Leveraging Qualitative Reasoning to Learning Manipulation Tasks

Abstract

Learning and planning are powerful AI 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 describe how both methods can be combined using an expressive qualitative knowledge representation. We argue that the crucial step in this integration is to employ a representation based on a well-defined semantics. This article proposes the qualitative spatial logic QSL, a representation that combines qualitative abstraction with linear temporal logic, allowing us to represent relevant information about the learning task, possible actions, and their consequences. Doing so, we empower reasoning processes to enhance learning performance beyond the positive effects of learning in abstract state spaces. Proof-of-concept experiments in two simulation environments show that this approach can help to improve learning-based robotics by quicker convergence and leads to more reliable action planning.