Swinging Up and Balancing a Pendulum on a Vertically Moving Cart Using Reinforcement Learning

Abstract

Underactuated systems occur frequently in robotics and legged locomotion. Unactuated pendulum on an actuated cart is a classic example used for designing and testing control algorithms for underactuated systems. While pendulum balancing on a horizontally moving cart is popular and environments available for reinforcement learning, pendulum on vertically moving cart is rarely discussed due to relatively higher difficulty level in balancing it. This paper presents a model environment for a pendulum on a vertically moving cart and trains a neural network controller using reinforcement learning to balance it in vertical position indefinitely without exceeding the displacement limits. Results presented for both con-tinuous and discrete force control input for the cart system show that the neural network controllers can successfully swing up and balance the pendulum.