Clark Embodiment in cognitive science Opinion

An embodied cognitive
science?
Andy Clark

The last ten years have seen an increasing interest, within cognitive science, in issues
concerning the physical body, the local environment, and the complex interplay between
neural systems and the wider world in which they function. Yet many unanswered
questions remain, and the shape of a genuinely physically embodied, environmentally
embedded science of the mind is still unclear. In this article I will raise a number of
critical questions concerning the nature and scope of this approach, drawing a distinction
between two kinds of appeal to embodiment: (1) Simple cases, in which bodily and
environmental properties merely constrain accounts that retain the focus on inner
organization and processing; and (2) More radical appeals, in which attention to bodily
and environmental features is meant to transform both the subject matter and the
theoretical framework of cognitive science.

T alk of embodiment and situatedness has become increas-

ingly frequent in philosophy15, psychology6,7, neuroscience8,9,
cylinder with a kind of foot. To get the hopper to locomote
to move, balance and turn involved solving a control prob-
robotics10,11, education1214, cognitive anthropology15,16, lin- lem that was radically impacted by the mechanical details,
guistics4,17, and in dynamical systems approaches to behavior such as the elastic rebound that occurs when the leg hits the
and thought18,19. There is clearly a shift in thinking but the floor. The crucial control parameters included items such as
nature and importance of the shift is surprisingly hard to pin the resting length of the leg spring, and degree of sideways tilt.
down. What is embodied cognitive science and how far can To understand how the robots brain controls the robots
it take us? Let us look first at some of the varied roles that motions, a shift towards an embodied perspective is required.
embodiment can play, as illustrated in the following examples. The controller must learn to exploit the rich intrinsic dynam-
ics of the system. As Fred Keijzer recently put it, instead
Fish of thinking about [the] control system as a center for com-
Consider first the swimming ability of the Bluefin tuna. The mands to be executed by actuators, the body and its move-
Bluefin tuna is a swimming prodigy, but its aquatic capabilities ments are taken as a system with its own dynamic charac-
its ability to turn sharply, to accelerate quickly, and to reach teristics (F. Keijzer, doctoral dissertation, University of
such high speeds have long puzzled biologists. Physically Leiden, 1998). A similar shift in thinking can be applied to
speaking, so it seemed, the fish should be too weak (by about a action routines in human infants and adults (see Box 1).
factor of seven) to achieve these feats. However, an explanation
for this prodigious ability can be found in the use of embodied, Vision
environmentally embedded action by the tuna. Fluid dy- A further example can be found in research in animate or
namicists have suggested that the fish uses bodily action to interactive vision22,23. The key insight here is that the task of
manipulate and exploit the local environment (the water) so vision is not to build rich inner models of a surrounding
as to swim faster, accelerate more quickly, and so on20. It ap- 3-D reality24, but rather to use visual information efficiently
pears that the tuna find and exploit naturally occurring cur- and cheaply in the service of real-world, real-time action.
rents so as to gain speed, and use tail flaps to create addi- Researchers in animate and interactive vision thus reject
tional vortices and pressure gradients, which are then used what Churchland et al.23 dub the paradigm of pure vision A. Clark is at the
Philosophy/
for rapid acceleration and turning. The physical system the idea (associated with work in classical AI and in the
Neuroscience/
whose functioning explains the prodigious swimming ca- use of vision for planing) that vision is largely a means of
Psychology (PNP)
pacities of the Bluefin tuna is thus the fish-as-embedded-in, creating a world model rich enough to let us throw the
Program, Washington
and as actively exploiting, its local environment. world away, allowing reason and thought to be focused University in St Louis,
upon the inner model instead. Real-world action, in these USA.
Robots pure vision paradigms, functions merely as a means of im-
Next consider a hopping robot, designed and built by Raibert plementing solutions arrived at by pure cognition. The ani- tel: +1 314 935 7146
fax: +1 314 935 7349
and Hodgins in 1993 (Ref. 21). Their robots were designed mate vision paradigm, by contrast, gives action a primary e-mail: andy@
to balance and move by hopping on a single leg a pneumatic role22. Computational economy and temporal efficiency are twinearth.wustl.edu

1364-6613/99/$ see front matter 1999 Elsevier Science Ltd. All rights reserved. PII: S1364-6613(99)01361-3
345
Trends in Cognitive Sciences Vol. 3, No. 9, September 1999
Opinion Clark Embodiment in cognitive science

