The Teaching-Box: A universal robot learning framework

Abstract

There exist many powerful machine learning software libraries, which help the engineer to build robots that learn autonomously. However, engineering of an autonomous robot still is a challenging and time consuming task even with these learning libraries. With the open source Teaching-Box presented here, the “training” of a robot becomes easier due to the following features. The Java library of the Teaching-Box provides algorithms for reinforcement learning as well as for learning by demonstration (utilizing supervised learning algorithms) and data structures for exchanging policies between the different ways of learning. As an initial policy one can even take a manually coded behaviour and then improve it for example with reinforcement learning. A human trainer feedback (e.g. via the speech interface) can be used to increase the learning speed. The Eclipse based GUI facilitates the design of the robot learning projects and visualizes the learning process. For connecting the various modules of a project, open interface standards such as RL-Glue are used and an easy integration of the Teaching-Box into standard robot middleware is possible.