Learning to look

How can autonomous agents with access to only their own sensory-motor experiences learn to look at visual targets? We explore this seemingly simple question, and find that naïve approaches are surprisingly brittle. Digging deeper, we show that learning to look at visual targets contains a deep, rich problem structure, relating sensory experience, motor experience, and development. By capturing this problem structure in a generative model, we show how a Bayesian observer should trade off different sources of uncertainty in order to discover how their sensors and actuators relate. We implement our approach on two very different robots, and show that both of them can quickly learn reliable intentional looking behavior without access to anything beyond their own experiences.