Abstract
with its application to the control of a simulated robot, serves to demonstrate that (i) the simulation hypoth-
esis can be elegantly reconciled with global workspace theory and (ii) a robot controller which draws on these
E-mail address: m.shanahan@imperial.ac.uk.
Consciousness and Cognition 15 (2006) 433–449
Consciousness
and
Cognition
www.elsevier.com/locate/concog1053-8100/$ - see front matter  2005 Elsevier Inc. All rights reserved.Cotterill (1998, 2001) advances the proposal that thought is ‘‘internally simulated interaction with the envi-
ronment,’’ and Hesslow (2002) argues that this ‘‘simulation hypothesis’’ can explain our experience of an inner
world. However, while the simulation hypothesis has the potential to account for the content of conscious
thought, it does not supply an answer to the question of what it is that distinguishes conscious from non-con-
scious activity in the brain. By contrast, global workspace theory can account for this distinction by appealing
to an information processing architecture that features both competition among, and broadcast to, different
brain processes (Baars, 1988, 1997).
The present article effects a marriage between these two proposals by presenting a neural-level cognitive
architecture that realises an internal sensorimotor loop in which information passes through multiple compet-
ing cortical areas and a global workspace. This architecture, whose implementation is described here alongThis paper proposes a brain-inspired cognitive architecture that incorporates approximations to the concepts of con-
sciousness, imagination, and emotion. To emulate the empirically established cognitive efficacy of conscious as opposed
to non-conscious information processing in the mammalian brain, the architecture adopts a model of information flow
from global workspace theory. Cognitive functions such as anticipation and planning are realised through internal simu-
lation of interaction with the environment. Action selection, in both actual and internally simulated interaction with the
environment, is mediated by affect. An implementation of the architecture is described which is based on weightless neu-
rons and is used to control a simulated robot.
 2005 Elsevier Inc. All rights reserved.
Keywords: Global workspace theory; Simulation hypothesis; Models of consciousness
1. IntroductionA cognitive architecture that combines internal simulation
with a global workspace
Murray Shanahan
Department of Computing, Imperial College London, 180 Queen’s Gate, London SW7 2AZ, UK
Received 26 April 2005
Available online 27 December 2005doi:10.1016/j.concog.2005.11.005

contemporary ideas from the scientific study of consciousness is also viable from an engineering point of
view.
1
For much of its history, mainstream cognitive science assumed language and reason to be the right concep-
tual foundations on which to build a scientific understanding of cognition. By contrast, the brain-inspired
architecture described here, instead of manipulating declarative, language-like representations in the manner
434 M. Shanahan / Consciousness and Cognition 15 (2006) 433–449of classical AI and cognitive science, realises cognitive function through topographically organised maps of
neurons, which can be thought of as a form of analogical (or diagrammatic or iconic) representation whose
structure is close to that of the sensory input of the robot whose actions they mediate (Barsalou, 1999; Glas-
gow, Narayanan, & Chandrasekaran, 1995; Sloman, 1971).
Analogical representations are especially advantageous in the context of spatial cognition, which, though
not the focus of the present paper, is a crucial capacity for both animals and robots. While common sense
inferences about shape and space are notoriously difficult with traditional logic-based approaches (Shanahan,
2004), in an analogical representation basic spatial properties such as distance, size, shape, and location are
inherent in the medium itself and require negligible computation to extract. Furthermore, traditional lan-
guage-like representations bear a subtle and contentious relationship to the world they are supposed to rep-
resent, and raise difficult questions about intentionality and symbol grounding (Harnad, 1990; Shanahan,
2005a). With analogical representations, which closely resemble raw sensory input, this semantic gap is small
and these questions are more easily answered.
In addition to these representational considerations, the design of the proposed architecture reflects the
view, common among proponents of connectionism, that parallel computation should be embraced as a foun-
dational concept rather than sidelined as a mere implementation issue. Specifically, the present paper advo-
cates a computational architecture based on the global workspace model of information flow, in which a
serial procession of states emerges from the interaction of many separate, parallel processes (Baars, 1988,
2002). This serial procession of states, which includes the unfolding of conscious content in human working
memory (Baars & Franklin, 2003), facilitates anticipation and planning and enables a cognitively enhanced
form of action selection. Yet the robustness and flexibility of these cognitive functions depends on the
behind-the-scenes performance of extremely large numbers of parallel computations, only the most relevant
of which end up making a contribution to the ongoing serial thread (Shanahan & Baars, 2005).
The architecture presented here is intended to be neurologically plausible at the level of large-scale neural
assemblies, and contains analogues of a variety of brain structures and systems, including multiple motor-cor-
tical populations (that compete for access to the global workspace), internal sensorimotor loops (capable of
rehearsing trajectories through sensorimotor space), the basal ganglia (to carry out action selection), and
the amygdala (to guide action selection through affect). But the central component is the global workspace
itself, for which there are a number of candidate homologues in the vertebrate brain, including higher-order
thalamocortical relays, and long-range corticocortical fibres.
In its overall conception, the architecture appeals to the notions of imagination and emotion as well as con-
sciousness. Although perhaps only rough approximations to their humanly applicable counterparts, the way
these concepts are deployed here is inspired by their increasingly important role in the brain sciences (Dama-
sio, 2000). As such, the architecture described builds on the work of a number of other authors who have
applied these ideas in the context of robotics or artificial intelligence.
• Consciousness. As already touched on, global workspace theory proposes a model of information flow in
which conscious information processing is cognitively efficacious because it integrates the results of the
brain’s massively parallel computational resources (Baars, 1988, 2002). The global workspace architecture
has previously been used in the design of software agents (Franklin, 2003; Franklin & Graesser, 1999), but
its application to robotics has so far been neglected in the literature.
1
The focus of the paper is the intersection of global workspace theory and the simulation hypothesis. It makes no claim to implement
every aspect of the multi-faceted theory presented in Baars’s, 1988 book. In particular, the global workspace architecture is not heredeployed to assemble an effective reactive response to a novel situation, but only in the service of deliberation.