Task re-encoding in vision-based control systems

Feedback control systems employing video cameras as sensors have been studied in the robotics community for many years. An especially interesting feature of such systems is that both the process output {e.g., the position of the robot in its workspace} and the reference set-point {e.g., visually determined target} are typically observed through the same sensors {i.e., cameras}. Because of this unusual architectural feature, it is sometimes possible to achieve precise positioning {in the absence of measurement noise}, despite sensor/actuator and process model imprecision, just as it is in the case of a conventional set-point control system with a perfect loop-integrator and precise output and exogenous reference sensing. But in contrast to a set-point control system where what to choose for an error is usually clear, in vision-based systems there are many choices for errors, each with different attributes. The aim of this paper is to discuss these issues in a fairly general setting and to provide concrete examples to illustrate the concepts involved in geometrical terms.