Research map / 3D vision / 3D generation

VolCo: Volumetric Contact for High-Fidelity Human Grasp Generation

NeurIPS 2026 · 3D vision › 3D generation · approach: Hand-object interaction generation

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VolCo, a hierarchical volumetric contact representation (local 8^3 grids around surface points with contact likelihood and CSE), and VolCoDiff, a VAE plus prior-guided latent diffusion model for grasp generation that reduces penetration and improves stability on GRAB and HO3D.

Problem

Human grasp generation for MANO hands suffers from severe penetrations and implausible results because existing surface-based contact maps lose hand detail and require hand-crafted, non-differentiable nearest-neighbor penetration losses during test-time adaptation.

Approach

Replace surface-only contact maps with local volumetric grids sampled around object surface points, where each grid point stores contact likelihood and a Continuous Surface Embedding; a 3D VAE compresses each volume conditioned on SDF, and a latent diffusion model with an auxiliary hand-prior branch generates the volume-wise latents, with a spring-damper force-based stability loss.

Key results

• On GRAB grasp generation, VolCoDiff reaches SD 0.52 cm and PD 0.27 cm, with CA 31.1 cm^2 (SD down 14.8% and PD down 32.5% versus the previous best).
• On out-of-domain HO3D, PD drops 28.6% and CA rises 18.2% relative to the best prior method, with SD 1.06 cm vs 1.21 for FAGrasp.
• In reconstruction from ground-truth contact on GRAB, the 8^3 volume setting reduces EPE to 6.33 mm from 11.19 mm for ManiDext (a contact-map baseline).

Surprising findings

• Stability loss reduces simulation displacement noticeably only when prior guidance is used, suggesting geometric plausibility is a prerequisite for physical stability.

Limitations

• Prefers only tighter grasps and offers limited controllability of the generated grasps.
• Fixed number and scale of volumes limit coverage of different object scales.

Experimental setup

Datasets: GRAB HO3D YCB

Benchmarks: GRAB HO3D

Metrics: EPE F-score AUC simulation displacement penetration depth intersection volume contact area contact ratio entropy cluster size

Baselines: ContactOpt ContactGen ManiDext GrabNet GraspTTA FastGrasp FAGrasp

Builds on: Scene Diffuser Mosaic-SDF ContactOpt ContactGen GrabNet

Closest papers by meaning

• LatentHOI: On the Generalizable Hand Object Motion Generation with Latent Hand Diffusion. (CVPR 2025 · similarity 0.87)
• Learning Dense Hand Contact Estimation from Imbalanced Data (NeurIPS 2025 · similarity 0.87)
• How Do I Do That? Synthesizing 3D Hand Motion and Contacts for Everyday Interactions (CVPR 2025 · similarity 0.86)
• Learning Explicit Contact for Implicit Reconstruction of Hand-Held Objects from Monocular Images (AAAI 2024 · similarity 0.86)
• Text2HOI: Text-guided 3D Motion Generation for Hand-Object Interaction (CVPR 2024 · similarity 0.86)
• MANUS: Markerless Grasp Capture using Articulated 3D Gaussians (CVPR 2024 · similarity 0.86)