Rob otics Philosophy of Mind using a

Screwdriver

Inman Harvey

Scho ol of Cognitive and Computing Sciences

University of Sussex

Brighton BN QH UK

Abstract

The design of autonomous rob ots has an intimate relationship

with the study of autonomous animals and humans rob ots pro

vide a convenient pupp et show for illustrating currentmyths ab out

cognition Like it or not any approach to the design of autonomous

rob ots is underpinned by some philosophical p osition in the designer

Whereas a philosophical p osition normally has to survive in debate

in a pro ject of building situated rob ots ones philosophical p osition

aects design decisions and is then tested in the real world doing

philosophy of mind with a screwdriver

Traditional Go o d Old Fashioned Articial Intelligence GOFAI

approaches have b een based on what is commonly called a Cartesian

split b etween b o dy and mind though the division go es backat

least to Plato The Dynamical Systems approach to cognition and

to rob ot design draws on other philosophical paradigms We shall

discuss how suchvaried philosophers as Heidegger MerleauPonty

or Wittgenstein in the improbable event of them wanting to build

rob ots mightbetemptedtosetaboutthetask

Intro duction

Car manufacturers need rob ots that reliably and mindlessly rep eat sequences

of actions in some wellorganised environment For many other purp oses au

tonomous rob ots are needed that will b ehave appropriately in a disorganised

environment that will react adaptively when faced with circumstances that

they havenever faced b efore Planetary exploring rob ots such as the So

journer rob ot sent to Mars cannot aord to wait the long time needed for

radio communication with p eople on earth for consultation on every indi

vidual movetheymake The user of a semiautonomous wheelchair should

b e able to delegate the same sort of decisions that a horse rider delegates to

herhorsehow to mano euvre around obstacles and react instinctively to

other trac Wewantsuch rob ots to b ehave to some extentintelligently

or adaptivelyinfacttobehave in some small part as if they had a mind

of their own

It has b een tempting to think of this as merely a technical scientic

problem we should study in an ob jective scientic fashion the basic re

quirements for adaptiveintelligence and then systematically engineer into

our rob ots what wehave found to b e necessary But likeitornotany

approach to the understanding of cognition and adaptiveintelligence and

hence to the design of autonomous rob ots is inevitably framed within some

philosophical p osition in the scientist or designer In a pro ject of building

situated rob ots ones philosophical p osition aects design decisions and is

then tested in the real world doing philosophy of mind with a screw

driver

We use basic working metaphors to make sense of scientic theories

billiard balls and waves on a p ond have b een much used in physics The

metaphor of animals or even humans as machines as comparable to the

technical artefacts that we construct is a p owerful one When we try to

build autonomous rob ots they are almost literally pupp ets acting to illus

trate our currentmyths ab out cognition The word myth sounds p ossibly

derogatory as it often implies a ction or halftruth it is not intended as

such here I am merely trying to emphasise that our view of cognition is a

humancentred view from the end of the second millennium some billion

years after the origin of life on this planet

Someone coming from the conventional scientic p ersp ectivemay sug

gest that our current cultural context is irrelevant After all ob jectivityin

science is all to do with discounting the accidental p ersp ectives of an ob

server and discovering universal facts and laws that all observers can agree

on from whatever place and time in which they are situated However to

pursue this line to o far leads one intoaparadox What theories if any

did the organisms of billion years ago have ab out the cognitive abilities of

their contemp oraries Clearly nothing like ours What theories mightour

descendants if any billion years hence have ab out the cognition of their

contemp oraries It would b e arrogant indeed unscientic to assume that

they would b e similar to ours

The Cop ernican revolutions in science have increased the scop e of our

ob jective understanding of the world by recognising that our observations

are not contextfree Cop ernicus and Galileo used their imagination and

sp eculated what the solar system mightlooklike if our view from the planet

Earth was not a privileged view from the xed centre of the universe but

merely one p ossible p ersp ective amongst many Darwin op ened up the way

to appreciating that Homo Sapiens is just one sp ecies amongst manywitha

common mo de of evolutionary development from one common origin Ein

stein brought ab out a fresh Cop ernican revolution with Sp ecial Relativity

showing how our understanding is increased when we abandon the idea of

some unique xed frame of reference for measuring the sp eed of anyobject

The history of science showsanumber of advances now generally accepted

stemming from a relativist p ersp ective that surprisingly is asso ciated with

an ob jective stance toward our role as observers

Cognitive science seems one of the last bastions to hold out against a

Cop ernican relativist revolution In this pap er I will broadly distinguish

between the preCop ernican views asso ciated with the computationalist ap

proach of classical Go o d Old Fashioned Articial Intelligence GOFAI and

the contextual situated approaches of nouvelle AI The two sides will b e

rather crudely p ortrayed with little attempt to distinguish the many dier

ing factions that can b e group ed under one ag or the other The dierent

philosophical views will b e asso ciated with the direct implications that they

