Pierre Sermanet

Title: Self-Supervision and Play

Abstract: Real-world robotics is too complex to supervise with labels or through reward functions. While some amount of supervision is necessary, a more scalable approach instead is to bootstrap learning through self-supervision by first learning general task-agnostic representations. Specifically, we argue that we should learn from large amounts of unlabeled play data. Play serves as a way to explore and learn the breadth of what is possible in an undirected way. This strategy is widely used in nature to prepare oneself to achieve future tasks without knowing in advance which ones. In this talk, we present methods for learning vision and control representations entirely from unlabeled sequences. We demonstrate these representations self-arrange semantically and functionally and can be used for downstream tasks, without ever using labels or rewards.