Neural-network based force planning for multifingered grasp

The real-time control of multifingered grasp involves the problem of the force distribution which is usually underdetermined. It is known that the results of the force distribution are used to provide force or torque setpoints to the actuators, so they must be obtained in real-time. The objective of this paper is to develop a fast and efficient force planning method to obtain the desired joint torques which will allow multifingered hands to firmly grasp an object with arbitrary shape. In this paper, the force distribution problem in a multifingered hand is treated as a nonlinear mapping from the object size to joint torques. We represent the nonlinear mapping using artificial neural networks (ANNs), which allow us to deal with the complicated force planning strategy. A nonlinear programming method, which optimizes the contact forces of fingertips under the friction constraints, unisense force constraints and joint torque constraints, is presented to train the ANNs. The ANNs used for this research are based on the functional link (FL) network and the popular back-propagation (BP) network. It is found that the FL-network converges more quickly to the smaller error by comparing the training process of the two networks. The results obtained by simulation show that the FL-network is able to learn the complex nonlinear mapping to an acceptable level of accuracy and can be used as a real-time grasp planner.