Object-dependent estimation of grasp posture and contact region of hand based on cluster analysis

This paper presents a data-driven framework for estimating grasp posture and contact region between hand and object by using clustered lower dimensional grasp features. Our framework extracts features of the grasp posture in order to estimate the contact region. Redundant dimension of the hand posture is removed depending on task, which we regard typical for the object. For this purpose, we used mixture principal component analysis (MPCA). This method enables to estimate clusters that can approximate grasp postures by using lower dimensional features. Estimation results of the contact region was obtained by using clustered lower dimensional features of MPCA and object features. These results show that our clusters by MPCA have stronger correlation with the contact region than clustered lower dimensional features obtained by PCA and k-means. Finally, estimation results of the grasp posture and the contact region for three objects are demonstrated. One of the result was compared with grasp posture synthesized by another method presented in the previous research. This comparison revealed that our framework was able to synthesize human-like natural grasp posture than the previous one.