A framework for vision-guided manipulation of a moving target

The authors present a general framework for reasoning about robotic manipulation tasks involving a moving target, where manipulation is loosely defined to include catching, hitting, interception, etc. This framework may be used to achieve robust vision-based control. The different levels at which visual input is involved in pursuing the dynamically defined goal are considered. A given task is first transformed into one of constrained trajectory planning on a topological space defined by a set of image parameters. A learning phase first learns the qualitative features of this perceptual control surface so that further operations may be carried out autonomously without precise calibration of different parts of the system. The framework facilitates the incorporation of learning strategies to automate the control mechanism.<<ETX>>