An Improved Approach for Inverse Kinematics and Motion Planning of an Industrial Robot Manipulator with Reinforcement Learning

Abstract

Robot manipulators have become very popular in many application areas in the last decades. Typically, simplified robot models are used to apply machine learning algorithms, so the particular challenges of using real robots are not taken into account (e.g., joint angle constraints). This work contributes a new approach for controlling manipulators that solves both the inverse kinematics problem and the path planning problem.The paper uses a Deep Deterministic Policy Gradient (DDPG) agent to learn the motion of a robot manipulator designed for industrial use. It introduces a new state space to reliably learn general motions with user-selectable start and target positions with a high success rate. A new reward function improves the robot’s path to the target position. In addition, the learned motions result in densely occupied paths, so that the presented learning approach can also be used for path planning of robot manipulators.The new approach outperforms other learned state-of-the-art approaches for solving both the problems of inverse kinematics and of path planning of arm-like robots in terms of success rate and mean error. The approach was applied to a physical robot device and can be easily transferred to other path planning problems.