Learning end-to-end policy for navigation with a focus on intelligent exploration has been found to be a challenging task in embodied AI. While methods like soft-Q learning and ensembles of policies have demonstrated navigation behaviors in completely observed maps, we currently do not have ways of extending these policies to unexplored or partially explored environments. To this end, we propose a modular hierarchical formulation by decomposing the navigation task into two sub-problems: selecting the next best goal in the visible space, followed by efficiently navigating to this space in the partial map setting. When navigating in unknown environments, humans often invoke a decision criteria. This could be the next sub-goal, and the decision could either be based on some semantic context, or structural priors. Learning how to select the next sub-goal for higher level task such as coverage of the map
is a challenging open question. We consider frontiers, which are the points at the boundary of known regions of a map, as points to facilitate exploration in a map. Therefore, we define such frontiers as sub-goals, and learn a policy that can select such frontiers on partial grid maps.