Tractability and Attention: Key Roles in Robotic Visual Search

Abstract

Visual search for objects, locations, or events of interest is a central capability for a robot with real-world utility. This capability cannot be limited to yes-no detection; it must include an ability to measure, describe, and compare within the context of a task. We have been investigating this problem since the late 1980s and regardless of the prevailing trends in machine vision, learning, or robotics, have not found reason to ignore the roles of attention, nor a deep understanding of the computational nature of the problem. This presentation will briefly trace our journey. Along the way, we emphasize a number of major points, including the roots of our approach in issues of tractability, the design and evaluation of our subsumptive search algorithm, the development of the AIM saliency model, the confounding nature of sensor bias, the integration of saliency within the object search algorithm, and the need for an overarching framework for attentive behavior, which we have named “Cognitive Programs.”