Gaussian processes for flexible robot learning.

摘要

As robotic systems are becoming more complex and as they are developed for increasingly realistic environments, the role of machine learning within robotics becomes even more central than it has been. Robot learning tasks typically raise a number of hard requirements for learning methods including accurate estimation of uncertainty, dealing with inhomogeneously sampled data, computational efficiency, and high modeling accuracy.
<br>
<br>We show in this thesis that Gaussian processes as a machine learning method have the potential to address these issues. Gaussian processes are a Bayesian approach to function regression, that is, they allow to place a prior distribution over the space of functions. The approach has become one of the standard tools for solving non-linear regression problems, which are central to many machine learning problems. While the modeling capabilities of Gaussian processes are as high as for state-of-the-art alternatives, the basic concept is exceptionally clear and easy to implement.
<br>
<br>We present Gaussian process-based solutions to a wide variety of problems encountered in robotics. This includes the modeling of laser and gas sensors, the interpretation of camera images, the detection of failures during mobile robot navigation, and the estimation of probabilistic terrain maps from noisy range measurements. Furthermore, we introduce an approach that allows a robot to learn a model of its own body from scratch using only visual self-observation. In each of the studied problem domains, our solutions have been evaluated empirically using real and simulated data.
<br>
<br>As a contribution to machine learning in general, we study and improve extensions of the standard Gaussian process model, which can be used flexibly to solve non-robotics learning tasks also. This includes an approach to dealing with spatially-varying observation noise and one for adapting the function smoothness locally. We furthermore derive sparse approximations for these extensions to be able to apply the models to large data sets and in online settings.