This paper describes a vision-based landmark learning and recognition system for use in mobile robot navigation tasks. The system uses genetic algorithms (GA) for both learning and recognition processes. It is able to learn new landmarks with very little human intervention. The recognition system can read text inside landmarks, when present. This learning ability has been tested with two very different landmarks that have been successfully used for indoor topological robot navigation. In addition, some new landmarks are learnt, that will be tested for indoor-outdoor navigation in future works. The presented experimental results show the performances of the proposed algorithm.