Contextual Policy Search for Ball-Throwing on a Real Robot

Contextual policy search is a popular approach to learn versatilely applicable skills with a robot. This work develops an implementation for learning such skills directly on a real robot, with the objective to throw a ball at varying target positions on the floor. Several approaches for the representation of the policy and for contextual policy search are examined in theory and in a simulated environment. With the goal to compare the qualities of different concepts for skill learning and determine a set of preferences for the application of these methods in a real experimental setup. In this talk I present and theoretically compare the most promising approaches for policy representation and contextual policy search in the test scenario and give an overview of the steps necessary to implement this skill learning task.