Throwing Motion by Flexible Robot Arm using Deep Reinforcement Learning

Abstract

This paper addresses how to control a robot arm with variable stiffness actuators (VSAs) using deep reinforcement learning. Each VSA has two manipulated variables, target position and stiffness, thereby making learning speed inefficient due to increase of action space. To avoid the increase of action space including deviations of actuators’ stiffness, two policies for position control and stiffness control are explicitly distinguished, instead of end-to-end learning. In a ball throwing motion, the robot therefore learns how to swing its end effector firstly, and later, how to adjust stiffness to accelerate swinging motion. As a result, the robot in a dynamical simulator succeeded in acquiring the ball throwing motion without suffering from huge action space on end-to-end learning.