Reinforcement-Learning-based Control of an Industrial Robotic Arm for Following a Randomly-Generated 2D-Trajectory

Abstract

This paper introduces a reinforcement-learning-based method to control an industrial robotic arm. The goal is to improve the process of teaching the robotic arm to perform a full-scan of the surface of free-formed components during an automatic quality inspection. As today’s standard approach, the teaching of a robotic arm to follow a complicated trajectory is being done manually by experts. However, this approach is not always reliable and might not necessarily result in an optimal trajectory. As a consequence, a lot of time and effort by a human expert is required. In addition, when a new component with a modified design has to be inspected, the human expert must create a suitable inspection trajectory for this component and meanwhile the whole process of automatic inspection has to be stopped for a considerably long time. The latter forms the biggest motivation of the presented work. The main focus of this work is to examine whether the teaching task can be accelerated by employing reinforcement learning methodologies to use the available information about the shape of the component to control the robot’s tool center point (TCP). As an initial step, a simulation environment has been developed, in which the position and orientation of the robot’s TCP can be controlled. During the simulation episodes a randomly generated 2D trajectory appears on the panel in front of the robot, and while observing the points on the trajectory, the robot trained with Deep Deterministic Policy Gradient algorithm follows the trajectory, and its goal is to reach the end point, while minimizing the time and the deviation from the trajectory. The initial results from the simulation environment and the steps to be taken in the future are explained in the given paper.