Research map / 3D vision / 3D generation
CADGrasp: Learning Contact and Collision Aware General Dexterous Grasping in Cluttered Scenes
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 CADGrasp, a two-stage dexterous grasping method that predicts a sparse interaction bisector surface (IBS) representation from single-view point clouds with an occupancy-diffusion model, then optimizes grasp poses under IBS-based energies for cluttered scenes.
Problem
Dexterous grasping in cluttered scenes from single-view partial point clouds is hard due to high DoF, occlusion, and collisions; end-to-end mappings from observations to grasp poses generalize poorly, and prior two-stage methods need complete object geometry.
Approach
Predict a scene-decoupled, contact- and collision-aware sparse IBS volume (occupancy of IBS surface plus thumb and other-finger contact voxels) as an intermediate target, generated by a voxel-conditioned occupancy-diffusion model with force-closure ranking, and then optimize the Leap Hand pose with IBS-based energies.
Key results
- On GraspNet-1Billion dense scenes, CADGrasp reaches 86.5% success vs 83.3% for DexGraspNet2.0 (and 77.6% on Allegro zero-shot).
- Real-world clutter tests: 93.8% (30/32) overall success vs 83.9% for DexGraspNet2.0.
- Ablation: removing thumb/other-finger decomposition drops success from 86.5% to 56.1%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CADGrasp instead of DexGraspNet2.0 | Dexterous grasping in cluttered scenes · GraspNet-1Billion dense | success rate (%): 83.3 -> 86.5 | Helped |
| Use CADGrasp instead of DexGraspNet2.0 | Dexterous grasping in cluttered scenes · GraspNet-1Billion random | success rate (%): 79.5 -> 85.5 | Helped |
| Use CADGrasp instead of DexGraspNet2.0 | Dexterous grasping in cluttered scenes · GraspNet-1Billion loose | success rate (%): 73.9 -> 80.1 | Helped |
| Use CADGrasp instead of DexGraspNet2.0 | Dexterous grasping in cluttered scenes · ShapeNet dense | success rate (%): 81.5 -> 79.3 | Hurt |
| Use CADGrasp instead of DexGraspNet2.0 | Dexterous grasping in cluttered scenes · ShapeNet random | success rate (%): 77.1 -> 77.7 | Helped |
| Use CADGrasp instead of DexGraspNet2.0 | Dexterous grasping in cluttered scenes · ShapeNet loose | success rate (%): 73.5 -> 75.7 | Helped |
| Use CADGrasp instead of DexGraspNet2.0 | Real-world cluttered grasping · Real-world clutter scenes overall | success rate: 83.9% (26/31) -> 93.8% (30/32) | Helped |
| Add Decompose to model thumb contact separately from other fingers (ablation) | CADGrasp · GraspNet-1Billion dense | success rate (%): 56.1 -> 86.5 | Helped |
| Add IBS Ranking (ablation) | CADGrasp with Grasp Pose Ranking and Decompose · GraspNet-1Billion dense | success rate (%): 83.9 -> 86.5 | Helped |
| Add Grasp Pose Ranking (ablation) | CADGrasp with IBS Ranking and Decompose · GraspNet-1Billion dense | success rate (%): 75.7 -> 86.5 | Helped |
| Set voxel size to 2.5 mm instead of 5 mm (ablation) | CADGrasp · Voxel resolution ablation | success rate (%): 81.0 -> 81.2 | No clear effect |
| Set voxel size to 10 mm instead of 5 mm (ablation) | CADGrasp · Voxel resolution ablation | success rate (%): 81.0 -> 72.4 | Hurt |
Surprising findings
- Decomposing thumb and other-finger contact representation is critical: without it success falls from 86.5% to 56.1%.
- Single-hand IBS predicted for Leap Hand transfers zero-shot to Allegro Hand with comparable success.
Limitations
- Struggles with small objects (lower success for volumes below 0.00025 m³).
- Second-stage optimization is slow due to DDPM sampling of IBS; DDIM suggested as fix.
- Not real-time at 6.51 s per grasp.
Experimental setup
Datasets: GraspNet-1Billion ShapeNet DexGraspNet2 scenes
Benchmarks: DexGraspNet2 test scenes (GraspNet-1Billion and ShapeNet, dense/random/loose) real-world 5-scene clutter test with Leap Hand
Metrics: success rate runtime
Baselines: DexGraspNet2.0 HGC-Net ISAGrasp GraspTTA
Builds on: DexGraspNet 2.0 Interaction Bisector Surface GraspTTA DDPM ResUNet14
Closest papers by meaning
- DexGrasp Anything: Towards Universal Robotic Dexterous Grasping with Physics Awareness (CVPR 2025 · similarity 0.90)
- Contact Map Transfer with Conditional Diffusion Model for Generalizable Dexterous Grasp Generation (NeurIPS 2025 · similarity 0.87)
- Scaffolding Dexterous Manipulation with Vision-Language Models (NeurIPS 2025 · similarity 0.87)
- DiffGrasp: Whole-Body Grasping Synthesis Guided by Object Motion Using a Diffusion Model (AAAI 2025 · similarity 0.87)
- ZeroGrasp: Zero-Shot Shape Reconstruction Enabled Robotic Grasping (CVPR 2025 · similarity 0.87)
- SparseDFF: Sparse-View Feature Distillation for One-Shot Dexterous Manipulation (ICLR 2024 · 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.