have for the design of rob ots It is worth mentioning that Bro oks recent

collection of his early pap ers on rob otics Bro oks explicitly divides

the eight pap ers into four under the heading of Technology and four as

Philosophy though the division is somewhat arbitrary as the two asp ects

go together throughout

Cartesian or Classical approaches to Rob otics

Descartes working in the rst half of the seventeenth century is considered

by many to b e the rst mo dern philosopher A scientist and mathematician

as much as philosopher his ideas laid the groundwork for much of the

waywe view science to day In cognitive science the term Cartesian has

p erhaps rather unfairly to Descartes come to exclusively characterise a

set of views that treat the division b etween the mental and the physical

as fundamental the Cartesian cut Lemmen One form of the

Cartesian cut is the dualist idea that these are two completely separate

substances the mental and the physical which can exist indep endently of

each other Descartes prop osed that these twoworlds interacted in just one

place in humans the pineal gland in the brain Nowadays this dualism is not

very resp ectable yet the common scientic assumption rests on a variant

of this Cartesian cut that the physical world can b e considered completely

ob jectively indep endent of all observers

This is a dierent kind of ob jectivity from that of the Cop ernican sci

entic revolutions mentioned ab ove Those relied on the absence of any

privileged p osition on intersub jective agreementbetween observers inde

p endentofany sp ecic observer The Cartesian ob jectivity assumes that

therejustisawaytheworld is indep endentofany observer at all The

scientists job then is to b e a sp ectator from outside the world with a

Go dseye view from ab ove

When building rob ots this leads to the classical approach where the

rob ot is also a little scientistsp ectator seeking information from outside

ab out how the world is what ob jects are in which place The rob ot takes in

information through its sensors turns this into some internal representation

or mo del with which it can reason and plan and on the basis of this

formulates some action that is delivered through the motors Bro oks calls

this the SMPA or sensemo delplanact architecture Bro oks

The brain or nervous system of the rob ot can b e considered as a Black

Box connected to sensors and actuators such that the b ehaviour of the ma

chine plus brain within its environment can b e seen to b e intelligent The

question then is What to put in the BlackBox The classical compu

tationalist view is that it should b e computing appropriate outputs from

its inputs Or p ossibly they maysay that whatever it is doing should b e

interpretable as doing such a computation

The astronomer and her computer p erform computational algorithms

in order to predict the next eclipse of the mo on the sun mo on and earth

do not carry out such pro cedures as they drift through space The co ok

follows the algorithm recip e for mixing a cake but the ingredients do

not do so as they rise in the oven Likewise if I was capable of writing a

computer program which predicted the actions of a small creature this do es

not mean that the creature itself or its neurons or its brain was consulting

some equivalent program in deciding what to do

Formal computations are to do with solving problems suchaswhenis

the eclipse But this is an astronomers problem not a problem that the

solar system faces and has to solve Likewise predicting the next movement

of a creature is an animal b ehaviourists problem not one that the creature

faces However the rise of computer p ower in solving problems naturally

though regrettably led AI to the view that cognition equalled the solving of

problems the calculation of appropriate outputs for a given set of inputs

The brain on this view was surely some kind of computer What was

the problem that the neural program had to solve the inputs must b e

sensory but what were the outputs

Whereas a rob oticist would talk in terms of motor outputs the more

cerebral academics of the infantAIcommunity tended to think of plans or

representations as the prop er outputs to study They treated the brain as

the manager who do es not get his own hands dirty but rather issues com

mands based on highlevel analysis and calculated strategy The manager

sits in his command p ost receiving a multitude of p ossibly garbled messages

from a myriad sensors and tries to work out what is going on Prop onents of

this view tend not to admit explicitly indeed they often denyvehemently

that they think in terms of a homunculus in some inner chamb er of the

brain but they have inherited a Cartesian split b etween mind and brain

and in the nal analysis they rely on such a metaphor

What is the Computer Metaphor

The concepts of computers and computations and programs haveavariety

of meanings that shade into each other On the one hand a computer is a

formal system with the same p owers as a Turing Machine assuming

the memory is of adequate size On the other hand a computer is this

ob ject sitting in frontofmenow with screen and keyb oard and indenite

quantities of software

A program for the formal computer is equivalent to the presp ecied

marks on the Turing machines tap e For a given starting state of this

machine the course of the computation is wholly determined by the program

and the Turing machines transition table it will continue until it halts with

the correct answer unless p erhaps it continues forever usually considered

a bad thing

On the machine on my desk I can write a program to calculate a suc

cession of coordinates for the parab ola of a cricketball thrown into the

air and display these b oth as a list of gures and as a curvedrawn on the

screen Here I am using the machine as a convenient fairly userfriendly

Turing machine

However most programs for the machine on my desk are very dierent

