Grasping Gesture Optimization of Multi-fingered Dexterous Hands Based on Deep Neural Networks

Based on deep neural network model, a grasping gesture optimization method is proposed for multi-fingered dexterous hands. Firstly, a grasp dataset is constructed in simulation environment, and then a convolutional neural network is trained on this basis to predict the grasp quality function from the monocular visual information of the target object and the grasp configuration of the multi-fingered dexterous hand. Therefore, the grasp planning problem of the multi-fingered dexterous hands is transformed into an optimization problem about maximizing the grasping quality. Further the backpropagation and gradient ascent algorithm in deep learning is used to iterate and optimize the grasping gestures of the multi-fingered dexterous hands. In simulation, the evaluation results of the grasping quality, separately computed by the proposed network and the simulation platform for the same grasp configuration, are compared. Then the proposed method is implemented to optimize the initial gestures searched randomly, and the force closure metrics of the gestures before and after optimization are compared. Finally, the optimization performance of the proposed method is validated on the actual robot platform. The results show that the grasping success rate of the proposed method for the unknown objects is more than 80%, and for the failed grasps, the success rate after optimization reaches 90%.