Application of reinforcement learning to dexterous robot control

We consider the use of reinforcement learning for control of a nonlinear dexterous robot. The control problem dictates that the learning is performed online, based on a binary reinforcement signal from a critic without knowing the system nonlinearity. The learning algorithm consists of an action and critic units that learned to keep the multifinger hand of the dexterous robot within expected limits. The multifinger hand is based on an “artificial muscle” concept, whereby the hand receives a probabilistic reinforcement signal (reward or penalty) and selects best control actions. The objective is to apply forces so as to keep the finger within the limits of the angular position and velocity at each link. The nonlinear sigmoidal transfer function has been chosen for replacing the original discontinuous binary threshold function during the learning rule evaluation