EMIKNet: Expanding Multiple-Instance Inverse Kinematics Network for Multiple End-Effector and Multiple Solutions

Abstract

Inverse Kinematics (IK) is essential for robot control but suffers from computational complexity as the number of links and end-effectors increases, thereby affecting real-time control accuracy and performance. In this letter, we propose a neural network-based IK approach to efficiently address the IK challenges in complex robotic systems with multiple end-effectors and methods for obtaining multiple solutions. The proposed method is constructed via interconnected Gaussian Mixture Models (GMM), which consist of two main components: (i) a hierarchical computation framework for calculating joint angles while considering joint limits, and (ii) a 'sub-link' concept to reduce the hierarchical computation. The experimental results of 7-Degrees of Freedom (DOF) robot arm manipulator with a single end-effector and 22-DOF robot arm-hand manipulator with multiple end-effectors show that the proposed method reduces both Euclidean Distance and Angular Distance as well as the Run-time, compared with the model-based numerical and earlier neural network methods. We also verify multiple solutions obtained through hierarchical computation of GMM.