Browsing by Author "Kulick, Johannes"

Now showing 1 - 1 of 1

Open Access
Information driven exploration in robotics
(2016) Kulick, Johannes; Toussaint, Marc (Prof. Dr.)
Imagine an intelligent robot entering an unknown room. It starts interacting with its new surroundings to understand what properties the new objects have and how they interact with each other. Finally, he gathered enough information to skillfully perform various tasks in the new environment. This is the vision behind our research towards intelligent robots. An important role in the described behavior is the ability to chose actions in order to learn new things. This ability we call exploration. It enables the robot to quickly learn about the properties of the objects. Surprisingly autonomous exploration has been mostly neglected by robotics research so far, because many fundamental problems like motor control and perception were still not satisfactory solved. The developments of recent years have, however, overcome this hurdle. State of the art methods enable us now to conduct research on exploration in robotics. On the other hand the machine learning and statistics community has developed methods and the theoretical background to lead learning algorithms to the most promising data. Under the terms active learning and experimental design many methods have been developed to improve the learning rate with fewer training data. In this thesis we combine results from both fields to develop a framework of exploration in robotics. We base our framework on the notion of information and information gain, developed in the field of information theory. And although we show that optimal exploration is a computational hard problem, we develop efficient exploration strategies using information gain as measure and Bayesian experimental design as foundation. To test the explorative behavior generated by our strategies we introduce the Physical Exploration Challenge. It formalizes the desired behavior as exploration of external degrees of freedom. External degrees of freedom are those the robot can not articulate directly but only by interacting with the environment. We present how we can model different exploration tasks of external degree of freedom: Exploring the meaning of geometric symbols by moving objects, exploring the existence of joints and their properties, and exploring how different joints in the environment are interdependent. Different robots show these exploration tasks in both simulated and real world experiments.