Box 1. The situated infant

Thelen and Smith have shown in some detail that infant walking fully back. Flat-foot contact with the moving treadmill creates
abilities depend on complex interactions between neural states, this condition and initiates stepping, whereas toe contact yields
the biomechanics of the legs, and local environmental parameters insufficient back stretch. Infant stepping is thus a complex,
(Ref. a). This view is then contrasted with the vision of walking multi-factor affair in which the target behavior emerges only
as the simple expression of a temporally staged set of genetically when the central elements cooperate with the effectors the
specified instructions (Ref. a, pp. 820, 263266). Experimental muscles, joints, tendons in the appropriate physical context
support for the multi-factor view includes a demonstration of (Ref. a, p. 113).
induced stepping in some (non-stepping) seven-month-olds Scott Kelso has demonstrated similar multi-factor profiles in
held upright upon a motorized treadmill. In this condition, the adult motor routines such as the production of rhythmic finger
non-stepping infants could even compensate for twin belts, each motions (Ref. b.) For further commentary on both of these cases,
driving a leg at a different speed! The treadmill stepping, it was see Clark (Ref. c.)
further discovered, depended crucially on the type of contact
made between foot and belt: flat-foot contact induced stepping, References

whereas mere toe-contact failed. The explanation seems to be a Thelen, E. and Smith, L. (1994) A Dynamic Systems Approach to the
Development of Cognition and Action, MIT Press
that the stepping behavior depends heavily on a spring-like
b Kelso, S. (1995) Dynamic Patterns, MIT Press
biomechanical response (Ref. a, pp. 111112). To uncoil the
c Clark, A. (1997) The dynamical challenge Cognit. Sci. 21, 461481
spring and propel the leg forward, it must first be stretched

purchased by a variety of tricks and ploys that exploit bodily image in the optic array) accurately predicts time-to-impact
action and local environment, including: and can be used to time wing closure. Other behaviors, such
(1) the use of cheap, easy-to-detect environmental cues as the timing of an athletes jump or a stroke in tennis, make
(e.g. searching for Kodak film in a drug store? Seek Kodak use of a similar quantity (see Ref. 32 for a review).
yellow); A similar approach can help explain how an outfielder
(2) the use of active sensing (using motor action, guided in baseball positions him/herself to catch a fly ball. It used to
by rough perceptual analysis, to seek further inputs yielding be thought that this problem required complex calculations
better perceptual data. Move head and eyes for better depth of the arc, acceleration and distance of the ball. More recent
perception, etc.); work, however, suggests a computationally simpler strategy33.
(3) the use of repeated consultations of the outside world Put simply, the fielder continually adjusts his or her run so
in place of rich, detailed inner models. that the ball never seems to curve towards the ground, but
This approach suggests that vision is a highly active and instead appears to move in a straight line in his or her visual
intelligent process. It is not the passive creation of a rich inner field. By maintaining this strategy, the fielder should be guar-
model, but rather the active retrieval (typically by moving the anteed to arrive in the right place at the right time to catch
high resolution fovea to sequential locations in the visual the ball.
scene) of useful information as it is needed from the constantly Notice the difference between these two models. In the
present real-world scene. Ballard et al.25 speak of just-in-time traditional model, the brain takes in data, performs a com-
representation, while the roboticist, Rodney Brooks, has plex computation that solves the problem (where will the ball
coined the slogan The world is its own best model11. Bio- land?) and then instructs the body where to go. This is a linear
logical vision thus gears its computational activity closely processing cycle: perceive, compute and act. In the second
and sparingly to the task at hand, making the most efficient model, the problem is not solved ahead of time. Instead, the
use of the persisting external scene. task is to maintain, by multiple, real-time adjustments to the
run, a kind of co-ordination between the inner and the outer
Action and affordance worlds. Such co-ordination dynamics constitute something of
Related insights stem from the work of J.J. Gibson26,27 and the a challenge to traditional ideas about perception and action:
ecological psychology movement7,28,29. This approach stresses they replace the notion of rich internal representations and
bodily movement, ecological context and the action-relevant computations, with the notion of less expensive strategies
information available in the perceptual array. A central orga- whose task is not first to represent the world and then rea-
nizing construct is the concept of an affordance26. Affordances son on the basis of the representation, but instead to maintain
are the possibilities for use, intervention and action which the a kind of adaptively potent equilibrium that couples the agent
physical world offers a given agent and are determined by the and the world together. Whether such strategies are genuinely
fit between the agents physical structure, capacities and skills non-representational and non-computational, or suggestive
and the action-related properties of the environment itself30. of different kinds of representation (action-oriented represent-
A simple but illustrative example is Lee and Reddishs ac- ations) and more efficient forms of computation, is a difficult
count of how diving birds, such as plovers and gannets, are able question whose resolution remains uncertain2,6,7,18,3436.
to close their wings at exactly the right moment before hitting What is clear, however, is that tuning to higher-order in-
the surface of the water in pursuit of a fish31. Such behavior is variants can help explain a wide variety of adaptive responses,
possible because there is available in the optic array, a higher- including visually guided locomotion37,38, rhythmic move-
order invariant that allows the control of such action. This ment18,39, and the capacity to grasp and wield objects7 (e.g.
quantity (which involves the relative rate of expansion of the hammers, golf clubs). In all these cases behavioral success