At the moment it is amongst many other things running an editor or word

pro cessing program It sits there and waits sometimes for very long p erio ds

indeed until I hit a key on the keyb oard when it virtually immediately p ops

asymbol into an appropriate place on the screen unless particular control

keys are pressed causing the le to b e written or edits to b e made Virtually

all of the time the program is waiting for input which it then pro cesses near

instantaneously In general it is a good thing for such a program to continue

for ever or at least until the exit command is keyed in

The cognitivist approach asserts that something with the p ower of a

Turing machine is b oth necessary and sucient to pro duce intelligence

b oth human intelligence and equivalentmachine intelligence Although not

usually made clear it would seem that something close to the mo del of a

wordpro cessing program is usually intended ie a program that constantly

awaits inputs and then nearinstantaneously calculates an appropriate out

put b efore settling down to await the next input Life so I understand the

computationalists to hold is a sequence of such individual events p erhaps

pro cessed in parallel

Time in Computations and in Connection

ism

One particular asp ect of a computational mo del of the mind whichderives

from the underlying Cartesian assumptions common to traditional AI is

the way in which the issue of time is swept under the carp et only the

sequential asp ect of time is normally considered In a standard computer

op erations are done serially and the lengths of time taken for each program

step are for formal purp oses irrelevant In practice for the machine on

my desk it is necessary that the timesteps are fast enough for me not to

get b ored waiting Hence for a serial computer the only requirementisthat

individual steps take as short a time as p ossible In an ideal world anygiven

program would b e practically instantaneous in running except of course for

those unfortunate cases when it gets into an innite lo op

The common connectionist assumption is that a connectionist network is

in some sense a parallel computer Hence the time taken for individual pro

cesses within the network should presumably b e as short as p ossible They

cannot b e considered as b eing eectively instantaneous b ecause of the ne

cessityofkeeping parallel computations in step The standard assumptions

made fall into two classes

The timelag for activations to pass from any one no de to another it

is connected to including the time taken for the outputs from a no de

to b e derived from its inputs is in all cases exactly one unit of time

eg a backpropagation or an Elman network

Alternatively just one no de at a time is up dated indep endently of the

others and the choice of whichnodeisdealtwithnextisstochastic

eg a Hopeld net or a Boltzmann machine

The rst metho d follows naturally from the computational metaphor

from the assumption that a computational pro cess is b eing done in parallel

The second metho d is closer to a dynamical systems metaphor yet still

computational language is used It is suggested that a network after ap

propriate training will when presented with a particular set of inputs then

sink into the appropriate basin of attraction which appropriately classies

them The network is used as either a distributed contentaddressable mem

ory or as a classifying engine as a mo dule taking part in some largerscale

computation The sto chastic metho d of relaxation of the network maybe

used but the dynamics of the network are thereby made relatively simple

and not directly relevant to the wider computation It is only the stable

attractors of the network that are used It is no coincidence that the at

tractors of suchastochastic network are immensely easier to analyse than

any nonsto chastic dynamics

It might b e argued that connectionists are inevitably abstracting from

real neural networks and inevitably simplifying In due course so this ar

gument go es they will slowly extend the range of their mo dels to include

new dimensions such as that of time What is so sp ecial ab out time

whycannotitwait Well the simplicity at the formal level of connection

ist architectures which need synchronous up dates of neurons disguises the

enormous complexity of the physical machinery needed to maintain a uni

versal clo cktickover distributed no des in a physically instantiated network

From the p ersp ectiveadvo cated here clo cked networks form a particular

complex subset of all realtime dynamical networks ones need b e and if

anything they are the ones that should b e left for later van Gelder

Amuch broader class of networks is that where the timelags on indi

vidual links b etweennodesisarealnumb er whichmay b e xed or may

vary in a similar fashion to weightings on such links A pioneering attempt

at a theory that incorp orates such timelags as an integral part is given in

Malsburg and Bienensto ck

In neurobiological studies the assumption seems to b e widespread that

neurons are passing information b etween each other enco ded in the rate

of ring By this means it would seem that real numb ers could b e passed

even though signals passing along axons seem to b e allornone spikes This

assumption is very useful indeed p erhaps invaluable in certain areas such

as early sensory pro cessing Yet it is p erverse to assume that this is true

throughout the brain a p erversitywhich while p erhaps not caused by the

computational metaphor is certainly aided by it Exp eriments demonstrat

ing that the individual timing of neuronal events in the brain and the

temp oral coincidence of signals passing down separate synre chains can

For a simple mo del without loss of generalityany time taken for outputs to b e derived

from inputs within a no de can b e set to zero by passing any nonzero value on instead

