Autonomous Knowledge Representation for Efficient Skill Learning in Cognitive Robots

This work explores the effects of the introduction of variational autoencoder based representation learning, and of its resulting latent spaces, within a robotic cognitive architecture to be able to efficiently learn models and policies when raw perceptual dimensionality is very high. The main focus of the paper is on the decision processes of the robots used for action selection. To this end we propose a procedure to obtain from autonomously produced latent state spaces the world and utility models necessary for deliberative operation as a first type of decision process. Additionally, we present a neuroevolutionary based approach to generate policies, for reactive operation, based on the information of the latent state space and using the previously obtained world and utility models to permit offline learning. A set of experiments over a real robot using vision, with the consequent high dimensional raw perceptual space, are carried out in order to validate the proposal.