Robotic Grasping: A Generic Neural Network Architecture

Abstract

In this chapter, a new model was proposed to define the kinematics of various robotic structures including an anthropomorphic arm and hand as well as industrial or service robots (MANUS). The proposed method is based on two neural networks. The first one is dedicated to finger inverse kinematics. The second stage of the model uses reinforcement learning to define the appropriate arm configuration. This model is able to define the whole upper limb configuration to grasp an object while avoiding obstacles located in the environment and with noise and uncertainty. Several simulation results demonstrate the capability of the model. The fact that no information about the number, position, shape and size of the obstacles is provided to the learning agent is an interesting property of this method. Another valuable feature is that a solution can be obtained after a relatively low number of iterations. One can consider this method as a part of a larger model to define robotic arm postures that tackles the “kinematical part” of the problem and can be associated with any grasp synthesis algorithm. In future work, we plan to develop algorithms based on unsupervised learning and Hopfield networks to construct the upper-limb movement. In this way, we will be able to generate an upper-limb collision free trajectory in joint coordinate space from any initial position to the collision free final configuration obtained by the method described in this article.