to the links connected to that no de

b e of critical imp ortance are discussed in Ab eles

What is a Representation

The concept of symb olic reference or representation lies at the heart of

analytic philosophy and of computer science The underlying assumption

of many is that a real world exists indep endently of anygiven observer and

that symb ols are entities that can stand for ob jects in this real world

in some abstract and absolute sense In practice the role of the observer in

the act of representing something is ignored

Of course this works p erfectly well in worlds where there is common

agreement amongst all observers explicit or implicit agreement on the

usages and denitions of the symb ols and the prop erties of the world that

they represent In the worlds of mathematics or formal systems this is the

case and this is reected in the anonymity of tone and use of the passive

tense in mathematics Yet the dep endency on such agreement is so easily

forgotten or p erhaps ignored in the assumption that mathematics is the

language of Go d

Asymbol P is used by a p erson Q to represent or refer to an ob ject

R to a p erson S Nothing can b e referred to without someb o dy to do the

referring Normally Q and S are memb ers of a community that have come

to agree on their symb olic usages and training as a mathematician involves

learning the practices of such a communityThevo cabulary of symb ols can

b e extended by dening them in terms of alreadyrecognised symb ols

The English language and the French language are systems of symbols

used by p eople of dierent language communities for communicating ab out

their worlds with their similarities and their dierentnuances and cliches

The languages themselves havedevelop ed over thousands of years and the

induction of eachchild into the use of its native language o ccupies a ma jor

slice of its early years The fact that nearly all the time we are talking

English we are doing so to an Englishsp eaker including when we talk

to ourselves makes it usually an unnecessary platitude to explicitly draw

attention to the community that sp eaker and hearer b elong to

Since symb ols and representation stand rmly in the linguistic domain

another attribute they p ossess is that of arbitrariness from the p ersp ective

of an observer external to the communicators When I raise my forenger

with its backtoyou and rep eatedly b end the tip towards me the chances

are that you will interpret this as come here This particular Europ ean and

American sign is just as arbitrary as the Turkish equivalent of placing the

hand horizontally facing down and apping it downwards Dierent actions

or entities can represent the same meaning to dierent communities and the

same action or entity can represent dierent things to dierentcommunities

In Mao TseTungs China a red trac light meant GO

In the more general case and particularly in the eld of connectionism

and cognitive science when talking of representation it is imp erativeto

make clear who the users of the representation are and it should b e p ossible

to at a minimum suggest how the convention underlying the representation

arose In particular it should b e noted that where one and the same entity

can represent dierent things to dierentobservers conceptual confusion

can easily arise When in doubt always make explicit the Q and S when P

is used by Q to represent R to S

In a computer program a variable popsize may b e used by the pro

grammer to represent to herself and to any other users of the program

the size of a p opulation Inside the program a variable i may b e used to

representa counter or internal variable in manycontexts In each of these

contexts a metaphor used by the programmer is that of the program de

scribing the actions of various homunculi some of them keeping countof

iterations some of them keeping trackofvariables and it is within the

context of particular groups of suchhomunculi that the symb ols are repre

senting But how is this notion extended to computation in connectionist

networks

Representation in Connectionism

When a connectionist network is b eing used to do a computation in most

cases there will b e input hidden and output no des The activations on

the input and output no des are decreed by the connectionist to represent

particular entities that have meaning for her in the same wayaspopsize

is in a conventional program But then the question is raised what ab out

internal representations

If a connectionist network is providing the nervous system for a rob ot a

dierentinterpretation might b e put on the inputs and outputs But for the

purp ose of this section the issues of internal representation are the same

All to o often the hidden agenda is based on a Platonic notion of repre

sentation what do activations or patterns of activations represent in some

absolute sense to Go d The b ehaviour of the innards of a trained network

are analysed with the same eagerness that a sacriced chickens innards are

interpreted as representing ones future fate There is however a more prin

cipled way of talking in terms of internal representations in a network but

away that is critically dep endent on the observers decomp osition of that

network Namely the network must b e decomp osed by the observer into

two or more mo dules that are considered to b e communicating with each

other by means of these representations

Where a network is explicitly designed as a comp osition of various mo d

ules to do various subtasks for instance a mo dule could b e a layer or a

group of laterally connected no des within a layer then an individual ac

tivation or a distributed group of activations can b e deemed to represent

an internal variable in the same way that i did within a computer program

However unlike a program whichwears its origins on its sleeve in the form

of a program listing a connectionist network is usually deemed to b e inter

nally nothing more than a collection of no des directed arcs activations

weights and up date rules Hence there will usually b e a large number of

p ossible ways to decomp ose such a network with little to cho ose b etween

them and it dep ends on just where the b oundaries are drawn just who is

representing what to whom

It might b e argued that some ways of decomp osing are more natural

than others a p ossible criterion b eing that two sections of a network should

have a lot of internal connections but a limited numb er of connecting arcs

between the sections Yet as a matter of interest this do es not usually hold

for what is p erhaps the most common form of decomp osition into layers

The notion of a distributed representation usually refers to a representation

b eing carried in parallel in the communication from one layer to the next

where the layers as a whole can b e considered as the Q and S in the formula

P is used by Q to represent R to S

An internal representation according to this view only makes sense

relative to a particular decomp osition of a network chosen byanobserver

To assert of a network that it contains internal representations can then only

b e justied as a rather to o terse shorthand for asserting that the sp eaker

prop oses some such decomp osition Regrettably this do es not seem to b e

the normal usage of the word in cognitive science yet I am not aware of

anywelldened alternative denition

Are Representations Needed

With this approach to the representation issue then any network can b e

decomp osed in a varietyofways into separate mo dules that the observer

considers as communicating with each other The interactions b etween such

mo dules can ipso facto b e deemed to b e mediated by a representation

Whether it is useful to do so is another matter

Asso ciated with the metaphor of the mind or brain or an intelligent

machine as a computer go assumptions of functional decomp osition Since

a computer formally manipulates symbols yet it is lightwaves that impinge

on the retina or the camera surely so the story go es some intermediate

agency must do the necessary translating Hence the traditional decomp o

sition of a cognitivesysteminto a p erception mo dule which takes sensory

inputs and pro duces a world mo del this is passed onto a central planning

mo dule which reasons on the basis of this world mo del passing on its deci

sions to an action mo dule which translates them into the necessary motor

actions This functional decomp osition has b een challenged and an alter

native b ehavioural decomp osition prop osed by Bro oks in eg Bro oks

In particular the computationalist or cognitivist approach seems to im

ply that communication b etween anysuch mo dules is a oneway pro cess any

feedback lo ops are within a mo dule Within for instance backpropagation

the backward propagation of errors to adjust weights during the learning

pro cess is treated separately from the forward pass of activations This

helps to maintain the computational ction by conceptually separating the

two directions and retaining a feedforward network But consider the fact

that within the primate visual pro cessing system when visualised as a net

work there are many more bres coming back from the visual cortex into

the Lateral Geniculate Nucleus LGN than there are bres going from the

retina to the LGN in the correct direction How do es the computationalist

make sense of this

Marr in Marr reprinted in Bo den classies AI theories

into Typ e and Typ e where a Typ e theory can only solve a problem

bythesimultaneous action of a considerable numb er of pro cesses whose in

teraction is its own simplest descriptionItwould seem that typ e systems

can only b e decomp osed arbitrarily and hence the notion of representation

is less likely to b e useful This is in contrast to a Typ e theory where a

problem can b e decomp osed into a form that an algorithm can b e formu

lated to solve by divide and conquerTyp e theories are of course the more

desirable ones when they can b e found but it is an empirical matter whether

they exist or not In mathematics the colour theorem has b een solved in

a fashion that requires a large numb er of sp ecial cases to b e exhaustively

worked out in thousands of hours of computation App el and Haken

It is hop ed that there were no hardware faults during the pro of pro cedure

and there is no way that the pro of as a whole can b e visualised and assessed

bya human There is no a priori reason why the workings of at least parts



of the brain should not b e comparably complex or even more so This



For the purp oses of making an intelligentmachine or rob ot it has in the past seemed

obvious that only Typ e techniques could b e prop osed However evolutionary techniques

