Neural network control of an inverted pendulum on a two DoF cart moving in the vertical plane

Abstract

Inverted pendulum control finds similarities with control of legged robots such as bipedal or humanoid robots where the trunk is balanced in an upright position. This paper proposes and presents a two degree of freedom cart that can move in the vertical plane while holding an underactuated pendulum. A neural network controller is trained using reinforcement learning to swing up and balance the pendulum by applying appropriate forces along horizontal and vertical directions. Simulation results show successful swinging up and balancing of the underactuated pendulum by applying horizontal and vertical forces on the cart while simultaneously keeping the cart within horizontal and vertical limits on displacements.