The article discusses the development of a unified diffusion-based dexterous grasp generation model, UGG. This model operates within the object point cloud and hand parameter spaces, using an all-transformer architecture to unify data from the object, the hand, and the contacts. The model can generate diverse grasping postures with a high success rate, overcoming the limitations of existing regression-based and generation-based methods. In addition to grasp generation, the UGG model can generate objects based on hand information, providing valuable insights into object design. The model shows excellent performance on the DexGraspNet dataset.
Publication date: 29 Nov 2023
Project Page: https://jiaxin-lu.github.io/ugg/
Paper: https://arxiv.org/pdf/2311.16917
