Learn to swing up and balance a real pole based on raw visual input data

Abstract

For the challenging pole balancing task we propose a system which uses raw visual input data for reinforcement learning to evolve a control strategy. Therefore we use a neural network - a deep autoencoder - to encode the camera images and thus the system states in a low dimensional feature space. The system is compared to controllers that work directly on the motor sensor data. We show that the performances of both systems are settled in the same order of magnitude. © 2012 Springer-Verlag.