This paper offers a new view of stimuli in robot control. The robot is viewed as controlled by a set of processes, only some of which are active at a given time. The control model has three parts: a model for stimuli as perceptual functions, a score matrix that is used to decide on the response to a stimulus, and an assessment of the cost involved in switching the current process upon which a choice of action is based.

The stimulus model requires that the robot be trained to recognize when an input constitutes a stimulus--examples are loss of Wi-Fi, battery exhaustion, and sounds. The training described by the authors identifies regions in the input space that should be ignored as not rising to the level of a stimulus.

The next part uses a stimulus response matrix, which is partially filled by having the robot perform a set of short-term missions under the direction of “operators who are well informed about the task library, the processes, and the stimuli that the robot can handle.” However, since this does not provide values for all the matrix entries, the remaining values “are recovered by factorizing [the matrix] into the product of two smaller matrices.”

The decision to switch or not--for example, to reestablish a Wi-Fi connection--is a matter of “how many processes need to be interrupted, and how many new processes need to be activated.” To do this, tasks and processes are modeled using situation calculus. This cost analysis also takes into account what the authors call “time to failure,” which estimates whether the effort involved in computing the state updates is affordable.

The paper describes a number of experiments in detail. The task-switching model is compared with robot replanning models, and the experiments show improved performance in terms of time required to complete the mission. An appendix goes into greater detail on the implementation, providing tables that illuminate the preceding material--it is definitely not the kind of appendix that should be skipped. The paper is highly recommended to people interested in robot control.