Learning Optimal Grasping Posture of Multi-Fingered Dexterous Hands for Unknown Objects

Abstract

Inspired by sophisticated grasp of human, we propose novel multi-level Convolutional Neural Networks (CNNs) for finely grasping of unknown objects with multi-fingered dexterous hands, and design a quantitative evaluation method for grasping quality. The proposed multi-level CNNs consist of four levels with different structures and functions, which can effectively imitate the grasp planning process of human: locating the pose of the grasped object, selecting the grasping part and determining the optimal grasping posture. The complete multi-level CNNs achieve mapping from the RGB-D image of the grasped object to the pose and posture of the multi-fingered dexterous hand. Based on the force closure metric, a quantitative evaluation method is developed to analyze the grasping quality in the simulator GraspIt!. The grasping experiments are carried out on an actual Shadow hand, and the experimental results indicate that the proposed multi-level CNNs can achieve finely grasping of unknown objects, by calculating the success rates and the quantitative evaluation values.