A large-scale navigation system for autonomous off-road robots is presented which uses a topological map to navigate to a previously unseen target location. During path traversal, the system relies on a local navigation layer to avoid obstacles not modeled in the map. Data from this local layer is abstracted to learn both realistic topological edge cost measures and local traversability maps which allow more efficient route selection during topological exploration. On the topological level, a technique to handle impassable route segments in the map is presented and an exploration strategy that allows to discover new routes to the goal is introduced. The performance of the proposed concept is experimentally validated on the robot RAVON at the 2nd Military European Land Robot Trial 2008.