Abstract
In this study, we propose a method to avoid singularity by selecting solution types of inverse kinematics for a six-degree-of-freedom (6-DOF) manipulator based on reinforcement learning. A general 6-DOF manipulator has eight solution types of inverse kinematics for any position/posture of the end-effector. Owing to the complex structure of cooperative robots to prevent pinching during cooperative operations, inverse kinematics is often solved using numerical methods. Because the numerical solution depends on the initial values, it is difficult to select suitable solution types. According to the selection of the solution types, the robot may pass through a singularity, causing some joints to rotate rapidly. To avoid this, solution types must be selected considering the entire motion path of the robot. Therefore, we constructed Deep Q-Learning (DQN), a type of reinforcement learning, to select the solution types that minimize the angular velocity of each joint during the motion path. This was verified by a 6-DOF cooperative robot, where the robot was commanded to take a path through the singularity, and the solution types of inverse kinematics were selected by the DQN. Consequently, singularity was avoided by selecting suitable solution types, and the angular velocity was minimized.
