|[hawesetal06cosy] Nick Hawes, Jeremy Wyatt and Aaron Sloman. An Architecture Schema for Embodied Cognitive Systems. Technical report CSR-06-12, University of Birmingham, School of Computer Science. Submitted for project review as part of CoSy deliverable DR.1.2. November 2006. [pdf] [bib]|
The study of architectures to support intelligent behaviour is certainly the broadest, and arguably one of the most ill-defined enterprises in AI and Cognitive Science. In the CoSy project one of our goals is to develop and understand cognitive architectures suitable for the control of robots. This is not the same as developing an architecture for robot control, nor is it the same as developing a purely cognitive architecture unconnected to real sensors or actuators. We argue that work on architectures traditionally falls into two camps. First there are cognitive architectures which attempt to provide unified theories of cognition such as SOAR [Laird et al., 1987] and ACT-R [Anderson et al., 2004]. Their value is typically measured in terms of an ability to reproduce some of the characteristics of human like information processing. ACT-R for example is used extensively for creating models of cognition that are then evaluated against data from humans. In other words, they are evaluated as psychological theories. On the other hand, roboticists have been engaged with issues of how to enable robots to act reliably and robustly in a rapidly changing world when faced with limited computational power, uncertain sensing, and uncertain action. Architectures for robot control such as 3T [Bonasso et al., 1997] are therefore largely concerned issues of real- time control, uncertainty, sensory fusion (or the lack of it), and error recovery. They are evaluated in terms of the performance of the resulting robotic systems on a variety of tasks.
In CoSy we have interests in both cognitive science and engineering science, and consequently our work is related to both of the aforementioned camps whilst also looking at closely related issues. Our work is neither concerned with trying to model humans or any other specific type of animal, nor with trying to compete on practical design tasks. Rather it is concerned with trying to understand the possibilities and trade-offs involved in different designs in relation to different sets of requirements. In this paper we will describe an architecture schema which inherits some of the ambitions of classic cognitive architectures, and those of robot control architectures, whilst allowing us to explore these additional issues. It is important to note now that we don't present an empirical evaluation of an implementation of a scenario-specific instantiations of the architecture schema in this paper, we do describe two possible instantiations based on the CoSy demonstrator scenarios. It is also worth stating at this stage that our intention is not to perform extensive evaluation of the architecture schema against human behaviour. Instead we intend to evaluate it by profiling behaviour of scenario-specific tiations of it under varying conditions, such as internal failures and varying types of change in the world.
|Download: pdf (319 KB) bib|
|Links: [Google Scholar] [CiteSeer]|
|Available from: ftp://ftp.cs.bham.ac.uk/pub/tech-reports/2006/CSR-06-12.pdf|