Estimation of Inverse Kinematics Solutions of a 2D Planar Robotic Manipulator using Feed-Forward Neural Network

Abstract

This study explores the application of machine learning algorithms, including Linear Regression, Decision Trees, and a Feed-Forward Neural Network, to solve the inverse kinematics problem for a 2R planar robotic arm. The dataset was generated using inverse kinematics equations for various end-effector positions within a rectangular workspace, resulting in 10,011 data points. The Neural Network model, consisting of an input layer, a hidden layer, and an output layer, was trained to predict joint angles from the end-effector coordinates. The performance of the Neural Network model was compared to traditional regression models using metrics such as Mean Square Error, Root Mean Square Error, Mean Absolute Error, and R2. The Neural Network model outperformed both Linear Regression and Decision Trees, achieving a Mean Square Error of 0.000031 on the validation set. The model was tested on new data from a semicircular path, demonstrating strong generalization capabilities. This method presents a promising approach to real-time, accurate robotic control without the need for explicit inverse kinematic equation derivation.