can b e interpreted as there is no a priori reason why all parts of the brain

should b e in such a mo dular form that representationtalk is relevant The

answer to the question p osed in the title of this section is no This do es

not rule out the p ossibility that in some circumstances representationtalk

might b e useful but it is an exp erimental matter to determine this

Alternatives to Cartesianism

Hub ert Dreyfus came out with a trenchant criticism of the classical AI

computationalist approach to cognition in the s He came from a very

dierent set of philosophical traditions lo oking for inspiration to twentieth

century philosophers such as Heidegger MerleauPonty and Wittgenstein

Initially he pro duced a rep ort for the RAND Corp oration provo catively

entitled Alchemy and Articial Intelligence in Dreyfus Later

more p opular b o oks are Dreyfus and with his brother Dreyfus and

Dreyfus These are amongst the easiest ways for someb o dy with a

conventional background in computer science cognitive science or rob otics

to approach the alternative set of philosophical views Nevertheless the

views are suciently strange to those broughtupinto the Cartesian wayof

thinking that at rst sight Dreyfus app ears to b e mystical or antiscientic

This is not the case

Another view of cognition from a phenomological or Heideggerian p er

sp ectiveisgiven in Winograd and Flores Winograd was instrumen

tal in some of the classical early GOFAI work b efore coming around to a

very dierent viewp oint A dierent Heideggerian p ersp ective is given in

Wheeler within Bo den The relevance of MerleauPontyis

drawn out in Lemmen A dierent p ersp ective that is similarly op

p osed to the Cartesian cut is given in Maturana and Varela Varela

et al Amuch more general textb o ok on rob otics that is written

from a situated and emb o died p ersp ective is Pfeifer and Scheier

Heidegger rejects the simplistic ob jective view that the ob jectivephys

ical world is the primary reality that we can b e certain of He similarly

rejects the opp osite idealistic or sub jective view that our thoughts are the

primary reality Instead the primary reality is our exp erience of the world

that cannot exist indep endently of one or the other Our everyday practical

lived exp erience as we reach for our coee or switch on the light is more

fundamental than the detached theoretical reection that we use as rational

scientists Though Heidegger himself would not put it this way this makes

sense from a Darwinian evolutionary p ersp ective on our own sp ecies From

need not restrict themselves in this fashion Harvey et al

this p ersp ective our language using and reasoning p owers probably arrived

in H sapiens over just the last million or so years in our billion year evo

lutionary history and is merely the thin layer of icing on the cake From

b oth a phylogenetic and an ontogenetic view we are organisms and animals

rst reasoning humans only later

As humans of course detached theoretical reasoning is one of our hall

marks indeed the Darwinian view presented in the previous paragraph is

just one such piece of reasoning However practical knowhow is more fun

damental than suchdetached knowingthat This is a complete reversal

of the typical approach of a Cartesian cognitive scientist or rob oticist who

would attempt to reduce the everyday action of a human or rob ot reaching

for the coee mug into a rational problemtob esolved hence the Cartesian

sensemo delplanact cycle

The archetypal Heideggerian example is that of hammering in a nail

When we do this normally the arm with the hammer naturally and without

thought go es through its motions driving the hammer home It is only when

something go es wrong such as the head of the hammer ying o or the nail

b ending in the wood that wehave to concentrate and start reecting on the

situation rationalising what the b est plan of action will b e Information

pro cessing knowingthat is secondary and is built on top of our everyday

rhythms and practices of practical knowhow knowhowwhich cannot b e

reduced to a set of rules that we implement This is true Wittgenstein

suggests even for our language skills

In general we dont use language according to strict rules it

hasnt b een taughtusby means of strict rules either Wittgen

stein

For the rob oticist this antiCartesian alternative philosophy seems at

rst sight negative and unhelpful For everyday rob ot actions this implies

that we should do without planning without the computational mo del

without internal representations but nothing has yet b een oered to replace

such metho ds The two lessons that need to b e learnt initially is that

cognition is

Situated a rob ot or human is always already in some situation rather

than observing from outside

and Emb o died a rob ot or human is a p erceiving b o dy rather than a

disemb o died intelligence that happ ens to have sensors

One nice example of a situated emb o died rob ot is the simple walking

machine of McGeer which uses passive dynamic walking McGeer

Figure McGeers passive dynamic walker

McGeer This is a twolegged walking rob ot with each leg made of

just an upp er and lower limb connected by a knee joint This knee acts as a

human knee allowing b ending freely in one direction but not in the other

At the b ottom of eachlower limb is a fo ot shap ed in an arc and the two

legs are hinged together at the waist The dimensions are carefully worked

out but then that is the complete walking rob ot with no control system

needed at all When it is started o on a gently sloping incline the legs will

walk along in a remarkably natural humanlike motion all the control has

come from the natural dynamics through b eing situated and emb o died

The walking rob ot do es not take in information through sensors and