346
Trends in Cognitive Sciences Vol. 3, No. 9, September 1999
 Clark Embodiment in cognitive science Opinion

involves locking on to simple (but often far from obvious) mechanisms thus record landmarks as a combination of ro-
properties of the environment made available in the perceptual botic motion and sonar readings, so that a corridor might be
array (see also Box 1). encoded as a combination of forward motion and a sequence
of short, lateral sonar distance readings. The stored map
Beyond adaptive coupling? is thus perfectly formatted to act as a direct controller of
The full implications and significance of these embodied and embodied action: using the map to find a route and gener-
embedded approaches remain to be determined, and there ating a plan of actual robot movements is therefore a single
are a number of difficulties that clearly remain to be resolved. computational task.
An immediate question is, to what extent, if at all, can the em- TOTO is adept at interacting with the local environment,
bodied, embedded approach contribute to our understanding and can even, in a weak sense, track that which is not present-
of so-called representation-hungry problem-solving40? To to-hand: it can return, on command, to a previously encoun-
illustrate this idea, consider the much simpler notion of adap- tered location. TOTO cannot, however, be prompted to track
tive coupling. Adaptive coupling occurs when a system (typi- or think about any location that it has not previously visited.
cally a plant or animal) evolves a mechanism that allows it to METATOTO46 builds on the original TOTO architecture
track the behavior of another system (a predator, or a source to create a system capable of finding its way, on command, to
of food or energy). locations that it has never previously encountered. It does so
To borrow an example from Brian Cantwell Smith, the by using the TOTO architecture off-line, so as to support
sunflower has evolved to track the daily motion of the sun the exploration, in imagination, of a totally virtual environ-
across the sky41. The sunflower thus has states that co-vary ment. When METATOTO is imagining, it deploys exactly
with solar position, and this is what they are meant (evolu- the same machinery that (in TOTO, and in METATOTO
tionarily speaking) to do. But does the sunflower thereby on-line) normally supports physical interactions with the real
exhibit cognition or mentality or intentionality; does it have world. The difference lies solely at the lowest-level interface:
internal representations? A common (and I think correct) where TOTO uses sonar to act and navigate in a real world,
intuition is that it does not. There is nothing cognitive oc- METATOTO uses simulated sonar to explore a virtual world
curring. One reason for thinking this is that cognition has (including a virtual robot body). Stein then goes on to imag-
been taken to involve the capacity to relate to an intentional ine linguistic directions interfacing with this virtual realm by
object42 and this means, in part, an object that might not be translating descriptions such as the second left into TOTO
present-at-hand or that might not even exist. The sunflower, (METATOTO)-style action-based encodings, such as short
by contrast, tracks the sun only when the sun is (in a certain sonar left, long sonar left, short sonar left, long sonar left
sense, at least) actually there. More precisely, the sunflower ex- (Ref. 46, p. 404).
hibits its behavior when there is an ongoing external physical METATOTO uses the basic behavior-producing archi-
trace to which it can adaptively couple. tecture of TOTO, but includes a program that can take, for
The mark of the cognitive, then, is the capacity to engage example, a floor plan or map and use it to stimulate the robots
in something like off-line reason43 reasoning in the absence sensors in the way they would be stimulated if the robot were
of that which our thoughts concern. Classical (disembodied) actually locomoting along a given route on the map. The map
cognitive science accounted well for this, by positing an inner can thus induce sequences of experiences that are qualita-
realm richly populated with internal tokens that stood for ex- tively similar to those generated by real sensing and acting,
ternal objects and states of affairs. Thus, it was able to offer a and this allows METATOTO to profit from virtual experi-
simple account of behavioral co-ordination in the absence of ences, just as TOTO profits from real experience. Once the
any external physical trace or perceptually available higher- sensors and motors are restored to real world input and ac-
order invariant (for some excellent discussion of the issues tion, on-line METATOTO can immediately find its way to
confronting a Gibsonian approach to cognition, see Kirsh44 a target location it has not actually (but merely virtually)
and Van Leeuwen45). visited.
One promising advance is the suggestion that embodied We should now ask two, related questions. How different
cognitive science might treat off-line reason as something is this account from more traditional solutions? And will it
like simulated sensing and acting, thus preserving the special work for all kinds of off-line reasoning or only some? The first
flavor of embodied problem-solving alongside a high degree question, it seems to me, leads to a mild dilemma. For the
of ability to decouple from the environment. The most devel- simulation-based account looks most clearly different (from
oped version of this strategy is probably the mobile robotics traditional accounts involving inner-world models) only in-
work of Lynn Andrea Stein at the MIT Artificial Intelligence sofar as it treats planning as, quite literally, imagined inter-
lab46. Stein uses as her platform a mobile robot (designed and action. Thus, Stein notes that While traditional planners use
implemented by Maja Mataric47) named TOTO. TOTO an abstracted world and plan operators distinct from the ac-
uses ultrasonic range sensors to detect walls, corridors and tual robot controls, our system uses the robotic architecture
other obstacles and is able to use its physical explorations to itself (Ref. 46, p. 396). To support this claim, Stein reminds
build an inner map of its environment, which it can then use us that METATOTO works by simulating both sensors and
to revisit previously encountered locations on command. actuators, and that simulation runs create the kinds of feed-
TOTOs internal map is, however, rather special in that it back (short and long sonar signals, etc.) that would be received
encodes geographic information in an action-oriented way from the actual world, were the robot actually to change pos-
(Ref. 2, p. 49), combining information about the robots ition. There are, of course, some idealizations: the simulated
movement and correlated perceptual input. TOTOs inner motion is, for example, straight and precise, unaffected by

