Exploration of Neural Network Architectures for Inertia Parameter Identification of a Robotic Arm

Abstract

We propose a machine-learning-based approach for identifying inertia parameters of robotic systems. We evaluate the method in simulation and compare it against classical methods. Specifically, we implement parameter identification based on numerical optimization and test it using ground truth data. For a case study, we set up a physical simulation of a four-degree-of-freedom robot arm, formulating the problem with Newton-Euler equations as opposed to the conventional Lagrangian formulation at the joint level. Additionally, we derive a test methodology for assessing various Artificial Neural Network architectures.