MODEL-FREE ONLINE REINFORCEMENT LEARNING OF A ROBOTIC MANIPULATOR

Jerry Sweafford, Jr. and Farbod Fahimi

Keywords

Reinforcement learning, neural networks, trajectory tracking, model-free control, robotic manipulator

Abstract

For robotic systems that do not have joint torques as direct inputs, the computed torque approach, often based on a general dynamic model that is linear with respect to joint velocities and accelerations, is very dicult to implement. We have applied a general model-free online reinforcement learning control methodology for discrete non-affine nonlinear multiple-input-multiple-output systems to a second-order robotic system. The controller produced effective trajectory tracking in simulations of a two-degree- of-freedom robotic arm, and in actual experiments with a three degree-of-freedom robotic arm.

Important Links:



Go Back