The paper focuses on a fundamental learning problem in adaptive, embodied cognitive systems: Namely, how to learn discrete models of situated, embodied experience which can act as a mediation between sensori-motoric experience and high-level cognitive processes. The paper suggests to address the problem using a combination of bottom up active learning of embodied concepts solely on the basis of the actions and perceptions of the robot, and top-down information obtained through interaction with other agents. The embodied concepts are constructed to be informative for the robot in terms of its sensorimotor prediction capability. From that point the effort of constructing humanlike concepts is shifted towards producing a translation between the sensorimotor based bottom-up ontology and more conventional top-down constructed ontologies. The suggested framework is based on a parameter free rule extraction algorithm that successfully has been applied to the problem of creating finite state descriptions of large, complex and even chaotic simulated dynamic systems. We will briefly describe how this algorithm can be ported to an autonomous robot domain.