Power grasp force distribution control using artificial neural networks

Abstract

In this article, methods for force distribution control of power grasp are developed. A power grasp is characterized by multiple points of contact between the object grasped and the surfaces of the fingers and palm. The grasp is highly stable because of form closure. However, modeling power grasps is difficult because of the resulting closed kinematic structure and the complexity of multiple contacts. The first method used to obtain the desired force distribution is based on linear programming. In particular, a model of the DIGITS grasping system, under development at The Ohio State University, is used, and constraint equations are formulated for force balance and actuator torque limits. Supervisory control of the desired forces at the contacts is achieved by prescribing a desired clinch level. The objective function is designed to achieve the desired clinch level, except in cases where the specified force is inadequate to stably hold the object. Although this method yields the desired force distribution, a second method based on artificial neural networks (ANNs) is developed to achieve constant-time solutions. Linear programming solutions are used to generate training data for a set of ANNs. Two techniques, modular networks and adaptive slopes, are also developed and employed in the training to improve the training time and accuracy of the ANNs. The results show that the ANNs learn the appropriate nonlinear mapping for the force distribution and provide stable grasp over a wide range of object sizes and clinch levels.