Artificial 3D vision is computationally intensive. It takes an impractically long time for a robot to analyze a video frame in order to accurately estimate the locations of nearby obstacles. This project will develop new techniques for selecting only the most important regions of each video frame to analyze at each moment, so that a robot can update its knowledge of obstacle locations a number of times per second. CrossWing Inc. is developing a telepresence robot that needs this capability in order to support rapid semi-autonomous navigation.