Learning Visual Affordances for Robotic Manipulation