), Eq. (10)] The surface-alignment refinement depends on thresholds τ_l, τ_h, τ_f and weights λ_s, λ_n, λ_p, which are never reported. These parameters control the plate-shaped, surface-aligned Gaussian primitives that define the anisotropic Mahalanobis interaction field; without them the method cannot be reproduced. Please provide a full hyperparameter table for this stage (and, ideally, release the code or a configuration file).
minor comments (5)
  1. [Abstract] The URL appears as "athttps://connor-zh.github.io/MANGO-Grasp/" with a missing space after "at".
  2. [Sec. III-C3, Eq. (10)] The symbol τℓ is inconsistent with the text's τ_l; use τ_l throughout. Also specify whether the hinge penalties are one-sided or two-sided for each term.
  3. [Sec. III-C2] The "maximum projected area" A_j is not defined precisely. Please specify over which views or rotations it is computed, as this directly affects the fixed-budget selection score.
  4. [Sec. III-D, Eq. (13)] The derivation of c_e^i from "near-contact frequency statistics" is underspecified. Please describe how the statistics are accumulated over hand-mesh vertices and propagated to keypoints, since these weights are load-bearing for the seen-hand results.
  5. [Sec. III-F] The phrase "one weight setting" should be qualified, because Eq. (20) contains the embodiment-dependent w_e^i. State explicitly which w_e^i values are used for each hand in the experiments and how the claimed shared setting is defined.

Circularity Check

0 steps flagged

No circular derivation: benchmark success is measured externally, and the training-derived contact prior is a weighting, not a definition of success.

full rationale

MANGO-Grasp's derivation chain is self-contained. Object primitives are fit from meshes by 3DGS-style optimization (Sec. III-C); the ground-truth interaction targets M* are computed by Eq. (16) from grasp-state keypoints and those primitives; the network is trained to regress these targets and at inference optimizes Eqs. (19)-(22) to match its predictions under penetration, self-collision, and joint-limit constraints. Success is evaluated by Isaac Gym physics perturbation and real-robot trials, not by the model's own objective, so the central claim is not defined in terms of its own outputs. The contact-aware keypoint weights w_i^e (Eq. (13)) are derived from training-grasp contact statistics and reused in Eqs. (17) and (20); this is a legitimate fairness concern for seen-hand comparisons, but it is not circular because the weights only modulate the optimization and do not by construction determine the externally measured success. The zero-shot SharpaWave results set w=1 (Sec. IV-C), so the central zero-shot claim is unaffected. The paper's self-citation [10] appears only in a related-work list and is not load-bearing. Stated limitations (open-loop execution, mesh-based construction, pose-estimation errors) are acknowledged in Secs. IV-E and V and do not indicate circularity; however, the missing ablation of w=1 on seen hands is a noteworthy missing support for the interpretation that the representation, rather than the privileged contact prior, drives the seen-hand margins.

Axiom & Free-Parameter Ledger

13 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical entities. The Gaussian primitives and keypoint descriptors are learned representations. The central claim rests on many hand-selected hyperparameters and on assumptions about forward kinematics, the adequacy of Mahalanobis fields for contact, the fixed-budget selection heuristic, the transferability of a training-grasp-derived contact prior, and the validity of the simulation protocol.

