A foundation of the developmental approach to robotics is that learning must be grounded on sensorimotor interaction. In order to behave autonomously, a robot has to build its own model of the world by searching and exploiting statistical regularities in his sensorimotor domain. Self-supervised learning consists in relying on previous knowledge to acquire new skills. We propose to mix self-supervised learning with our probabilistic programming method, the Bayesian Robot Programming Framework. This idea corresponds to achieve feature selection for searching for relevant sensors. We compare several feature selection algorithms and validate them on a real robotic experiment