Developing Interpretable Machine Learning for Forward Kinematics of Robotic Arms

Abstract

Machine learning (ML) is becoming increasingly sought after in diverse domains. Unfortunately for this objective, most ML research has focused on improving performance on evaluation metrics such as accuracy. However, to make important decisions, ML models need to be interpretable. We propose an approach to approximate kinematics of a robotic arm using interpretable artificial neural networks (ANNs). This work is based on approximating nonlinear functions where domain knowledge and visually observable features of the data are used to design ANNs. After analyzing existing work, we present a feasibility study approximating the kinematics of a simplified robotic arm and extend the work to multiple hidden layers. We generalize the existing work and extend its use for a different application, noting the challenges that arise while extending this work to multiple hidden layers.