free parameters (13)
  • G (number of object primitives) = 256
    Number of object Gaussian primitives; fixed to balance coverage and interaction matrix size (Sec. III-A).
  • N (number of hand keypoints) = 256
    Number of hand keypoints per embodiment; fixed empirically (Sec. III-A).
  • beta (curvature boost cap) = 0.2
    Caps the curvature boost in the fixed-budget retention score (Sec. III-C2).
  • lambda_C, lambda_S = 0.8, 0.2
    Weights in the photometric loss of geometry-driven densification, Eq. (5).
  • lambda_opa = 0.02
    Opacity sparsity weight in Eq. (6).
  • Surface alignment hyperparameters (tau_l, tau_h, tau_f, lambda_s, lambda_n, lambda_p) = not reported
    Weights and thresholds in Eqs. (7)-(10); values absent from the text, blocking exact reproduction.
  • tau (contrastive temperature) = 0.1
    Temperature in the morphology-identity contrastive loss, Eq. (11).
  • alpha (contact prior floor) = 0.1
    Floors near-zero contact prior weights in Eq. (13).
  • lambda_reg, lambda_kl = 10, 0.01
    Weights in the interaction fields training loss, Eq. (18).
  • lambda_guide, lambda_pen, lambda_self = 50, 60, 10
    Weights in the grasp realization objective, Eq. (19).
  • m, K (penetration) = -0.001, 20
    Penetration margin and top-K selection in Eq. (21).
  • epsilon_s, K (self-collision) = 0.01, 20
    Self-collision threshold and top-K selection in Eq. (22).
  • Contact prior weights w_e^i = per-keypoint values
    Computed from training-grasp contact statistics (Eq. 13); used in both training and realization, potentially biasing seen-hand results.
axioms (5)
  • domain assumption Analytical forward kinematics (Eq. (2)) maps canonical keypoints to configuration-dependent positions accurately for all embodiments, including the unseen SharpaWave hand.
    This underpins both descriptor pretraining (Sec. III-D) and grasp realization (Sec. III-F); if the URDF/mesh-based keypoint mapping does not reflect the real hand, the optimized configuration would not be realizable.
  • domain assumption The Mahalanobis field (Eq. (16)) over surface-aligned Gaussian plates is an adequate proxy for contact compatibility; minimizing field mismatch under the chosen energies yields physically stable grasps.
    This is the core inductive bias of the method, used in training targets (Eq. (17)) and optimization guidance (Eq. (20)).
  • ad hoc to paper The fixed-budget selection score o_j A_j (1 + beta H_j) (Sec. III-C2) preserves the contact-relevant geometry needed for grasp success.
    The score is a hand-designed heuristic; there is no proof that the top-256 primitives retain the surface regions that determine grasp stability.
  • ad hoc to paper The contact prior w_e^i (Eq. (13)) computed from training grasps of seen hands is appropriate for weighting the interaction losses, and setting w_i=1 for the unseen hand is a valid zero-shot protocol.
    Stated in Sec. IV-C; if the prior leaks training statistics, seen-hand results could be inflated.
  • domain assumption Isaac Gym simulation with the 0.5 m_o g perturbation protocol is a valid proxy for real-world grasp success.
    Used for all benchmark evaluations; real-world validation on 10 objects partially supports it.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of MANGO-Grasp: Mahalanobis Fields over Geometry-Oriented 3D Gaussians for Cross-Embodiment Dexterous Grasping." pith.science (2026). https://pith.science/paper/ODWBLV7B

@misc{pith2026260802014,
  author       = {Pith},
  title        = {Pith review of: MANGO-Grasp: Mahalanobis Fields over Geometry-Oriented 3D Gaussians for Cross-Embodiment Dexterous Grasping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ODWBLV7B}},
  note         = {Machine review of arXiv:2608.02014}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Cross-embodiment dexterous grasping aims to synthesize stable grasps across heterogeneous multi-fingered hands with little or no embodiment-specific tuning. Existing interaction-centric methods achieve promising results, but their object representations often underrepresent local surface geometry, while their robot descriptors do not explicitly encode both robot morphology and kinematics. We propose MANGO-Grasp, an anisotropic interaction framework that represents objects as geometry-oriented 3D Gaussian primitives and robot hands as surface keypoints encoded into morpho-kinematic descriptors. The object primitives are adaptively allocated by geometric complexity and shaped as surface-aligned plates with outward normals, encoding local geometry. Mahalanobis fields over keypoint--primitive pairs serve as interaction prediction targets during training and as optimization guidance for grasp realization at inference. These fields rise sharply for displacement along the surface normal but only gently within the tangent plane, matching the directional structure of contact. Grasps are realized with one shared optimization formulation and hyperparameter setting across all embodiments. On the CMAP and MultiGripperGrasp benchmarks, MANGO-Grasp outperforms the strongest seen-hand baseline by up to 8.24 percentage points in simulation. It also transfers zero-shot to the unseen SharpaWave hand, improving over the strongest zero-shot baseline by up to 16.57 percentage points, and achieves 86% success in real-world experiments. The code and additional materials will be made available upon publication at https://connor-zh.github.io/MANGO-Grasp/.

Figures

Figures reproduced from arXiv: 2608.02014 by Heng Zhang, Kevin Yuchen Ma, Mike Zheng Shou, Weisi Lin, Yan Wu.

