VolCo: Volumetric Contact
for High-Fidelity Human Grasp Generation

1University of Birmingham, 2Beijing Institute of Technology
NeurIPS 2026

*Indicates Corresponding Author

What is Volumetric Contact?

Contact is the likelihood of an object point in contact with the hand. Voumetric contact samples the points in small volumes around the surface.

Binoculars
Mug

Why Volume Contact?

A one-sentence summary of the core motivation goes here.

01

Local Info Preserving

VolCo Preserves local information to the largest extent: contact positions, finger orientations, and exact penetration depth. It can restore the 3D local hand configurations.

Ground truth
ContactOpt
ContactGen
ManiDext
VolCo (ours)
02

Robust Penetration Avoidance

VolCo reconstructs the local hand surface, which allows a more robust hand fitting based on Euclidean distance, not contact value. Thus it does not heavily rely on the non-robust penetration loss based on nearest-neighbor query.

ContactOpt · surface contact
Surface contact
Surface contact illustration
Contact on the surface
Nearest-neighbor penetration queries
Penetration Loss
Grasp
VolCo (ours) · volumetric contact
Volumetric contact
Volumetric contact illustration
Contact in 3D volume
Local hand surface
Euclidean distance fitting
Euclidean fitting
Grasp

Drag any 3D view to rotate all of them together.

03

Force Modelling

VolCo encodes subtle penetrations, which allows us to model contact forces via spring-damper models, like in simulators. This connects the geometric clues with physical forces, which is crucial for stable grasp generation.

Summing per-voxel spring forces into a contact force
$$F = \sum k\,\delta$$

How to Generate with VolCo

VolCo \(\rightarrow\) Low-dimensional latent \(\rightarrow\) Diffusion \(\rightarrow\) Grasp

01

Volume VAE

Volume VAE framework

Describe the Volume VAE here: what it takes as input, how the volumetric contact is encoded into a compact latent, and what the decoder reconstructs. Two or three sentences are enough; point readers to the key components in the figure above.

02

VolCo Diffusion

VolCo diffusion framework

Describe the diffusion stage here: how volumetric contact latents are generated conditioned on the object, and how the generated contact is turned into the final grasp (e.g. local hand reconstruction followed by fitting). Mention any key design choices that make generation robust.

Generation Procedure

Starting from the object, VolCo generates volumetric contact and hand correspondence, reconstructs local hand surfaces, and fits the final grasp. Drag the slider to step through each stage.

Potted meat can
Toothpaste

BibTeX

@inproceedings{chen2026volco,
        title={VolCo: Volumetric Contact for High-Fidelity Human Grasp Generation},
        author={Chen, Zhuo and Cheng, Yihua and Leonardis, Ales and Chang, Hyung Jin},
        booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
        year={2026},
}