A neural network interface to the DIGITS Grasping System

A neural-network-based interface between an operator and the DIGITS (dexterous integrated grasping with intrinsic tactile sensing) grasping system is proposed, and the initial results of the network training are presented. The neural network is responsible for accepting the description of an object to be held in a power grasp, and mapping these data into a set of actuator torques which will allow DIGITS to firmly grasp the object. The network should attempt to maximize the normal forces on the object to provide the best possible grasp while not exceeding a set level provided by the operator. The backpropagation neural network was trained with various quantities of hidden nodes and learning rates and then tested for stability and error with respect to the optimal solution. Useful results concerning the effect of learning rate and number of hidden nodes were obtained, as well as results indicating that the network can accurately determine torques for both trained and untrained objects