347
Trends in Cognitive Sciences Vol. 3, No. 9, September 1999
Opinion Clark Embodiment in cognitive science

Box 2. Dynamical systems theory

The dynamical systems approach has been gaining ground in cog- recognized by van Gelder and Port (Ref. b) and pursued by
nitive scientific treatments of cognition and adaptive behavior Melanie Mitchell and her colleagues (Ref. g), is that the dynami-
(Refs ad). The dynamical approach focuses on the evolution of cal approaches might transform and enrich (rather than displace)
a system over time, and is particularly well-suited to dealing with our computational and representational understandings, perhaps
cases in which a system or component a is constantly affecting by identifying complex, temporally extended dynamical patterns
and being affected by another system or component b (which (chaotic attractors, limit cycles, values of collective variables, etc.)
might likewise be continuously sensitive to item c and so on). as the vehicles of specific representational contents and as the
An example might be the process of returning a tennis serve: the implementation of so-called emergent computational processes
location of the ball and the other player (and perhaps your partner, (Ref. h).
in doubles) are constantly changing, and simultaneously, you are
References
moving and acting, which is affecting the other players, and so
a Port, R. and van Gelder, T., eds (1995) Mind as Motion: Dynamics,
on. Put simply, everything is simultaneously affecting everything Behavior, and Cognition, MIT Press
else (Ref. a, p. 23). Such densely coupled events, insofar as they b van Gelder, T. and Port, R. (1995) Its about time: an overview of the
prove characteristic of all or some human cognitive activity, dynamical approach to cognition, in Mind as Motion: Explorations in
seem ill-understood in traditional cognitive scientific terms. By the Dynamics of Cognition (Port, R. and van Gelder, T., eds), pp. 144,
contrast, the apparatus of dynamical systems theory, with its MIT Press
notions of state spaces, coupled systems, trajectories in state space, c Kelso, S. (1995) Dynamic Patterns, MIT Press

collective variables and more is expressly designed to deal with d Thelen, E. and Smith, L. (1994) A Dynamic Systems Approach to the
Development of Cognition and Action, MIT Press
such simultaneous interactive complexity (Ref. e). It provides a
e Abraham, R. and Shaw, C. (1992) Dynamics: The Geometry of
set of mathematical and conceptual tools that support a geometric
Behavior, AddisonWesley
understanding of the space of possible total system behaviors.
f Beer, R. (1995) A dynamical systems perspective on environment
Such analyses have proven useful in understanding the activity agent interactions Artif. Intell. 72, 173215
of simple robots (Ref. f), infants (Ref. d), and adults (Ref. c). g Mitchell, M., Crutchfield, J. and Hraber, P. (1994) Evolving cellular
But what mileage can we get from such analyses once we leave automata to perform computations Physica D 75, 361391
the domain of on-the-spot adaptive coupling and turn to various h Crutchfield, J. and Mitchell, M. (1995) The evolution of emergent
forms of off-line reason and cogitation? One exciting possibility, computation Proc. Natl. Acad. Sci. U. S. A. 92, 1074210746

