A Fundamental Study of Sim-to-Real Deep Reinforcement Learning for Picking Motion using a Vacuum Suction Cup by Desktop Robotic Arm

Abstract

In this study, we propose a reinforcement learning method for motion instruction that aims to reduce the burden of learning on the actual robot. In this method, we construct a virtual environment for learning on a simulator, and the motion is instructed on the simulator as pre-training. The actual robot learns the motion based on the results of the pre-training. By utilizing the pre-training results, we aim to shorten the time required for learning on the actual robot and reduce the burden on the actual robot. As an example problem, we apply the proposed method to the picking motion using a vacuum suction cup of the desktop robotic arm. Through the application to the example problem, the feasibility and validity of the proposed method are basically investigated.