Traditional robot planning relies on human-crafted logic representations, but this paper introduces a method to autonomously learn abstract representations from raw robot data. Results show these learned models enable scalable planning for complex tasks without human intervention.