Deep Convolutional AutoEncoders as a Minimal State Representation for Reinforcement Learning in Industrial Robot Manipulators

Reinforcement Learning has become a very powerful technique in the last decade thanks to the technological advances in computational power and efficient learning algorithms in neuronal networks. Reinforcement Learning allows an autonomous agent to learn through experience how to solve a problem in an optimal way, with a minimal information of its environment, just a reward signal. This approach has lightened the way to an artificial intelligence capable of reaching superhuman performance in games such as Go and Backgammon, beating world champion players. Unfortunately this is far from a real world robot application due to many technological challenges. So in this paper we propose how to get a low-dimensional state representation for industrial robot manipulators, using visual information and Convolutional Autoencoders to speed up the information processing in the Reinforcement Learning training phase.