Figure 1
Figure 1. Figure 1: Overview of the proposed anisotropic interaction formulation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Method overview. The object mesh is converted into a fixed-budget set of surface-aligned, plate-like 3D Gaussian primitives [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Pretraining of morpho-kinematic robot descriptors. For two [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] 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

30 extracted references · 6 linked inside Pith

  1. [1]

    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 Robot. Autom. Lett., vol. 7, no. 1, pp. 470–477, 2021

    2021

  2. [2]

    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,”arXiv preprint arXiv:2210.02697, 2022

    Pith/arXiv arXiv 2022

  3. [3]

    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 Proc. IEEE Int. Conf. Robot. Autom. (ICRA), 2023, pp. 8082–8089

    2023

  4. [4]

    Dex- terous grasp transformer,

    G.-H. Xu, Y .-L. Wei, D. Zheng, X.-M. Wu, and W.-S. Zheng, “Dex- terous grasp transformer,” inCVPR, 2024

    2024

  5. [5]

    Dexdiffuser: Generating dexterous grasps with diffusion models,

    Z. Weng, H. Lu, D. Kragic, and J. Lundell, “Dexdiffuser: Generating dexterous grasps with diffusion models,” 2024. [Online]. Available: https://arxiv.org/abs/2402.02989

    Pith/arXiv arXiv 2024

  6. [6]

    Dexgrasp anything: Towards universal robotic dexterous grasping with physics awareness,

    Y . Zhong, Q. Jiang, J. Yu, and Y . Ma, “Dexgrasp anything: Towards universal robotic dexterous grasping with physics awareness,” 2025. [Online]. Available: https://arxiv.org/abs/2503.08257

    Pith/arXiv arXiv 2025

  7. [7]

    Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy,

    Y . Xu, W. Wan, J. Zhang, H. Liu, Z. Shan, H. Shen, R. Wang, H. Geng, Y . Weng, J. Chen,et al., “Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy,” inCVPR, 2023

    2023

  8. [8]

    Unidexgrasp++: Improving dexterous grasping policy learning via geometry-aware curriculum and iterative generalist-specialist learn- ing,

    W. Wan, H. Geng, Y . Liu, Z. Shan, Y . Yang, L. Yi, and H. Wang, “Unidexgrasp++: Improving dexterous grasping policy learning via geometry-aware curriculum and iterative generalist-specialist learn- ing,”arXiv preprint arXiv:2304.00464, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Unigrasptransformer: Simplified policy distillation for scalable dexterous robotic grasping,

    W. Wang, F. Wei, L. Zhou, X. Chen, L. Luo, X. Yi, Y . Zhang, Y . Liang, C. Xu, Y . Lu,et al., “Unigrasptransformer: Simplified policy distillation for scalable dexterous robotic grasping,” inCVPR, 2025

    2025

  10. [10]

    Machagrasp: Morphology-aware cross-embodiment dexterous hand articulation generation for grasping,

    H. Zhang, K. Y . Ma, M. Z. Shou, W. Lin, and Y . Wu, “Machagrasp: Morphology-aware cross-embodiment dexterous hand articulation generation for grasping,” 2026. [Online]. Available: https://arxiv.org/abs/2510.06068

    arXiv 2026

  11. [11]

    Unigrasp: Learning a unified model to grasp with multifingered robotic hands,

    L. Shao, F. Ferreira, M. Jorda, V . Nambiar, J. Luo, E. Solowjow, J. A. Ojea, O. Khatib, and J. Bohg, “Unigrasp: Learning a unified model to grasp with multifingered robotic hands,”IEEE Robot. Autom. Lett., vol. 5, no. 2, pp. 2286–2293, 2020

    2020

  12. [12]

    Gendexgrasp: Generalizable dexterous grasping,

    P. Li, T. Liu, Y . Li, Y . Geng, Y . Zhu, Y . Yang, and S. Huang, “Gendexgrasp: Generalizable dexterous grasping,” inProc. IEEE Int. Conf. Robot. Autom. (ICRA). IEEE, 2023, pp. 8068–8074

    2023

  13. [13]

    Geometry matching for multi-embodiment grasping,

    M. Attarian, M. A. Asif, J. Liu, R. Hari, A. Garg, I. Gilitschenski, and J. Tompson, “Geometry matching for multi-embodiment grasping,” in CoRL, 2023

    2023

  14. [14]

    Geomatch++: Morphol- ogy conditioned geometry matching for multi-embodiment grasping,

    Y . Wei, M. Attarian, and I. Gilitschenski, “Geomatch++: Morphol- ogy conditioned geometry matching for multi-embodiment grasping,” arXiv preprint arXiv:2412.18998, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    D(R,O)grasp: A unified representation of robot and object interaction for cross-embodiment dexterous grasping,

