Performing the Inverse Kinematics of an ABB 6 DOF Robotic Arm with Feedforward and Recurrent Neural Networks

Okoh, Christopher and Secco, Emanuele Lindo (2026) Performing the Inverse Kinematics of an ABB 6 DOF Robotic Arm with Feedforward and Recurrent Neural Networks. Journal of Artificial Intelligence & Control Systems. ISSN 3135-0682 (Accepted for Publication)

Text JAICS-18 Revised_v2_ES.pdf - Accepted Version Restricted to Repository staff only until 1 March 2027. Download (723kB)

Abstract

Conventional numerical Inverse Kinematics (IK) methods—such as the Jacobian Pseudo-inverse, Levenberg-Marquardt (LM), and Newton-Raphson algorithms—are known to be applied when solving the non-linear IK problems of a robot arm. This is done by taking an initial guess of the joint angles and iteratively or sequentially calculating small adjustments using the robot's Jacobian matrix until the end-effector converges onto the target 6-DoF pose. While highly accurate, these traditional techniques suffer from several fundamental limitations such as high computational cost and slower inference, high dependency on initial guess and inability to handle multi-modal solutions and thus make them challenging to deploy in dynamic, real-time, or highly complex robotic environments.
This paper leverages on the dataset of an ABB IRB120 6 Degree Of Freedom (DOF) robot arm from Kaggle where end-effectors’ positions obtained by means of Forward Kinematics (FK) are fed into the neural network models. And the derived joint angles were mapped to the end-effectors’ position thereby showing a faster method of providing IK solutions. The datasets were trained individually on Long-Short Term Memory (LSTM) neural network, Feed-forward Neural Network (FNN), and Recurrent Neural Network (RNN) using Keras on TensorFlow.
According to a performance analysis of the results, the LSTM model showed an MSE of 0.0001 rad2 with an execution time of 0.4 secs which translates into a highly precise inverse mapping capable of accurately guiding the robot through its Cartesian trajectory while satisfying latency constraints during robot deployment. Other experimental evaluations made are the MAE, R2 score and the End-effector error (Mean/Max) to show a high precision and minimal Euclidean distance error of the End-effector’s 3D Cartesian coordinates (x, y, z). Comparisons were made with previous results that performed IK evaluations using neural networks.

Item Type:	Article
Faculty / Department:	Faculty of Human and Digital Sciences > School of Computer Science and the Environment
Depositing User:	Emanuele Secco
Date Deposited:	30 Oct 2026 10:19
Last Modified:	30 Oct 2026 10:19
URI:	https://hira.hope.ac.uk/id/eprint/5009

Actions (login required)

View Item