do es not compute what its current p osition is and what its next move

should b e The designers of course carefully computed the appropriate

dimensions In the natural world organisms and animals have their b o dily

dimensions designed through natural evolution

The Dynamical Systems alternative

This last example is a rob ot designed on nonCartesian principles using an

alternative view that has gained favour in the last decade within AI circles

though its origins date back at least to the early cyb ernetics movement

One description of this is the Dynamical Systems view of cognition

animals are endowed with nervous systems whose dynamics

are such that when coupled with the dynamics of their b o dies

and environments these animals can engage in the patterns of

behavior necessary for their survival Beer Gallagher

Atthisstagewedowngrade the signicance of intel ligence for AI in

favour of the concept of adaptive behaviour Intelligence is nowjustone

form of adaptivebehaviour amongst many the ability to reason logically

ab out chess problems may b e adaptive in particular rened circles but the

ability to cross the road safely is more widely adaptive We should note the

traditional priorities of AI the computationalists emphasis on reasoning led

them to assume that everydaybehaviour of sensorimotor co ordination must

b e built on top of a reasoning system Sensors and motors in their view

are merely to ols for informationgathering and planexecution on b ehalf

of the central executive where the real work is done Many prop onents of

an alternative view including myself would want to turn this on its head

logical reasoning is built on top of linguistic b ehaviour whichisbuilton

prior sensorimotor abilities These prior abilities are the fruit of billions of

years of evolution and language has only b een around for the last few tens

of thousands of years

A dynamical system is formally any system with a nite numb er of state

variables that can change over time the rate of change of any one suchvari

able dep ends on the currentvalues of any or all of the variables in a regular

fashion These regularities are typically summed up in a set of dierential

equations A Watt governor for a steam engine is a paradigmatic dynamical

system van Gelder and we can treat the nervoussystemplusbody

of a creature or rob ot as one also The b ehaviour of a dynamical system

suchasthegovernor dep ends also on the currentvalue of its external in

puts from the steam engine whichenter the relevant dierential equations

as parameters In a complementary way the output of the governor acts

as a parameter on the equations which describ e the steam engine itself as

a dynamical system One thing that is very rapidly learnt from handson

exp erience is that two such indep endent dynamical systems when coupled

together into eg steamengineplusgovernor treated now as a single dy

namical system often b ehave in a counterintuitive fashion not obviously

related to the uncoupled b ehaviours

Treating an agent creature human or rob ot as a dynamical sys

tem coupled with its environment through sensors and motors inputs and

outputs leads to a metaphor of agents b eing perturbed in their dynam

ics through this coupling in contrast to the former picture of such agents

computing appropriate outputs from their inputs The view of cognition

entailed by this attitude ts in with Varelas characterisation of cognition

as embodied action

By using the term embodied we mean to highlighttwopoints

rst that cognition dep ends up on the kinds of exp erience that

come from having a body with various sensorimotor capacities

and second that these individual sensorimotor capacities are

themselves emb edded in a more encompassing biological psy

chological and cultural context By using the term action we

mean to emphasise once again that sensory and motor pro cesses

p erception and action are fundamentally inseparable in lived

cognition Indeed the two are not merely contingently linked

in individuals they havealsoevolved together Varela et al

Evolutionary Rob otics and Behaviourism

Moving from natural agents to articial rob ots the design problem that a

rob ot builder faces is now one of creating the internal dynamics of the rob ot

and the dynamics of its coupling its sensorimotor interactions with its en

vironment such that the rob ot exhibits the desired b ehaviour in the right

context Designing such dynamical systems presents problems unfamiliar

to those who are used to the computational approach to cognition

A primary dierence is that dynamics involves time real time Whereas

a computation of an output from an input is the same computation whether

it takes a second or a minute the dynamics of a creature or rob ot has to b e

matched in timescale to that of its environment A second dierence is that

the traditional design heuristic of divide and conquer cannot b e applied in

the same way It is not clear how the dynamics of a control system should

b e carved up into smaller tractable pieces and the design of any one small

comp onent dep ends on an understanding of howitinteracts in real time

with the other comp onents suchinteraction p ossibly b eing mediated via

the environment This is true for b ehavioural decomp osition of control sys

tems Bro oks as well as functional decomp osition However Bro oks

subsumption architecture approach oers a dierent design heuristic rst

build simple complete rob ots with b ehaviours simple enough to understand

and then incrementally add new b ehaviours of increasing complexityorva

riety one at a time which subsume the previous ones Before the designer

adds a new control system comp onent in an attempt to generate a new

