Evolutionary synthesis of grasping through self-exploratory movements of a robotic hand

Abstract

This paper explores an evolutionary approach extended by developmental processes (emhryogenic evolution) to evolve adaptive neural controllers for different robotic platforms. These controllers are able to grow, learn, and adapt to different tasks. We use a PC cluster and a physically realistic simulator of a robotic hand to synthesize grasping from random movements. We present the ligand-receptor concept that can be used by artificial evolution to explore (a) the growth of a neural network, (b) value systems, and (c) learning mechanisms for a given task (grasping). Different objects require different grasps, when we pick up a glass, manipulate a screwdriver, or turn the pages of a book, our fingers move very differently. The position of the hand also varies. That is a fundamental problem for a robot, because it either needs to be pre-programmed to handle every object it might encounter in the future and its possible orientations, or it must be able to learn to adjust its grasp according to what it sees and feels. Thus why a neural controller should be capable to explore its own movement capabilities, reconfigure itself to cope with environmental and morphological changes, and coherently adapt its behavior to new situations. The results show that this self exploratory activity can make the robot more robust and adaptive, and that grasping can be produced from totally random and independent movements of the fingers generated intrinsically by the neural controller.