Deep Learning Framework for Inverse Kinematics Mapping for a 5 DoF Robotic Manipulator

Abstract

Robotic manipulators have several applications, such as in manufacturing, surgery, transport, etc. Appropriate control techniques are essential to avoid undesirable consequences. Deep learning has been shown to be useful in robotic manipulator control. This paper presents a deep learning frame-work for the mapping of inverse kinematics (IK) for as-degree of freedom robotic manipulator. The framework provides a mapping from joint angles to end-effector position and orientation. Inputs used for the networks are the desired trajectory points and outputs are the joint angles. Additionally, a vector-based mean absolute error loss function is proposed for the training of different deep learning networks. The framework is investigated based on the position error and orientation error between the calculated and actual trajectory, and the computational time required to predict the joint angle values for the reference trajectory. The results show that the implementation of neural networks facilitated the quicker prediction of the joint angles. The best joint angle prediction in terms of minimum position error with the least amount of time is provided by the Deep Neural Network, whereas Long Short Term Memory performs better for orientation error.