Automatic Dexterous Functional Grasping Data Generation Based on Conditional Diffusion Model

This paper proposes an automatic dexterous functional grasping data generation framework for multi-finger robot hands. This work has important benefits for developing effective data driven dexterous grasping methods, because the data scarcity has delayed the relevant research progress currently. Taking the diffusion model as the main architecture, the proposed method leverages the advantages of high generation quality and diversity, designing a controlling condition by integrating the grasp pose and object category. After further candidates selection and point clouds refinement, the generated dataset have been validated through experiments to fulfill the requirements of functionality, diversity, scalability.