Comparing Data-Driven motion tracking controllers for a Flexible-Joint Robotic Manipulator

Abstract

Modeling and control of flexible robotic manipulators in collaborative robotics applications, face key issues when it comes to properly including non-linearities but keeping motion models and controllers easy to handle. Machine learning (ML) strategies stand as well suited solutions to obtain simplified models and derive controllers for flexible-joints or flexible-links manipulators. In the present paper data-driven dynamics analysis and controller design for a Flexible-Joint Robotic Manipulator (FJRM) are presented. The FJRM under study is a planar two-DOF manipulator with two flexible-joints and two rigid-links with a switched dynamics. The implementation hereby described is determined by a comparative analysis developed between direct and indirect data-driven controllers. Firstly, state-space feedback is proposed from an experimentally identified model as an indirect framework. Secondly, a Neural PID is designed and developed directly from data. The comparison results allowed to identify the most appropriate controller topology to implement.