Research map / 3D vision / 3D generation
GeoDexGrasp: Geometry-aware Generation for Data-efficient and Physics-plausible Dexterous Grasping
Bing Han, Weiyuan Liu, Changlong Zhang, Chenxi Wang, Zhibin Zhao, Zhi Zhai
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| decouple root rotation (SO(3)) from finger joints and translation (Euclidean) (ablation) | GeoDexGrasp · DexGRAB | success rate (%): 41.8 -> 56.1 | Improved |
| add pose and interaction-direction representations (ablation) | GeoDexGrasp (decoupled) · DexGRAB | success rate (%): 56.1 -> 58.6 | Improved |
| add shape and scale representations (ablation) | GeoDexGrasp (decoupled) · DexGRAB | penetration depth (mm): 31.5 -> 19.9 | Improved |
| add self-supervised disentanglement pretraining of the SIM(3) encoder (ablation) | GeoDexGrasp · DexGRAB | success rate (%): 63.2 -> 67.5 | Improved |
| use geometry-aware SIM(3)-equivariant representations instead of DexGraspAnything | dexterous grasp generation · Five datasets average | penetration depth (mm): 21.2 -> 13.6 | Improved |
| use geometry-aware representations with far fewer parameters | GeoDexGrasp vs DexGraspAnything · Five datasets average | success rate (%): 58.5 -> 60.1 with 159.7M -> 28.7M parameters | No clear effect |
| train on reduced data (25% of grasps per object) | GeoDexGrasp vs UGG and DexGraspAnything · UniDexGrasp and DexGraspNet | success rate | Improved |
| geometry-aware model on out-of-distribution object sizes | GeoDexGrasp · Size generalization (Cube, Cylinder, Torus) | success rate (%): 56.3 -> 68.8 on large cube | Improved |
| geometry-aware model on topologically complex Toy horse | GeoDexGrasp · Size generalization, Toy horse | success rate (%): 9.4 (large) | No clear effect |
| explicitly extract scale factor | GeoDexGrasp · DexGRAB | success rate | No clear effect |
| geometry-aware model on RealDex | GeoDexGrasp · RealDex | success rate (%): 44.8 (DexGraspAnything) vs 38.3 | Hurt |
Closest papers by meaning
- Contact Map Transfer with Conditional Diffusion Model for Generalizable Dexterous Grasp Generation (NeurIPS 2025 · similarity 0.89)
- GraspXL: Generating Grasping Motions for Diverse Objects at Scale (ECCV 2024 · similarity 0.89)
- DexGrasp Anything: Towards Universal Robotic Dexterous Grasping with Physics Awareness (CVPR 2025 · similarity 0.89)
- GEARS: Local Geometry-aware Hand-object Interaction Synthesis (CVPR 2024 · similarity 0.88)
- Grasp2Grasp: Vision-Based Dexterous Grasp Translation via Schrödinger Bridges (NeurIPS 2025 · similarity 0.88)
- CADGrasp: Learning Contact and Collision Aware General Dexterous Grasping in Cluttered Scenes (NeurIPS 2025 · similarity 0.88)
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.
