Application of reinforcement learning control to a nonlinear dexterous robot

In this paper, the effects of basic parameters in reinforcement learning control such as eligibility, action and critic network weights, system nonlinearities, gradient information, state-space partitioning, variance of exploration were studied in detail. We attempt to increase feasibility for practical applications, implementation, learning efficiency, and performance. Reinforcement learning is then applied for control of a nonlinear dexterous robot. This control problem dictates that the learning is performed online, based on binary and real valued reinforcement signal from a critic network, without knowing the system model nonlinearity. The learning algorithm consists of an action and critic networks that learn to keep the multifinger hand of the dexterous robot within desired limits