Unsupervised clustering of robot activities: A Bayesian approach

Our goal is for robots to learn conceptual systems sufficient for natural language and planning. The learning should be autonomous, without supervision. The first steps in building a conceptual system are to say some things are alike and others are different, based on how an agent interacts with them, and to organize similar things into classes or clusters. We use the BCD algorithm for clustering episodes experienced by our robots. The clusters contain episodes with similar dynamics, described by Markov chains.