Learning-based modeling and control of underactuated balance robotic systems

Abstract

Underactuated balance robots represent a broad class of mechanical systems, ranging from Furuta pendulum, autonomous motorcycles, and robotic bipedal walkers, etc. The control tasks of these systems include trajectory tracking and balancing requirements. We present a data-driven modeling and control framework of the underactuated balance robots. A machine-learning method is used to capture the dynamics and the balance equilibrium manifold that represents balancing task target. We combine the learning-based models with the structural properties of the external/internal convertible form of these underactuated systems. Applications of the proposed learning-based models and control design are applied to the Furuta pendulum by simulation and experiments.