Throwing a Ball of the Pneumatically-Controlled Continuum Robot Arm using Reinforcement Learning with Movement Primitives

Abstract

Previous research on reinforcement learning for continuum robot arms have been dealt with a relatively small number of active degrees of freedom and made experiments of simple tasks such as reaching. We aimed to learn to throw a ball by reinforcement learning in a pneumatically-controlled continuum robot arm that has nine actuators. We adopt Cost-regularized Kernel Regression (CrKR) which uses dynamic movement primitives (DMPs) which is one of the movement primitives. In the simulation, the continuum arm was able to learn how to throw a ball forward. We made the same experiment for our real continuum robot arm and found that the learning progressed in the experiment.