Postural Synergies and Neural Network for Autonomous Grasping: a Tool for Dextrous Prosthetic and Robotic Hands

In this paper, a neural network model has been designed for planning grasps of a cybernetic hand prototype by means of postural synergies. The synergies subspace is derived by means of a joint-to-joint mapping from a human hand set of grasps. A library of motor primitives of the hand in a synergy-based rendering has been built for a number of selected objects and tasks. The requirement of the task in a simplified approach is specified by the type of grasp, such as precision or power. A feedforward neural network has been trained using the grasping data from the library and running the Levenberg-Marquadt algorithm. By combining postural synergies and neural network the nonlinear relationship between the object geometric features and the hand configuration during grasping can be easily found with a good approximation. The experiments have been performed on the DEXMART hand prototype and the results demonstrate that integration of postural synergies and neural network is a promising tool toward simplified and autonomous grasping for artificial hands, such as anthropomorphic robotic hands and prostheses.