Explore innovative approaches to robot learning through human-derived abstractions in this insightful talk by Andi Peng. Delve into the process of creating representations that capture key task features for decision-making in robotics. Discover three methods for integrating human knowledge into abstraction learning: utilizing human feedback as a general prior for state abstractions in imitation learning, as a personalized interface for identifying implicit preferences, and as a pragmatic framework for learning user-aligned reward functions. Gain valuable insights into improving efficiency, generalizability, and interpretability in robot learning algorithms, drawing from Peng's research at MIT CSAIL and her background in AI safety and governance.