Detecting object boundaries using focus, zoom, shadows, and active probing

Our goal is to give our robot the perceptual abilities it needs to actively resolve ambiguity in its environment, when the passive vision algorithms it uses fail. A key example of this is figure/ground separation [2]. By equipping our robot with a camera with a narrow, controllable depth of field, the robot can bring a questionable edge into focus and blur out distracting background. By giving it the ability to zoom, the robot can look closely around the edge for hints of texture (see Figure 1). And finally, by bringing its manipulator into view, it can physically probe regions and cast its own controlled shadows.