Reproduce “ScaleADFG: Affordance-based Dexterous Functional Grasping via Scalable Dataset” 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
Sizhe Wang, Yifan Yang, Yongkang Luo, Daheng Li, Wei Wei, Yan Zhang, Peiying Hu, Yunjin Fu and 3 more
Researchers from CASIA, CasiaHand Robotics, and SenseTime developed ScaleADFG, an affordance-based framework for dexterous functional grasping that includes a large-scale, multi-hand, scale-variant dataset. This approach automates data generation using foundation models and demonstrates robust zero-shot transfer to real-world objects, achieving a 76.0% grasp success rate on a real robot.
Key insights
- Utilizing pre-trained foundation models for 3D asset generation and affordance retrieval provides an effective and scalable solution for constructing large, diverse datasets for robotic manipulation.
- Explicitly incorporating wide object scale variations into training data is crucial for developing grasp models that exhibit robust generalization to unseen object sizes in real-world scenarios.
- An affordance-based grasp synthesis approach, combined with efficient optimization and a lightweight prediction network, enables strong zero-shot transfer from synthetic training data to real-world functional grasping tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.