Long-living autonomous agents must be able to learn to perform competently in novel environments. One important aspect of competence is the ability to plan, which entails the ability to learn models of the agent’s own actions and their effects on the environment. In this paper we describe an approach to learn action models of environments with continuous-valued spatial states and realistic physics consisting of multiple interacting rigid objects. In such environments, we hypothesize that objects exhibit multiple qualitatively distinct behaviors we call modes, conditioned on their spatial relationships to each other. We argue that action models that explicitly represent these modes using a combination of symbolic spatial relationships and continuous metric information learn faster, generalize better, and make more accurate predictions than models that only use metric information. We present a method to learn action models with piecewise linear modes conditioned on a combination of first order Horn clauses that test symbolic spatial predicates and continuous classifiers. We empirically demonstrate that our method learns more accurate and more general models of a physics simulation than a method that learns a single function (locally weighted regression).