Evolution of Deep Belief Neural Network Parameters for Robot Object Recognition and Grasping

Abstract

Robot object recognition and grasping is an important research area in robotics. Recently, deep learning is gaining popularity as a powerful mechanism for object recognition. Deep learning has very complicated configurations including network structures and several parameters, such as the number of hidden units and the number of epochs, which influence the performance and computation time. Determining such parameters require high expertise in deep learning. Thus, the development of deep learning is limiting in the skilled experts. In this work, we combine Deep Belief Neural Network (DBNN) and evolutionary algorithm in order to improve the performance and reduce the computation time. To verify the performance, robot object recognition and grasping is considered. Experimental results show that our method outperforms on object recognition and robot grasping tasks.