Deep learning has driven substantial progress in robotics, with new ideas, models, and datasets published on a weekly (if not daily!) basis. This pace has led to impressive results but has made it hard for researchers to build a foundation for understanding and contextualizing the impact of new work. This graduate seminar aims to help students entering robot learning build that foundation, through structured paper presentations across four arcs spanning the lifecycle of a robot learning system: 1) How data is collected, and how actions and tasks are represented; 2) What changes once a policy has to make contact with the physical world; 3) What scale, evaluation, and prediction tell us about progress; and 4) How systems fail, how trust is calibrated, and new paradigms for human-robot collaboration.