An experiment in robot discovery with ILP Gregor Leban, Jure ˇ Zabkar, and Ivan Bratko Faculty of Computer and Information Science, University of Ljubljana, Trˇ zaˇ ska cesta 25, SI-1001 Ljubljana, Slovenia gregor.leban,jure.zabkar,ivan.bratko@fri.uni-lj.si Abstract. We describe an experiment in the application of ILP to au- tonomous discovery in a robotic domain. An autonomous robot is per- forming experiments in its world, collecting data and formulating pre- dictive theories about this world. In particular, we are interested in the robot’s “gaining insights” through predicate invention. In the first ex- perimental scenario in a pushing blocks domain, the robot discovers the notion of objects’ movability. The second scenario is about discovering the notion of obstacle. We describe experiments with a simulated robot, as well as an experiment with a real robot when robot’s observations contain noise. 1 Introduction In this paper we describe an application of an ILP system Hyper [3] to a simple robotic domain. The autonomous robot observes its environment which in our experiments consisted of two movable and two non-movable boxes (Fig. 1). It can perform experiments in this environment and collect data about its performed actions and the resulting observations. Our goal is to provide the robot a learning system that enables it to automatically discover new useful notions by exploring the domain. This is also known as gaining insights about the domain. Although there has not been a unique generally accepted definition of the notion of insight, one definition that corresponds well to the present case study is as follows. An insight is a new piece of knowledge that enables a simplification of the robot’s current theory. A variant of this definition is: Insight is the discovery of a new concept (e.g. in logic: the discovery of a new predicate) that enables the formulation of a new theory using the current theory as background knowl- edge and the observations as learning examples. This second definition precisely corresponds to the insight in the presented experiments. We present two scenarios in which the robot learns the notions of movability and obstacle, respectively. In the movability scenario, where the robot’s task is to push different objects, the robot discovers the notion of object’s movability which helps it to explain the observations. In the obstacle scenario, the robot is moving around in the environment and discovers the notion of an obstacle, which helps to explain why in some cases it is not able to reach its desired goal position.