Grasp synthesis for a dextrous robotic hand using a multi-layer perceptron

A neural-network-based approach to the synthesis of the grasp for a robotic dexterous hand is presented. After a discussion of categorization techniques for the representation of geometrical aspects of the grasp, two different experiments are illustrated, exploiting multi-layer perceptrons to choose a set of feasible grasps in terms of opposition modes or location of contacts on the hand. The information fed to the networks includes the geometry of the object, its location in the workspace, and the kinematics of the hand. Performance results reveal good generalization capabilities