Research map / Image generation / Autoregressive and latent generation
Contact Map Transfer with Conditional Diffusion Model for Generalizable Dexterous Grasp 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 a conditional diffusion framework that transfers contact, part, and direction maps from a shape template to novel same-category objects (conditioned on shape similarity and task text), then recovers ShadowHand grasps via robust joint-based optimization.
Problem
Dexterous grasp generation that is both stable and generalizable to unseen objects, categories, and tasks; generative methods rely heavily on training data quality while analytical methods are slow and task-agnostic.
Approach
Reformulates grasp generation as transferring object-centric contact maps from a shape template via a dual-branch conditional diffusion model with bidirectional cross-attention adaptation, extended to a cascaded diffusion over contact, part, and direction maps, plus a robust optimization that filters unreliable points by recovering per-part joint positions from direction intersections.
Key results
- On task-agnostic grasping, Ours reaches 84.65% SR and 38.16% coverage vs. DexGraspNet 83.63% SR and ContactGen 73.00% SR, and UGG 70.50% SR.
- On unseen categories, Ours achieves 74.14% SR vs. RealDex 29.90% and DexGYS 39.16%.
- Real-world experiments with an Inspire hand report an average grasp success of 70% across four categories.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use joint contact, part, and direction maps instead of contact map only (ablation) | Ours · Seen Categories | Cov.: 33.55 -> 36.77 | Helped |
| use joint contact, part, and direction maps instead of contact map only (ablation) | Ours · Unseen Categories | Cov.: 35.66 -> 30.05 | Hurt |
| remove the adaptation module (ablation) | dual-branch conditional diffusion model · novel objects | grasp generation performance | Hurt |
| replace cascaded diffusion with three independent diffusion models (ablation) | joint contact, part, and direction transfer · novel objects | grasp generation performance | Hurt |
| replace cascaded diffusion with one unified model concatenating all three maps (ablation) | joint contact, part, and direction transfer · novel objects | grasp generation performance | Hurt |
| remove textual task descriptions (ablation) | conditional diffusion model · diverse manipulation tasks | task alignment | Hurt |
| remove robust optimization strategy (ablation) | grasp recovery using transferred maps · novel objects | SR: 76.14 -> 55.04 | Hurt |
| use contact, part, and direction maps with robust optimization rather than contact map only (ablation) | Ours · novel objects | SR: 76.14 vs. 55.04 | Helped |
| randomly select a different same-category template for each target (ablation) | Ours · Unseen categories | Cont. Err.: 0.0363 -> 0.0391 | Hurt |
| randomly select a different same-category template for each target (ablation) | Ours · Unseen categories | SR: 74.14 -> 75.36 | Helped |
| randomly select a different same-category template for each target (ablation) | Ours · Unseen categories | Consis.: 79.28 -> 77.19 | Hurt |
| transfer grasps from Ours rather than Ours-Contact (ablation) | Ours · Seen Categories | Consis.: 75.26 -> 83.60 | Helped |
Surprising findings
- Using the part and direction maps without the robust optimization filter lowers performance (55.04% SR vs. 76.14% with contact map only).
- Tink achieves the lowest penetration (1.14 mm) among task-agnostic methods despite its lower success rate.
Limitations
- Evaluated only on the ShadowHand (plus human MANO); cross-embodiment hands not tested.
- Relies on a template shape per category and on the quality of the retargeted template grasps.
- Tink outperforms it on penetration, and novel-category task alignment remains weaker for some metrics (e.g., coverage on unseen categories).
Experimental setup
Datasets: CapGrasp
Benchmarks: CapGrasp
Metrics: Success Rate Penetration Depth Contact Coverage Contact Error Consistency Chamfer Distance R-Precision P-FID
Baselines: DFC DexGraspNet ContactGen UGG Tink RealDex DexGYS
Builds on: DDPM ContactGen Tink UGG
Closest papers by meaning
- Grasp2Grasp: Vision-Based Dexterous Grasp Translation via Schrödinger Bridges (NeurIPS 2025 · similarity 0.89)
- RAGG: Retrieval-Augmented Grasp Generation Model (AAAI 2025 · similarity 0.89)
- DexGrasp Anything: Towards Universal Robotic Dexterous Grasping with Physics Awareness (CVPR 2025 · similarity 0.87)
- G-HOP: Generative Hand-Object Prior for Interaction Reconstruction and Grasp Synthesis (CVPR 2024 · similarity 0.87)
- CADGrasp: Learning Contact and Collision Aware General Dexterous Grasping in Cluttered Scenes (NeurIPS 2025 · similarity 0.87)
- DexHandDiff: Interaction-aware Diffusion Planning for Adaptive Dexterous Manipulation (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.