Research map / 3D vision / Human pose estimation
UniDexGrasp++: Improving Dexterous Grasping Policy Learning via Geometry-Aware Curriculum and Iterative Generalist-Specialist Learning
Weikang Wan, Haoran Geng, Yun Liu, Zikang Shan, Yaodong Yang, Li Yi, He Wang
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 |
|---|---|---|---|
| replace category-based object curriculum with geometry-aware curriculum (GeoCurriculum) (ablation) | state-based policy · UniDexGrasp train | success rate (%): 79.4 -> 82.7 | Improved |
| add curriculum learning versus none (ablation) | state-based policy · UniDexGrasp train | success rate (%): 30.5 -> 79.4 (object curriculum) and 82.7 (GeoCurriculum) | Improved |
| use 5 curriculum stages instead of 4 (ablation) | GeoCurriculum · UniDexGrasp train | success rate (%): 82.7 -> 82.9 | No clear effect |
| add iterative fine-tuning in state-based setting (ablation) | state-based policy · UniDexGrasp train | success rate (%): 82.7 -> 84.0 | Improved |
| add geometry-aware clustering for specialist assignment in state-based setting (ablation) | state-based policy · UniDexGrasp test unseen objects | success rate (%): 77.9 -> 84.3 | Improved |
| add geometry-aware clustering in vision-based stage (ablation) | vision-based policy · UniDexGrasp train | success rate (%): 78.6 -> 85.4 | Improved |
| distill state specialists directly to vision generalist instead of via state generalist (ablation) | vision-based policy · UniDexGrasp train | success rate (%): 82.7 -> 85.4 | Improved |
| add iterative fine-tuning in vision stage (ablation) | vision-based policy · UniDexGrasp train | success rate (%): 82.5 -> 85.4 | Improved |
| use DAgger-based distillation instead of BC, GAIL or DAPG (ablation) | iGSL distillation · UniDexGrasp train | success rate (%): 61.4 (DAPG) -> 87.9 | Improved |
| increase number of clusters from 20 to 50 (ablation) | GeoClustering · vision-based train | success rate (%): 85.4 -> 77.2 | Hurt |
| use geometry-aware clustering instead of category-label clustering (ablation) | GeoClustering · vision-based train | success rate (%): 79.7 -> 85.4 | Improved |
| use PointNet+Transformer backbone instead of PointNet (ablation) | vision-based policy · UniDexGrasp train | success rate (%): 82.1 -> 85.4 | Improved |
Closest papers by meaning
- UniGraspTransformer: Simplified Policy Distillation for Scalable Dexterous Robotic Grasping (CVPR 2025 · similarity 0.89)
- DexGrasp Anything: Towards Universal Robotic Dexterous Grasping with Physics Awareness (CVPR 2025 · similarity 0.88)
- CADGrasp: Learning Contact and Collision Aware General Dexterous Grasping in Cluttered Scenes (NeurIPS 2025 · similarity 0.88)
- Cross-Embodiment Dexterous Grasping with Reinforcement Learning (ICLR 2025 · similarity 0.87)
- Efficient Residual Learning with Mixture-of-Experts for Universal Dexterous Grasping (ICLR 2025 · similarity 0.87)
- Scaffolding Dexterous Manipulation with Vision-Language Models (NeurIPS 2025 · similarity 0.86)
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.
