Efficient Learning of Grasp Selection for Five-Finger Dexterous Hand

This paper proposes a novel learn-to-search method to efficiently learn grasp configuration for five-finger dexterous hand. The framework is divided into three stages. Firstly, a rectangle is learned for locating an approximate grasp region; secondly, the thumb position is searched within this rectangle; thirdly, the position of other four fingers is learned. In three different stages, different coordinate systems are established and different features are extracted, applying SVM-rank algorithm to train the three-stage learning model. The experimental results show that, the proposed method can generate feasible grasp configurations for the five-finger dexterous hand and avoid excessive computation when searching five grasp points directly within the entire image.