the dinks and slopes found in the real world. But, overall, Simple versus radical embodiment
METATOTO does indeed rely on the simulation of sensori- In addition to asking how far the embodied approach can
motor experience rather than on abstract kinds of reasoning go, we should also ask to what extent it is a genuinely radical
and planning. alternative to more traditional views. To focus this concern,
What remains unclear, however, is the scope of this kind I would like to distinguish two different ways to appeal to
of solution. The reader might, for example, try the following facts about embodiment and environmental embedding. The
exercise in abstract off-line reasoning: consider whether US first, which I will call simple embodiment, treats such facts
gun manufacturers should be held liable for having know- as, primarily, constraints upon a theory of inner organization and
ingly manufactured more guns than the legal market could processing. The second, which I will call radical embodiment
possibly account for? Here it is not clear how rich sensori- goes much further and treats such facts as profoundly altering
motor simulation could possibly account for all the kinds of the subject matter and theoretical framework of cognitive science.
moral and abstract reasoning required reasoning about The distinction between the simple and radical forms is, how-
rights, implications, responsibilities, economics, and so on. ever, not absolute, and many (perhaps most) good research
Nor is this just a point about how things seem introspec- programs end up containing elements of both.
tively. Rather, it is hard to see how sensorimotor simulation Examples of simple embodiment abound in the literature.
could in principle account for all the kinds of thought and A good deal of work in interactive vision, for example, still
reasoning that the problem demands. Simulated acting and relies heavily on internal representations, computational trans-
sensing may well play a role and perhaps even an essential formations, and abstract data structures22,25. There is much
role4 in our reasoning. But the capacity to examine argu- talk for example, of inner databases, of internal featural rep-
ments, to judge what follows from what, and to couch the resentations (of color, shape and so on), of high-dimensional
issues in the highly abstract terms of a fundamental moral feature vectors, and so on. Attention to the roles of body,
debate (using concepts like liability, reasonable expecta- world and action, in such cases, is merely a methodological
tion, acceptable risk, etc.) does not obviously lend itself to tool aimed at getting the internal data-structures and oper-
an analysis in terms of literal sensorimotor simulation as in ations right. Churchland et al.s vision of a motocentric
METATOTO. Perhaps there are other, less direct ways to rather than visuocentric cognitive science has, I suspect, a
depict high-level cognition as dependent on simulated sensing similar goal (Ref. 23, p. 60). Maja Matarics47 work on
and acting48. But (and this is the mild dilemma) it does seem TOTO has this flavor, insofar as it concentrates attention
that the more decoupled and abstract the target contents be- on an inner representational resource (the map/controller)
come, either the less applicable the sensorimotor simulation and is exploring the ways in which usefulness in the guid-
strategy is, or the less clearly it can then be differentiated from ance of real-world action can both constrain and inform the
the more traditional approaches it seeks to displace. nature of inner representations and processing. The same

348
Trends in Cognitive Sciences Vol. 3, No. 9, September 1999
 Clark Embodiment in cognitive science Opinion

applies to the recent important work on the role of bodily a reaching motion to a target. In such cases, Ballard et al.
metaphors4 in abstract, high-level cognition: here, too, the suggest, the external world is analogous to computer memory
goal is to give an account of the inner representational realm, and changing gaze is analogous to changing the memory ref-
but one informed by the evolutionary and developmental erence in a silicon computer (Ref. 28, p. 725). The compu-
roles of bodily experience. tational organization relevant to cognition is here depicted
The source of much recent excitement, however, are the as literally spread across neural, bodily and environmental
striking claims involving radical embodiment. Such claims elements.
can be found in work by Tim Van Gelder49, Thelen and In thinking about higher cognition and advanced human
Smith6, Kelso18, Varela et al.1, Turvey and Carello7 and others. reason, it might likewise prove fruitful to consider the literal
These accounts of radical embodiment all involve one or more extension of the cognitive system to include aspects of the
of the following claims: local environment. In this vein, Clark2,52 and Hutchins15,
(I) that understanding the complex interplay of brain, following Vygotsky53, Bruner54, Dennett55 and others, have
body and world requires new analytic tools and methods, such argued that just as basic forms of real-world success turn on
as those of dynamical systems theory6,18,49 (see Box 2) the interplay between neural, bodily and environmental fac-
(II) that traditional notions of internal representation and tors, so advanced cognition turns in crucial respects upon
computation are inadequate and unnecessary6,7,49 the complex interplay between individual reason, artifact
(III) that the typical decomposition of the cognitive sys- and culture. The simplest illustration of this idea is prob-
tem into a variety of inner neural or functional subsystems ably the use of artifacts such as pen and paper to support or
is often misleading, and blinds us to the possibility of alter- scaffold human performance34,5658. Most of us, armed with
native, and more explanatory, decompositions that cut across pen and paper, can, for example, solve multiplication prob-
the traditional brainbodyworld divisions3,5,6,15. lems that would baffle our unaided brains. In so doing we
Closely related to these three points is the idea that even create external symbols (numerical inscriptions) and use ex-
the subject matter of cognitive science needs to be re-thought. ternal storage and manipulation so as to reduce the complex
A mature science of the mind, it now seems, targets not (or problem to a sequence of simpler, pattern-completing steps
not only) the individual, inner organization of intelligence that we already command58. On this model, then, it is the
but the bodily and environmentally extended organizations combination of our biological computational profile with
responsible for adaptive success2,12,15. the fundamentally different properties of a structured, sym-
Some support for claims (I) and (II) may be found in bolic, external resource that is a key source of our peculiar
the work on infant motor development30,37 (Box 1), adult brand of cognitive success53,55. The external environment,
motor actions7,18, and mobile robotics911,21. The support is actively structured by us, becomes a source of cognition-
weak, however, because the solutions that appear most non- enhancing wideware52 external items (devices, media,
computational, representation-free, and open to dynamical notations) that scaffold and complement (but usually do not
analysis (see Box 2) usually involve cases of adaptive coupling, replicate) biological modes of computation and processing,
and do not directly confront representation-hungry prob- creating extended cognitive systems whose computational
lems. Here, we must simply suspend judgement and await profiles are quite different from those of the isolated brain.
empirical advances. Hutchins, for example, provides a lucid and detailed account
My own guess, however, is that as tasks become more of the way multiple biological brains, tools (such as sextants
representation-hungry more concerned with the distal, ab- and alidades), and media (such as maps and charts) combine
stract and non-existent we will see more and more evidence to make possible the act of ship navigation15. In Hutchins
of some kinds of internal representation and inner models. It words, such tools and media permit the users to do the tasks
is at exactly this point that the possibility of a middle ground that need to be done while doing the kinds of things people
between simple and radical versions of embodiment becomes are good at: recognizing patterns, modeling simple dynamics
apparent. For these new kinds of internal representation might of the world, and manipulating objects in the environment
differ from familiar forms both in their contents (being more (Ref. 15, p. 155).
deictic25 and action-oriented2) and in the nature of their In short, the world of artifacts, texts, media, and even
inner vehicles (perhaps using temporally extended processes cultural practices and institutions59, might be for us what
and complex dynamical regularities as inner tokens2,50,51). the actively created whorls and vortices are for the Bluefin
A good prospect for the supporters of radical embodi- tuna. Human brains, raised in this sea of cultural tools55
ment could, however, lie in claim (III) (that there might be might develop strategies for advanced problem solving that
alternative systemic decompositions). An example which factor in these external resources as profoundly and deeply
also demonstrates how a single research program can com- as the bodily motions of the tuna factor in and maximally
bine elements of simple and radical embodiment is Ballard exploit the reliable properties of the surrounding water.
et al.s25 use of a notion of deictic pointers. A pointer, in arti- Recognizing the complex ways in which human
ficial intelligence, is an inner state, which can act both as an thought and reason exploit the presence of external symbols
object of computation and as a key for retrieving additional and problem-solving resources, and unraveling the ways in
data-structures or information. Deictic pointers, as Ballard which biological brains couple themselves with these very
et al. describe them, are physical actions such as foveating special kinds of ecological objects, is surely one of the most
a certain location in visual space that play a similar kind of exciting tasks confronting the science of embodied cogni-
functional role. The very act of foveation, it is suggested, tion and one that might shed great light on the role of
may be used to temporarily bind color to location, or to direct embodiment in more abstract cognitive domains.

