An Overview of Data-Driven Paradigms for Identification and Control of Robotic Systems

Abstract

Fueled by the ever-growing availability of large-scale datasets and cutting-edge machine learning advances, data-driven approaches are revolutionizing the design, identification, and control of nonlinear robotic systems. This review paper examines this transformative paradigm, focusing on studies that utilize data-driven techniques involving the Koopman operator-theoretic framework, recurrent neural networks, and the Gaussian process regression for modeling and control of robotic systems. In particular, this study undertakes a review of these state-of-the-art data-driven methods, which have delivered significant performance improvement over a large class of robotic systems, including rigid manipulators, soft robots, and quadrotor aerial systems. The challenges, opportunities, and future directions across this dynamic landscape of data-driven robotics are also explored in this study with an emphasis on the interdisciplinary nature of this rapidly evolving field.