Soft Actor Critic Swing Up of a Real Inverted Pendulum on a Cart

Abstract

The inverted pendulum, is a classical experiment widely used as a benchmark for research in control systems, due to its challenging dynamics. In this paper, Deep Reinforcement Learning is used to control a real inverted pendulum on a cart. The Soft Actor Critic algorithm with automatic entropy tuning is used to train an agent capable of acting as a controller. The agent is trained on real data collected on an episodic basis and learns to carry out the swing up control task successfully.