Learning and motion planning are powerful methods that exhibit complementary strengths. While planning allows goal-directed actions to be computed when a reliable forward model is known, learning allows such models to be obtained autonomously. In this paper we outline how both methods can be combined using an expressive qualitative knowledge representation as a link. We argue that the crucial step in this integration is to employ a representation based on a well-defined semantics which empowers reasoning to boost performance of learning as well as of the resulting action plans. We present an architecture for learning based robotics that exploits qualitative reasoning. Expected beneficiaries of this approach are discussed, some of which are already demonstrated in proof-of-concept experiments.