Abstract
Grasping objects with a high degree of anthropomorphism is a critical component in the field of highly anthropomorphic robotic grasping. However, the accuracy of contact maps and the irrationality of the grasping gesture become challenges for grasp generation. In this paper, we propose a reasonably improved generation scheme, called Diffangle-Grasp, consisting of two parts: contact map generation based on a conditional variational autoencoder (CVAE), sharing the potential space with the diffusion model, and optimized grasping generation, conforming to the physical laws and the natural pose. The experimental findings demonstrate that the proposed method effectively reduces the loss in contact map reconstruction by 9.59% in comparison with the base model. Additionally, it enhances the naturalness by 2.15%, elevates the success rate of grasping by 3.27%, reduces the penetration volume by 11.06%, and maintains the grasping simulation displacement. The comprehensive comparison and qualitative analysis with mainstream schemes also corroborate the rationality of the improvement. In this paper, we provide a comprehensive account of our contributions to enhancing the accuracy of contact maps and the naturalness of grasping gestures. We also offer a detailed technical feasibility analysis for robotic human grasping.