349
Trends in Cognitive Sciences Vol. 3, No. 9, September 1999
Opinion Clark Embodiment in cognitive science

Outstanding questions References

1 Varela, F. et al. (1991) The Embodied Mind, MIT Press
Is cognition truly seamless implying a gentle, incremental trajectory 2 Clark, A. (1997) Being There: Putting Brain, Body and World Together
linking fully embodied responsiveness to abstract thought and off-line Again, MIT Press
reason? Or is it a patchwork quilt, with jumps and discontinuities and with 3 Hurley, S. (1998) Consciousness in Action, Harvard University Press
very different kinds of processing and representation serving different 4 Lakoff, G. and Johnson, M. (1999) Philosophy in the Flesh, MIT Press
needs? 5 Haugeland, J. (1998) Mind embodied and embedded, in Having Thought
What role does public language play in the transition from simple adaptive (Haugeland, J., ed.), pp. 338, MIT Press
coupling to heavy-duty cognition? 6 Thelen, E. and Smith, L. (1994) A Dynamic Systems Approach to the
Insofar as we depend heavily on cultural artifacts (pen, paper, PC) to Development of Cognition and Action, MIT Press
augment and enhance biological cognition, what should we say about 7 Turvey, M. and Carello, C. (1995) Some dynamical themes in perception
their origins? If we do indeed make our world smart so that our brains can and action, in Mind as Motion (Port, R. and Van Gelder, T., eds),
be dumb in peace2, just how did dumb brains create such a smart world? pp. 373401, MIT Press
Are the tools of dynamical systems theory replacements for, or merely 8 Damasio, A. (1994) Descartes Error, Grosset/Macmillan
additions to, the familiar arsenal of inner models, maps, representations 9 Chiel, H. and Beer, R. (1997) The brain has a body: adaptive behavior
and computations? emerges from interactions of nervous system, body and environment
If we follow the embodied, embedded approach to its natural conclusions, Trends Neurosci. 20, 553557
do we lose sight of the differences between perception, reason and action? 10 Arkin, R. (1998) Behavior-Based Robotics, MIT Press
If not, just how do we reconstruct them? Do we begin to lose sight of the 11 Brooks, R. (1991) Intelligence without representation Artif. Intell. 47,
distinction between agents and the worlds in which they think and act? 139159
12 Resnick, M. (1994) Turtles, Termites and Traffic Jams, MIT Press
13 Anderson, J., Reder, L and Simon, H. (1996) Situated learning and
education Educational Researcher 25, 511
Conclusions
14 Greeno, J. (1996) On claims that answer the wrong questions (Response
Embodied, environmentally embedded approaches have a lot
to Anderson et al.) Educational Researcher 25, 511
to offer cognitive science. It is increasingly clear that, in a wide 15 Hutchins, E. (1995) Cognition in the Wild, MIT Press
variety of cases, the individual brain should not be the sole 16 Suchman, A. (1987) Plans and Situated Actions, Cambridge University
locus of cognitive scientific interest. Cognition is not a phe- Press

