Learning from demonstration with Gaussian processes

Abstract

There is huge potential in the field of robotics for the application of machine learning methodologies, particularly in the case of learning by demonstration, which considerably reduces the time required to program robotic actions, and in addition, makes robotic movements more natural. Within machine learning domain supervised, unsupervised and reinforcement learning classifications can be found. Among these, the most widely used is supervised learning. This allows two learning tasks: classification and regression. The Gaussian process model is one of the methodologies used for regression. Through regression, a learning process can be performed, allowing to learn by demonstration from a given data set. In this article, the development of a learning method is presented, it is based on Gaussian process regression and intended to be applied in robotic platforms which require to learn quickly and incrementally, since the robots today maintain more contact with the environment and therefore with the human. That is why the Gaussian processes have the characteristics required to develop this type of control for robots. In this paper, a non-parametric regression model such Gaussian process is investigated, as well as how this can be applied to learning from demonstration framework.