In this paper, we use deep learning to identify critical regions and automatically construct hierarchical state and action abstractions. We use these hierarchical abstractions with a multi-source mutli-directional hierarchical planner to compute solutions for robot planning problem.