nomenon that can be successfully studied while marginalizing 17 Regier, T. (1996) The Human Semantic Potential, MIT Press
18 Kelso, S. (1995) Dynamic Patterns, MIT Press
the roles of body, world and action.
19 Port, R. and van Gelder, T., eds (1995) Mind as Motion: Dynamics,
The major challenge for the vision of radical embodiment Behavior, and Cognition, MIT Press
described here lies with the class of representation-hungry 20 Triantafyllou, M. and Triantafyllou, G. (1995) An efficient swimming
problems and the phenomena of off-line, abstract, and en- machine Sci. Am. 272, 6471

vironmentally-decoupled reason. It is important not to con- 21 Raibert, M. and Hodgins, T. (1993) Legged robots, in Biological Neural
Networks in Invertebrate Neuroethology and Robotics (Beer, R.D.,
clude, however, that facts about embodiment impact only our
Ritzmann, R.E. and McKenna, T., eds), pp 319354, Academic Press
ideas about low-level sensorimotor processes. In the human 22 Ballard, D. (1991) Animate vision Artif. Intell. 48, 5786
case, at least, we seem to find at all levels a mixture of highly 23 Churchland, P. et al. (1994) A critique of pure vision, in Large-Scale
embodied, embedded strategies and apparently much more Neuronal Theories of the Brain (Koch, C. and Davis, J., eds), MIT Press

abstract and potentially de-coupled strategies, with the cre- 24 Marr, D. (1982) Vision, W.H. Freeman
25 Ballard, D. et al. (1997) Deictic codes for the embodiment of cognition
ation and manipulation of external symbolic items often func-
Behav. Brain Sci. 20, 723767
tioning as a kind of bridge between the two. It thus seems 26 Gibson, J.J. (1996) The Senses Considered as Perceptual Systems,
likely that one key to understanding the nature and potency Houghton-Mifflin
of human thought and reason lies precisely in understanding 27 Gibson, J.J. (1979) The Ecological Approach to Visual Perception,
the complex relations and interactions between these various Houghton-Mifflin
28 Gibson, E.J. (1982) The concert of affordances in development, in
types of strategy and resource41. (Human language skills a
Minnesota Symposium on Child Psychology (Vol. 15) (Collins, W., ed.),
topic I have deliberately avoided in this review are a case in
pp. 5581, Huirdale
point; words and text are both real, external objects that we 29 Michaels, C. and Cavello, C. (1981) Direct Perception, Prentice-Hall
can encounter and manipulate and key instruments of inner, 30 Adolph, K. (1995) Psychophysical assessment of toddlers ability to cope
abstract, environmentally decoupled reason55,60.) with slopes J. Exp. Psychol. Hum. Percept. Perform. 21, 734750
31 Lee, D. and Reddish, P. (1981) Plummeting gannets: a paradigm of
The gulf between the embodied, embedded skills of the
ecological optics Nature 293, 293294
Bluefin tuna and the more de-coupled skills of the moralists
32 Tresilian, J.R. (1999) Visually timed action: time-out for tau? Trends
and mathematicians remains. But the size, nature and signifi- Cognit. Sci. 3, 301310
cance of this gap are matters for further research and debate. 33 McBeath, M., Shaffer, D. and Kaiser, M. (1995) How baseball outfielders
At the very least, an embodied cognitive science must now determine where to run to catch fly balls Science 268, 569573

look beyond the on-line production of tuned motor responses 34 Clark A. (1995) Moving minds: situating content in the service of real-
time success, in Philosophical Perspectives (Vol. 9: A-I, Connectionism
to the creation, maintenance and transformation of the inner
and Philosophical Psychology) (Tomberlin, J., ed.), pp. 89104, Ridgeview
and outer states that together allow us to know the world as 35 Rutkowska, J. (1993) The Computational Infant, Harvester Press
an arena for embodied action. 36 Neisser, U. (1976) Cognition and Reality, W.H. Freeman
37 Adolph, K. and Eppler, M. (1998) Development of visually guided
locomotion Ecol. Psychol. 10, 303321
Acknowledgements 38 Adolph, K., Eppler, M. and Gibson, E. (1993) Crawling versus walking
Thanks to Tim Van Gelder, Esther Thelen, Dan Dennett, Randy Beer, infants perception of affordance for locomotion over sloping surfaces
Michael Wheeler, Lynn Stein, Brian Cantwell Smith, Dominic Murphy, Child Dev. 64, 11581174
Timo Jarvilehto and the anonymous referees for invaluable help and 39 Hatsopoulis, N. and Warren, W. (1996) Resonance tuning in rhythmic
suggestions on these topics. arm movements J. Mot. Behav. 78, 314

