Learning and Reasoning with Action-Related Places for Robust Mobile Manipulation