Grasping Control of a Robot Hand by Reinforcement Learning

Abstract

It is very useful to apply a reinforcement learning for controlling a robot hand with tactile sensors which can grasp and manipulate an object delicately like a human hand. A reinforcement learning based on trial and error is proposed here, which is expected to learn autonomously the optimum manipulation from experiences. In computer simulations, the learning algorithm is applied to controlling a simple hand with two fingers and four fingers to investigate its validity . As a result, it has acquired autonomously almost the optimum control for the manipulation of the hand to grasp and convey an object. Therefore, the learning algorithm proposed may be useful basically for controlling a robot hand.