A major contributor to path planning failure is costmap and world model noise and errors, which result when the characteristics of the current environment deviate from the data used to train the sensor perception system. Our adaptive traversability capability improves costmap estimation from purely passive sensing using an online learning strategy that uses successful traversal by the vehicle as a training signal to improve estimation over time. This adaptation mitigates the negative effects of noise and errors in generation of the world model and costmap and decreases dependency and sensitivity to fixed or user-set parameter values. Our system learns the association between traversability and a diverse set of descriptors that characterize the world model and costmap. These descriptors include single voxel-based quantities such as color and temporal coherence and collections of voxels over small local areas such as ground smoothness. By measuring or learning the data characteristics in the fixed sensor perception system and resulting world model, principled estimation of the correctness of the world model and costmap can also be incorporated. We avoid duplicating sensor processing directly in favor of characterizing the derived world model data. This allows the system to directly mitigate the noise and errors that can negatively impact path planning. The system design is general and modular, allowing easy exploration of different learning approaches. Both linear and non-linear methods have been examined, with non-linear methods showing more stable and better performance. We show successful terrain traversability estimation in a diverse set of environments, including unimproved road and expeditionary.