InterMASH: A Unified Geometric Representation for Grasp Synthesis

Xuanze Yang; Yumeng Liu; Haiyang Xin; Changhao Li; Haowei Shen; Kai Xu; Ligang Liu; Ruizhen Hu

doi:10.1145/3829340.3842336

Teaser

ShadowHand, MANO, Barrett, and Allegro share anchor-indexed object, hand, and contact tokens, enabling grasp synthesis across embodiments. — **One representation, multiple embodiments.** InterMASH encodes local object geometry, hand geometry, and contact in a shared token space, connecting human and robotic grasp synthesis.

Abstract

Grasp synthesis aims to generate stable and physically plausible hand–object interactions. Yet a unified representation across human and robotic hands remains challenging because of differences in hand morphology and surface modeling. Contact maps alone leave hand geometry implicit, while dense interaction descriptors can be computationally expensive and redundant.

We introduce InterMASH, a unified geometric representation that establishes cross-embodiment correspondence using sphere-fixed anchors. At each anchor, low-degree spherical harmonics compactly encode local hand geometry, object geometry, and contact, forming an explicit and interpretable token sequence.

Building on this structure, a conditional Diffusion Transformer jointly generates hand geometry and contact. Physics-guided training and sampling improve physical plausibility, while patch-wise inverse kinematics recovers the final articulated hand pose. Experiments demonstrate strong physical feasibility on DexGraspNet, support for mixed-hand training, and improved robotic grasp success and diversity when fine-tuning with human grasp data.

Method

InterMASH framework: anchor-based hand and object representation, a conditional Diffusion Transformer with neighborhood-enhanced attention, and inverse kinematics for the final pose. — The InterMASH pipeline. Object geometry and a hand template condition the joint generation of grasp-specific hand geometry and contact. Neighborhood-enhanced attention incorporates local patch adjacency.

Align & encode

Align robotic-hand patches to a MANO reference ordering using keypoint-distance signatures. Shared sphere-fixed anchors encode local surface patches with spherical harmonics; coarse-to-fine fitting stabilizes reconstruction.

Jointly denoise

Generate hand geometry and contact tokens conditioned on the object and hand template. Neighborhood-enhanced attention uses patch adjacency, while physical guidance encourages stable contact and reduces penetration.

Recover the pose

Reconstruct patch-wise point sets, initialize global hand alignment, and solve inverse kinematics under the shared patch correspondence to recover the final articulated grasp.

Original meshes above their colored InterMASH patch reconstructions for Barrett, Allegro, ShadowHand, and MANO. — Original meshes (top) and InterMASH reconstructions (bottom), across four hand embodiments.

Each anchor stores a compact description of its local interaction. The continuous patch representation supports surface sampling beyond a fixed point-cloud resolution, while a consistent semantic ordering connects corresponding regions across hands.

Results

Twenty-one ShadowHand grasps generated by InterMASH on differently shaped objects from DexGraspNet. — Generated grasps on DexGraspNet across varied object geometries (paper Figure 8).

Barrett and ShadowHand grasp comparisons, with D(R, O) above InterMASH for each hand type. — Cross-hand comparisons on the filtered CMapDataset: D(R, O) in yellow and InterMASH in blue. Barrett results occupy the top two rows; ShadowHand results the bottom two (paper Figure 9).

Three groups of grasp comparisons: DexGrasp Anything in green, D(R, O) in orange, and InterMASH in blue. — DexGrasp Anything (green), D(R, O) (orange), and InterMASH (blue). Each group of three rows compares the same objects (paper Figure 7).

InterMASH supports single-hand generation, cross-embodiment robotic grasp synthesis, and human-to-robot prior transfer. The following results are reported in the paper, with each benchmark retaining its own evaluation protocol.

Single-hand grasp generation

DexGraspNet · ShadowHand · Paper Table 1

Comparison on DexGraspNet. Higher is better for success and diversity; lower is better for penetration.
Method	Suc.6 (%) ↑	Suc.1 (%) ↑	Pen. (mm) ↓	Div. ↑
UniDexGrasp (2023)	33.9	70.1	31.9	0.14
GraspTTA (2021)	18.6	67.8	24.5	0.13
SceneDiffuser (2023)	26.6	66.9	31.0	0.15
UGG (2024)	46.9	79.0	25.2	0.14
DexGrasp Anything (2025)	53.6	90.4	21.5	0.22
D(R, O) (2024)	46.9	89.7	17.5	0.20
InterMASH Ours	53.5	91.9	16.2	0.14

Suc.6 / Suc.1: success under all six / at least one disturbance direction. Pen.: maximum penetration depth. Div.: mean standard deviation of local pose parameters. Bold marks the best value in each column.

InterMASH achieves the best Suc.1 and lowest penetration among the compared methods. Its stricter six-direction success and diversity remain below the strongest baseline, reflecting a quality–diversity trade-off.

Cross-embodiment robotic grasp synthesis

Filtered CMapDataset · Barrett & ShadowHand · Paper Table 2

Cross-embodiment success rate and diversity on the filtered CMapDataset.
Method	Success rate (%) ↑		Diversity ↑
Method	Barrett	ShadowHand	Barrett	ShadowHand
DFC	86.30	58.80	0.532	0.435
GenDexGrasp	67.00	54.20	0.488	0.318
D(R, O)	87.30	83.00	0.513	0.441
InterMASH Shadow only	—	57.62	—	0.416
InterMASH Shadow + Barrett	90.30	64.15	0.480	0.396

Evaluated in Isaac Gym under the D(R, O) protocol. Diversity is the standard deviation of joint values among successful grasps. These success rates use a different protocol from the DexGraspNet table above.

Mixed-hand training improves ShadowHand success from 57.62% to 64.15% and yields the highest Barrett success among these methods. D(R, O) remains stronger on ShadowHand in this smaller mixed-hand dataset.

Human grasp priors benefit robotic hands

DexGRAB evaluation · Human-to-robot transfer · Paper Table 3

Starting from a model trained on ShadowHand DexGRAB samples, we fine-tune on a 1:1 mixture of DexGRAB and MANO-based human GRAB samples. Both settings are evaluated on the same DexGRAB test split.

Effect of adding human GRAB data to robotic DexGRAB training.
Training setting	Suc.6 (%) ↑	Suc.1 (%) ↑	Pen. (mm) ↓	Div. ↑
DexGRAB only	25.8	64.2	14.9	0.450
DexGRAB + GRAB fine-tuning	29.0	65.9	18.6	0.500

Human data improves success and diversity, while penetration increases from 14.9 to 18.6 mm. The paper links this trade-off to higher average penetration in the human source data.

BibTeX

@inproceedings{yang2026intermash,
  title     = {{InterMASH}: A Unified Geometric Representation for Grasp Synthesis},
  author    = {Yang, Xuanze and Liu, Yumeng and Xin, Haiyang and Li, Changhao
               and Shen, Haowei and Xu, Kai and Liu, Ligang and Hu, Ruizhen},
  booktitle = {SIGGRAPH Asia 2026 Conference Papers},
  series    = {SA Conference Papers '26},
  year      = {2026},
  publisher = {Association for Computing Machinery},
  address   = {New York, NY, USA},
  location  = {Kuala Lumpur, Malaysia},
  numpages  = {11},
  doi       = {10.1145/3829340.3842336}
}