b ehaviour the rob ot is fully tested and debugged for its earlier b ehaviours

then the new comp onent is added so as to keep to a comprehensible and

tractable minimum its eects on earlier parts

This approach is explicitly describ ed as b eing inspired by natural evolu

tion but despite the design heuristics it seems that there is a practical limit

to the complexitythata human designer can handle in this way Natural

Darwinian evolution has no such limits hence the more recentmoves to

wards the articial evolution of rob ot control systems Harvey et al

In this work a genetic enco ding is set up such that an articial genotyp e

typically a string of s and s sp ecies a control system for a rob ot This

is visualised and implemented as a dynamical system acting in real time

dierent genotyp es will sp ecify dierentcontrol systems A genotyp e may

additionally sp ecify characteristics of the rob ot b o dy and sensorimotor

coupling with its environment When wehave settled on some particular

enco ding scheme and wehave some means of evaluating rob ots at the

required task we can apply articial evolution to a p opulation of genotyp es

over successive generations

Typically the initial p opulation consists of a numb er of randomly gener

ated genotyp es corresp onding to randomly designed control systems These

are instantiated in a real rob ot one at a time and the rob ot b ehaviour that

results when placed in a test environment is observed and evaluated After

the whole p opulation has b een scored their scores can b e compared for an

initial random p opulation one can exp ect all the scores to b e abysmal but

some through chance are less abysmal than others A second generation

can b e derived from the rst by preferentially selecting the genotyp es of

those with higher scores and generating ospring which inherit genetic ma

terial from their parents recombination and mutation is used in pro ducing

the ospring p opulation which replaces the parents The cycle of instan

tiation evaluation selection and repro duction then continues rep eatedly

each time from a new p opulation which should have improved over the av

erage p erformance of its ancestors Whereas the intro duction of new variety

through mutation is blind and driven bychance the op eration of selection

at each stage gives direction to this evolutionary pro cess

This evolutionary algorithm comes from the same family as Genetic Al

gorithms and Genetic Programming whichhave b een used with success on

thousands of problems The technique applied to rob otics has b een exp eri

mental and limited to date It has b een demonstrated successfully on simple

navigation problems recognition of targets and the use of minimal vision

or sonar sensing in uncertain real world environments Harvey et al

Thompson One distinguishing feature of this approach using blind

evolution is that the resulting control system designs are largely opaque and

incomprehensible to the human analyst With some considerable eort sim

ple control systems can b e understo o d using the to ols of dynamical systems

theory Husbands et al However it seems inevitable that for the

same reasons that it is dicult to design complex dynamical systems it is

also dicult to analyse them

This is reected in the metho dology of Evolutionary Rob otics which

once the framework has b een established concerns itself solely with the

b ehaviour of rob ots if it walks likea duckandquacks likea duck it

is a duck For this reason wehave sometimes b een accused of b eing

the New Behaviourists but this emphasis on b ehaviour assumes that



there are signicantinternal states and in my view is compatible with

the attribution of adaptiveintelligence A ma jor conceptual advantage

that Evolutionary Rob otics has over classical AI approachestorobotics

is that there is no longer a mystery ab out how one can get a rob ot to have

needs and wants In the classical version the insertion of a value function

robot avoid obstacle often leaves p eople uncomfortable as to whether it

is the rob ot or the programmer who has the desires In contrast genera

tions of evolutionary selection that tends to eliminate rob ots that crash into

the obstacle pro duces individual rob ots that do indeed avoid it and here it

seems much more natural that it is indeed the robot which has the desire

Relativism

Itake a Relativist p ersp ective which contrary to the naive p opular view

do es not imply solipsism or sub jectivism or an anythinggo es attitude to

science The history of science shows a number of advances now gener

ally accepted that stem from a relativist p ersp ective which surprisingly

is asso ciated with an ob jective stance toward our role as observers The

Cop ernican revolution abandoned our privileged p osition at the centre of

the universe and to ok the imaginativeleapofwondering how the solar

system would lo ok viewed from the Sun or another planet Scientic ob

jectivity requires theories to b e general to hold true indep endently of our

particular idiosyncratic p ersp ective and the relativism of Cop ernicus ex

tended the realm of the ob jective Darwin placed humans amongst the

other living creatures of the universe to b e treated on the same fo oting

With Sp ecial Relativity Einstein carried the Cop ernican revolution fur

ther by considering the viewp oints of observers travelling near to the sp eed

of light and insisting that scientic ob jectivity required that their p ersp ec

tives were equally privileged to ours Quantum physics again brings the

observer explicitly into view



Not signicant in the sense of representational internal states are mentioned here

to dierentiate evolved dynamical control systems whichtypically haveplentyofinter

nal state from those control systems restricted to feedforward inputoutput mappings

