The handheld F3 metal detector, developed by the MineLab Corporation, measures the responses of buried objects to electromagnetic pulses. These responses can be processed to determine whether a landmine is present. The simplest processor calculates the total energy in the response, thereby reducing the entire spatial and temporal response to a single value. This value, proportional to the amount of metal in the object, can then be compared to a predetermined threshold. The drawback of this common approach is that, although the threshold may be set so that few, if any, mines are missed, doing so may result in a high false alarm rate. Previous work has demonstrated that incorporating physics-based features into a Bayesian detection framework and performing simple, one-dimensional regional processing can significantly reduce the false alarm rate while maintaining the desired level of detection. Based on these promising results, this approach has been extended to incorporate two-dimensional regional processing. At the test site, data was collected both manually and robotically using nearly identical protocols. Thus, in theory, measured responses should be similar and algorithm performance equivalent whether the detector was operated by a robot or a human. The robustness of various algorithms was evaluated by comparing performance across manual and robotic data sets. Certain physics-based feature detectors were relatively unaffected by the response variability introduced unintentionally by the human operator. However, other algorithms that incorporate more sensitive, often regional, features were able to provide greater gains for the robotic data set than for the manual data set. These results imply that there may be a tradeoff between performance and practical issues that need to be addressed when selecting an algorithm for implementation in a field setting.