From Speech Recognition to Instruction Learning

In robotics, learning from demonstration is a research field that has gained a lot of attention since it provides an easy and intuitive way to teach a robot complex movements. When teaching a human it is common to use natural language to provide additional information. Regardless of that, it is rarely investigated how speech can be used in addition for teaching a robot with learning from demonstration. Therefore this thesis examines how natural language is already used in robotics and how the learning process can be supported further by using speech. Furthermore a learning from demonstration system is implemented based on task spaces that are an intuitive abstraction of the state of the robot and its relation to the environment and therefore are more correlated with spoken instructions. Additional a task space selection instruction and a correction instruction are introduced to use natural language to support the learning process. They are evaluated on a real robot in an experiment where the robot has to draw a point.