Learning from Human Play Data

Human play data consists of unstructured, reset-free interactions with the environment, where the human is able to choose both the tasks and how the tasks are executed. Suneel Belkhale with the Stanford Intelligent and Interactive Autonomous Systems Group (ILIAD) discusses methods for learning from human play data and presents the lab's recent method - PLATO: Predicting Latent Affordances Through Object-centric Play - which views play data as sequential object interactions in order to learn a robust policy from play.