Reinforcement learning for manipulator control

Nataly P. Koshmanova, Dmitry S. Trifonov, Vladimir E. Pavlovsky

Keldysh Institute of Applied Mathematics RAS, Miusskaya st. 4, Moscow, 125047, Russia

Abstract: We present method for constructing manipulator control system with reinforcement learning algorithm. We construct learning algorithm which uses information about performed actions and their quality with respect to desired behaviour called «reward». The goal of the learning algorithm is to construct control system maximizing total reward. Learning algorithm and constructed control system were tested on the manipulator collision avoidance problem.

Keywords: reinforcement learning, manipulator, control, Newton–Euler algorithm.

UDC: 62-503.56

MSC: 68T40, 93C85

Received: 17.07.2012
Revised: 24.08.2012