typical of reactive rob otics

Cognitive scientists must b e careful ab ove all not to confuse ob jects that

are clear to them that have an ob jective existence for them with ob jects

that have a meaningful existence for other agents A rob oticist learns very

earlyonhow dicult it is to make a rob ot recognise something that is crystal

clear to us such as an obstacle or a do or It makes sense for us to describ e

such an ob ject as existing for that rob ot if the physical sensorimotor

coupling of the rob ot with that ob ject results in rob ot b ehaviour that can

b e correlated with the presence of the ob ject By starting the previous

sentence with It makes sense for us to describ e I am acknowledging our

own p osition here acting as scientists observing a world of cognitive agents

such as rob ots or p eople this ob jective stance means we place ourselves

outside this world lo oking in as go dlike creatures from outside Our theories

can b e scientically ob jective which means that predictions should not b e

dep endent on incidental factors such as the nationality or lo cation or star

sign of the theorist

When I see a red sign this red sign is an ob ject that can b e discussed

scientically This is another wayofsaying that it exists for me for you

and for other human observers of any nationality though it do es not exist

for a bacterium or a mole We construct these ob jects from our exp erience



and through our acculturation as humans through education Just as our

capacity for language is phylogenetically built up on our sensorimotor capac

ities so our ob jects our scientic concepts are built out of our exp erience

But our phenomenal exp erience itself cannot b e an ob jective thing that can

b e discussed or compared with other things It is primary in the sense that

it is only through having phenomenal exp erience that we can create things

ob jective things that are secondary

Conclusion

Likeitornotany approach to the design of autonomous rob ots is un

derpinned by some philosophical p osition in the designer There is no

philosophyfree approach to rob ot design though sometimes the phi

losophy arises through accepting unthinkingly and without reection the

approach within which one has b een brought up GOFAI has b een predi

cated on some version of the Cartesian cut and the computational approach

has had enormous success in building sup erb to ols for humans to use but



It makes no sense to discuss for us humans to discuss the existence of ob jects

in the absence of humans And in an attempt to forestall the predictable ob jections

this view do es not imply that we can just p osit the existence of any arbitrary thing as

our whim takes us

it is simply inappropriate for building autonomous rob ots

There is a dierent philosophical tradition which seeks to understand

cognition in terms of the priorityoflived phenomenal exp erience the pri

orityofeveryday practical knowhowover reective rational knowingthat

This leads to very dierent engineering decisions in the design of rob ots

to building situated and embodied creatures whose dynamics are suchthat

their coupling with their world leads to sensible b ehaviours The design

principles needed are very dierent Bro oks subsumption architecture is

one approach Evolutionary Rob otics is another Philosophydoesmakea

practical dierence

Acknowledgments

I thank the EPSRC and the University of Sussex for funding and Shirley

Kitts for philosophical orientation

References

Ab eles M Local Cortical Circuits an Electrophysiological Study

SpringerVerlag

App el K and Haken W Every planar map is four colorable

American Mathematical Society Contemporary Mathematics

Beer R D and Gallagher J C Evolving dynamic neural networks

for adaptive b ehavior Adaptive Behavior

Bo den M A The Philosophy of Articial Intel ligen ce Oxford Uni

versity Press

Bo den M A The Philosophy of Articial Life Oxford University

Press

Bro oks R Cambrian Intelligence the Early History of the New AI

MIT Press Cambridge MA

Dreyfus H L Alchemy and Articial Intelligence RAND Corp o

ration Pap er P December

Dreyfus H L What Computers Cant Do a Critique of Articial

Reason Harp er and Row New York

Dreyfus H L and Dreyfus S E Mind Over Machine

Harvey I Husbands P Cli D Thompson A and Jakobi N

Evolutionary rob otics the Sussex approach Journal of Robotics and

Autonomous Systems v pp

Husbands P Harvey I and Cli D Circle in the round State

space attractors for evolved sighted rob ots Journal of Robotics and

Autonomous Systems Special Issue on The Biology and Technology

of Intel ligent Autonomous Agents

Lemmen R Towards a NonCartesian Cognitive Science in the

light of the philosophy of MerleauPonty DPhil Thesis Universityof

Sussex

Malsburg C von der and Bienensto ck E Statistical co ding

and shortterm plasticity a scheme for knowledge representation in

the brain In Bienensto ck E FougelmanSoulie F and Wiesbuch

G eds Disordered Systems and Biological Organization Springer

Verlag

Marr D C Articial Intelligence a p ersonal view Articial Intel

ligence

Maturana H and Varela F The TreeofKnowledge The Biological

Roots of Human UnderstandingShambhala Press Boston

McGeer T Passive dynamic walking IntJRobotics Research

McGeer T Dynamics and Control of Bip edal Lo comotion J Theor

Biol

Pfeifer R and Scheier C Understanding Intel ligence MIT Press

Thompson A Evolving electronic rob ot controllers that exploit

hardware resources In Moran F Moreno A Merelo JJ and

Chacon P eds Proceedings of ThirdEuropean ConferenceonAr

ticial Life SpringerVerlag pp

van Gelder T What might cognition b e if not computation Tech

nical Rep ort Indiana University Cognitive Sciences Reprinted in

Journal of Philosophy

Varela F Thompson E and Rosch E The Embodied MindMIT

Press

Wheeler M From Rob ots to Rothko The Bringing Forth of Worlds

In Bo den M ed The Philosophy of Articial Life Oxford University

Press

Winograd T and Flores F Understanding Computers and Cog

nition A New Foundation for Design Ablex Publishing Corp oration

New Jersey

Wittgenstein L The Blue and Brown Books Basil Blackwell Oxford