Learning autonomous motion generating dynamical systems from demonstration

Abstract

This thesis studies dynamical systems based learning methods at a proof-of-concept level. The purpose of dynamical systems is to generate motion. In particular, three different methods are studied in detail and implemented in software to judge their applicability for a real robotic system. These methods were chosen for the stability they guarantee for the dynamical system. The software was used with a real manipulator arm to reproduce taught motions and the reproduction data was recorded and studied. The results indicate that the methods are viable for learning robotic motions but caution should be exercised when using the dynamical system for motion generation.