Towards autonomous design of experiments for robots

For understanding a real-world environment on a conceptual level, any agent requires the capability for autonomous, open-ended learning. One of the main challenges in Artificial Intelligence is to bias the learning phase sufficiently in order to obviate complexity issues, while at the same time not restricting the agent to a certain environment or to a particular task. In this paper we describe a framework for autonomous design of experiments for a robotic agent, which enables the robot to improve and increase its conceptual knowledge about the environment through open-ended learning by experimentation. We specify our implementation of this framework and describe how its modules can recognize situations in which learning is useful or necessary, gather target-oriented data and provide it to machine learning algorithms, thus reducing the search space for the learning target significantly. We describe the integration of these modules and the real world scenarios in which we tested them.