350
Trends in Cognitive Sciences Vol. 3, No. 9, September 1999
 Clark Embodiment in cognitive science Opinion
40 Clark, A. and Toribio, J. (1995) Doing without representing? Synthese 51 Mitchell, M., Crutchfield, J. and Hraber, P. (1994) Evolving cellular
101, 401431 automata to perform computations Physica D 75, 361391
41 Smith, B.C. (1996) On the Origin of Objects, MIT Press 52 Clark, A. (1998) Where brain, body and world collide Daedalus 127,
42 Brentano, F. (1874) Psychologie vom Empirischen Standpunkt (transl. 257280
1973, Psychology from an Empirical Standpoint), Taylor & Francis, 53 Vygotsky, L. (1986) Thought and Language, MIT Press
London 54 Bruner, T. (1990) Acts of Meaning, Harvard University Press
43 Clark, A. and Grush, R. Towards a cognitive robotics Adaptive Behav. 55 Dennett, D. (1996) Kinds of Minds, Basic Books
(in press) 56 Wood, D., Bruner, J. and Russ, G. (1976) The role of tutoring in problem
44 Kirsh, D. (1992) Today the earwig, tomorrow man?, in Foundations of solving J. Child Psychol. Psychiatry 17, 89100
Artificial Intelligence (Kirsh, D., ed.), pp. 161184, MIT Press 57 Norman, D.A. (1993) Things That Make Us Smart, AddisonWesley
45 Van Leeuwen, C. (1998) Perception, in A Companion to Cognitive Science 58 McClelland, J. et al. (1986) The appeal of parallel distributed processing,
(Bechtel, W. and Graham, G., eds), pp. 265281, Blackwell in Parallel Distributed Processing: Explorations in the Microstructure of
46 Stein, L. (1994) Imagination and situated cognition J. Exp. Theoret. Artif. Cognition (Vol. II) (McClelland, J., Rumelhart, D. and PDP Research Group,
Intell. 6, 393407 eds), pp. 344, MIT Press
47 Mataric, M. (1992) Integration of representation into goal-driven 59 Clark, A. (1997) Economic reason: the interplay of individual learning and
behavior based robots IEEE J. Robotics Automation 8, 304312 external structure, in The Frontiers of the New Institutional Economics
48 Barsalou, L. Perceptual symbol systems Behav. Brain Sci. (in press) (Drobak, J., ed.), Academic Press
49 van Gelder, T. (1995) What might cognition be, if not computation? 60 Clark, A. (1998) Magic words: how language augments human
J. Philos. XCII, 345381 computation, in Language and Thought (Boucher, J. and Carruthers, P.,
50 Clark, A. (1997) The dynamical challenge Cognit. Sci. 21, 461481 eds), pp. 162183, Cambridge University Press

Neuromodulation:
acetylcholine and
memory consolidation
Michael E. Hasselmo

Clinical and experimental evidence suggests that hippocampal damage causes more
severe disruption of episodic memories if those memories were encoded in the recent
rather than the more distant past. This decrease in sensitivity to damage over time
might reflect the formation of multiple traces within the hippocampus itself, or the
formation of additional associative links in entorhinal and association cortices.
Physiological evidence also supports a two-stage model of the encoding process in
which the initial encoding occurs during active waking and deeper consolidation occurs
via the formation of additional memory traces during quiet waking or slow-wave sleep.
In this article I will describe the changes in cholinergic tone within the hippocampus in
different stages of the sleepwake cycle and will propose that these changes modulate
different stages of memory formation. In particular, I will suggest that the high levels
of acetylcholine that are present during active waking might set the appropriate
dynamics for encoding new information in the hippocampus, by partially suppressing
excitatory feedback connections and so facilitating encoding without interference from
previously stored information. By contrast, the lower levels of acetylcholine that are
present during quiet waking and slow-wave sleep might release this suppression and
Michael E. Hasselmo
thereby allow a stronger spread of activity within the hippocampus itself and from the is at the Department
hippocampus to the entorhinal cortex, thus facilitating the process of consolidation of of Psychology,
Boston University,
separate memory traces.
64 Cummington
Street, Boston,

T his article will describe some of the physiological and

neurochemical mechanisms that might mediate episodic
memory consolidation. I will then describe electrophysiological
data that support this concept and suggest that the two stages
MA 02215, USA.

tel: +1 617 495 3875

fax: +1 617 495 3728
memory consolidation. Firstly, I will review the neuropsycho- are linked to different stages of the sleepwake cycle when en- e-mail: hasselmo@bu.
logical and computational evidence for a two-stage model of coding and consolidation might occur. In the final section, I edu

1364-6613/99/$ see front matter 1999 Elsevier Science Ltd. All rights reserved. PII: S1364-6613(99)01365-0
351
Trends in Cognitive Sciences Vol. 3, No. 9, September 1999