Reproduce “AdaDexGrasp: Adaptive Dexterous Grasping via 3D Visuo-Tactile Representation Fusion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xirui Liang, Jiaqi Liang, Jingkai Xu, Yuran Wang, Ruochong Li, Yuanpei Chen, Masayoshi Tomizuka, Wei Zhan and 1 more
AdaDexGrasp presents a unified framework for adaptive dexterous grasping by tightly fusing 3D visual geometry and tactile feedback for initial planning and post-contact refinement. This approach achieved 83% grasp success on unseen object categories in simulation and 81% in real-world settings, outperforming baseline methods.
Key insights
- Semantic contact awareness, achieved by labeling object points with specific hand part IDs (e.g., finger/palm) during grasp generation, significantly improves both initial planning and subsequent adaptation.
- The iterative, closed-loop refinement process, which leverages real-time visuo-tactile information, is crucial for enhancing grasp robustness and generalization, particularly for novel objects.
- A multi-modal representation combining object geometry (3D point cloud) with RGB and a 6-dimensional tactile intensity vector (per point) balances informational density and sim-to-real transfer feasibility.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.