What can cognitive architectures do for robotics?

Abstract

Abstract Research in robotic systems has traditionally been identified with approaches that are characterized by the use of carefully crafted representations and processes to find optimal solutions. The use of such representations and processes, which we refer to as the algorithmic approach, is uniquely suited for problems requiring strong models, i.e., tasks and domains that are well defined, and/or involve close interaction with the environment. These problems have historically been the focus of robotics research because they exercise perceptual, motor and manipulation capabilities that form the basic foundational abilities required for every robotic agent. Recent work (for example ROS and Tekkotsu) on the abstraction and encapsulation of perception and motor functionality has standardized the above mentioned foundational abilities and allowed researchers to study problems in less clearly defined and open-ended domains: problems that have previously been considered the province of AI and Cognitive Science. In this paper, we argue that the study of these problems (examples of which include multi-agent interaction, instruction following and reasoning in complex domains) referred to under the rubric of Cognitive Robotics is best achieved via the use of cognitive architectures – unified computational frameworks developed specifically for general problem solving and human cognitive modeling. We lay out the relevant architectural concepts and principles and illustrate them using nine cognitive architectures that are under active development – Soar, ACT-R, CLARION, GMU-BICA, Polyscheme, Co-JACK, ADAPT, ACT-R/E, and SS-RICS.