Abstract

We present a method for learning active vision skills, to move the camera to observe a robot's sensors from informative points of view, without external rewards or labels. We do this by jointly training a visual predictor network, which predicts future returns of the sensors using pixels, and a camera control agent, which we reward using the negative error of the predictor. The agent thus moves the camera to points of view that are most predictive for a chosen sensor, which we select using a conditioning input to the agent. We observe that despite this noisy learned reward function, the learned policies a exhibit competence by reliably framing the sensor in a specific location in the view, an emergent location which we call a behavioral fovea. We find that replacing the conventional camera with a foveal camera further increases the policies' precision.