We describe an approach for mobile robot localization based on geometric features extracted from ultrasonic data. As is well known, a single sonar measurement using a standard POLAROIDTM sensor, though yielding relatively accurate information regarding the range of a reflective surface patch, provides scant information about the location in azimuth or elevation of that patch. This lack of sufficiently precise localization of the reflective patch hampers any attempt at data association, clustering of multiple measurements or subsequent classification and inference. In previous work [15, 16] we proposed a multi-stage approach to clustering which aggregates sonic data accumulated from arbitrary transducer locations in a sequential fashion. It is computationally tractable and efficient despite the inherent exponential nature of clustering, and is robust in the face of noise in the measurements. It therefore lends itself to applications where the transducers are fixed relative to the mobile platform, where remaining stationary during a scan is both impractical and infeasible, and where deadreckoning errors can be substantial. In the current work we apply this feature extraction algorithm to the problem of localization in a partially known environment. Feature-based localization boasts advantages in robustness and speed over several other approaches. We limit the set of extracted features to planar surfaces. We describe an approach for establishing correspondences between extracted and map features. Once such correspondences have been established, a least squares approach to mobile robot pose estimation is delineated. It is shown that once correspondence has been found, the pose estimation may be performed in time linear in the number of extracted features. The decoupling of the correspondence matching and estimation stages is shown to offer advantages in speed and precision. Since the clustering algorithm aggregates sonic data accumulated from arbitrary transducer locations, there are no constraints on the trajectory to be followed for localization except that sufficiently large portions of features be ensonified to allow clustering. Preliminary experiments indicate the usefulness of the approach, especially for accurate estimation of orientation.