    Z. Wei, Z. Xu, J. Guo, Y . Hou, C. Gao, Z. Cai, J. Luo, and L. Shao, “D(R,O)grasp: A unified representation of robot and object interaction for cross-embodiment dexterous grasping,” inProc. IEEE Int. Conf. Robot. Autom. (ICRA), 2025, pp. 4982–4988

    2025

  16. [16]

    T(r,o) grasp: Efficient graph diffusion of robot-object spatial transformation for cross-embodiment dexterous grasping,

    X. Fei, Z. Xu, H. Fang, T. Zhang, and L. Shao, “T(r,o) grasp: Efficient graph diffusion of robot-object spatial transformation for cross-embodiment dexterous grasping,” 2025. [Online]. Available: https://arxiv.org/abs/2510.12724

    arXiv 2025

  17. [17]

    3d gaussian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, G. Drettakis,et al., “3d gaussian splatting for real-time radiance field rendering.”ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

    2023

  18. [18]

    Multi- grippergrasp: A dataset for robotic grasping from parallel jaw grippers to dexterous hands,

    L. F. Casas, N. Khargonkar, B. Prabhakaran, and Y . Xiang, “Multi- grippergrasp: A dataset for robotic grasping from parallel jaw grippers to dexterous hands,” inProc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), 2024

    2024

  19. [19]

    Cedex: Cross-embodiment dexterous grasp generation at scale from human-like contact representations,

    Z. Wu, R. A. Potamias, X. Zhang, Z. Zhang, J. Deng, and S. Luo, “Cedex: Cross-embodiment dexterous grasp generation at scale from human-like contact representations,” 2025. [Online]. Available: https://arxiv.org/abs/2509.24661

    arXiv 2025

  20. [20]

    Unimorphgrasp: Diffusion model with morphology-awareness for cross-embodiment dexterous grasp generation,

    Z. Wu, X. Zhang, Z. Chen, J. Deng, R. A. Potamias, and S. Luo, “Unimorphgrasp: Diffusion model with morphology-awareness for cross-embodiment dexterous grasp generation,” 2026. [Online]. Available: https://arxiv.org/abs/2602.00915

    arXiv 2026

  21. [21]

    Cross-embodiment dexterous grasping with reinforcement learning,

    H. Yuan, B. Zhou, Y . Fu, and Z. Lu, “Cross-embodiment dexterous grasping with reinforcement learning,” inICLR, 2025

    2025

  22. [22]

    Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering,

    A. Gu ´edon and V . Lepetit, “Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering,”

  23. [23]

    2d gaussian splatting for geometrically accurate radiance fields,

    B. Huang, Z. Yu, A. Chen, A. Geiger, and S. Gao, “2d gaussian splatting for geometrically accurate radiance fields,” inProc. ACM SIGGRAPH Conf. Papers, 2024, pp. 1–11

    2024

  24. [24]

    Nerfstudio: A modular framework for neural radiance field development,

    M. Tanciket al., “Nerfstudio: A modular framework for neural radiance field development,” inProc. ACM SIGGRAPH Conf., 2023

    2023

  25. [25]

    Dynamic graph cnn for learning on point clouds,

    Y . Wang, Y . Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, and J. M. Solomon, “Dynamic graph cnn for learning on point clouds,”ACM Trans. Graph., vol. 38, no. 5, pp. 1–12, 2019

    2019

  26. [26]

    On the generalised distance in statistics

    P. MAHALANOBIS, “On the generalised distance in statistics.” in Proc. Nat. Inst. Sci.(India), vol. 12, 1936, pp. 49–55

    1936

  27. [27]

    Learning structured output represen- tation using deep conditional generative models,

    K. Sohn, H. Lee, and X. Yan, “Learning structured output represen- tation using deep conditional generative models,”Adv. Neural Inf. Process. Syst., vol. 28, 2015

    2015

  28. [28]

    Gpu-accelerated robotic simulation for distributed reinforce- ment learning,

    J. Liang, V . Makoviychuk, A. Handa, N. Chentanez, M. Macklin, and D. Fox, “Gpu-accelerated robotic simulation for distributed reinforce- ment learning,” inCoRL, 2018

    2018

  29. [29]

    FoundationPose: Unified 6d pose estimation and tracking of novel objects,

    B. Wen, W. Yang, J. Kautz, and S. Birchfield, “FoundationPose: Unified 6d pose estimation and tracking of novel objects,” inCVPR, 2024

    2024

  30. [2023]

    Available: https://arxiv.org/abs/2311.12775

    [Online]. Available: https://arxiv.org/abs/2311.12775

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.