Autonomous robot navigation in unstructured outdoor environments is a challenging area of active research and is currently unsolved. The navigation task requires identifying safe, traversable paths that allow the robot to progress toward a goal while avoiding obstacles. Stereo is an effective tool in the near field, but used alone leads to a common failure mode in autonomous navigation in which suboptimal trajectories are followed due to nearsightedness, or the robot's inability to distinguish obstacles and safe terrain in the far field. This can be addressed through the use of machine learning methods to accomplish near-to-far learning, in which near-field terrain appearance and stereo readings are used to train models able to predict far-field terrain. This paper proposes to enhance existing, memoryless near-to-far learning approaches through the use of classifier ensembles that allow terrain models trained on data seen at different points in time to be preserved and referenced later. These stored models serve as memory, and we show that they can be leveraged for more effective far-field terrain classification on future images seen by the robot. A five-factor, full-factorial, repeated-measures experimental evaluation is performed on hand-labeled data sets taken directly from the problem domain. The experiments result in many statistically significant findings, the most important being that the proposed near-to-far Best-K Ensemble Algorithm, with appropriate parameter selection, outperforms the single-model, nonensemble baseline approach in far-field terrain classification. Several other findings that inform the use of near-to-far ensemble methods are also presented. © 2009 Wiley Periodicals, Inc.