Emotions from brain to robot1.pdf

Review

TRENDS in Cognitive Sciences Vol.8 No.12 December 2004

Emotions: from brain to robot

Michael A. Arbib 1 and Jean-Marc Fellous 2

1 Computer Science, Neuroscience and USC Brain Project, University of Southern California, Los Angeles, CA 90089-2520, USA

2 Biomedical Engineering Department and Center for Cognitive Neuroscience, Duke University, Durham, NC 27708-0281, USA

Some robots have been given emotional expressions in

an attempt to improve human–computer interaction. In

this article we analyze what it would mean for a robot to

have emotion, distinguishing emotional expression for

communication from emotion as a mechanism for the

organization of behavior. Research on the neurobiology

of emotion yields a deepening understanding of inter-

acting brain structures and neural mechanisms rooted in

neuromodulation that underlie emotions in humans and

other animals. However, the chemical basis of animal

function differs greatly from the mechanics and compu-

tations of current machines. We therefore abstract from

biology a functional characterization of emotion that

does not depend on physical substrate or evolutionary

history, and is broad enough to encompass the possible

emotions of robots.

noting that not all emotions need be like human emotions.

We analyze emotion in two main senses:

(1) Emotional expression for communication and social

coordination.

(2) Emotion for organization of behavior (action

selection, attention and learning).

The ﬁrst concerns ‘external’ aspect of emotions; the

second ‘internal’ aspects. In animals, these aspects have

co-evolved. How might they enter robot design? Both

robots and animals need to survive and perform efﬁciently

within their ‘ecological niche’ and, in each case, patterns of

coordination will greatly inﬂuence the suite of relevant

emotions (if such are indeed needed) and the means

whereby they are communicated.

A key function of emotion is to communicate simpliﬁed

but high impact information. A scream is extremely poor

in information (it says nothing about the cause for alarm),

but its impact on others is high. Moreover, neurobiology

shows that simpliﬁed but high impact information is

communicated between brain areas, through the very

different ‘vocabulary’ of neuromodulation.

The similarity in facial expressions between certain

animals and humans prompted classic evolutionary

analyses [2] , which support the view that mammals

(at least) have emotions (although not necessarily the

same as human emotions), and work reviewed below

explores their (neuro)biological underpinnings. What of

robots? Robots are mechanical devices with silicon

‘brains’, not products of biological evolution. But as we

better understand biological systems we will extract ‘brain

operating principles’ that do not depend on the physical

medium in which they are implemented. These principles

might then be instantiated for particular robotics archi-

tectures to the point where we might choose to speak of

robot-emotions.

Many researchers (see for example [3,4] ) have proposed

explicit functions for emotions: coordinating behavioral

response systems; shifting behavioral hierarchies; com-

munication and social bonding; short-cut cognitive proces-

sing; and facilitating storage and recall of memories.

However, emotions are not always beneﬁcial [5] – if one is

caught in a trafﬁc jam, anger can easily set in, but anger in

this case has no apparent usefulness. How does the brain

maximize the beneﬁts of emotion yet minimize its occa-

sional inappropriateness? And how would the understand-

ing of such tradeoffs affect our ideas about robot designs?

Interest in the creation of robots with emotions is fourfold.

First, current technology already shows the value of pro-

viding robots with ‘emotional’ expressions (e.g. computer

tutors) and bodily postures (e.g. robot pets) to facilitate

human–computer interaction. Second, this raises the

question of the possible value (or inevitability) of future

robots not only simulating emotional expression but

actually ‘having emotions’. Third, this in turn requires

us to re-examine the neurobiology of emotion to generalize

concepts ﬁrst developed for humans and then extended to

animals so that the question of robot emotions becomes

meaningful. And fourth, this suggests in turn that build-

ing ‘emotional robots’ could also provide a novel test-bed

for theories of biological emotion.

This article samples the state of the art on current robot

technology, and examines recent work on the neurobiology

of emotions, to ground our suggestions for a scientiﬁc

framework in which to approach robot ‘emotions’. The

question of ‘emotional robots’ being used to test theories of

biological emotion is of great interest, but beyond the

scope of this article.

Different kinds of emotions

There is a wide spectrum of feelings, from the ‘motivation’

afforded by drives such as the search for food afforded by

hunger [1] to ‘emotions’ in which, at least in humans,

cognitive awareness might be linked to feeling the ‘heat’

of love, sorrow or anger, and so on. But as we have no

criterion for saying that a robot has ‘feelings’, we will seek

here to understand emotions in their functional context,

From ethology to robot motivation and emotion

We advance our discussion by reviewing work that has

added emotion-like features to robotic systems [6] , some

Corresponding authors: Michael A. Arbib (arbib@pollux.usc.edu),

Jean-Marc Fellous (fellous@duke.edu).

Available online 2 November 2004

Review

TRENDS in Cognitive Sciences Vol.8 No.12 December 2004

555

work inspired by ethology (the study of animal behavior),

and we then sample attempts to analyze the role of

emotion in general ‘cognitive architectures’ at the inter-

face between artiﬁcial intelligence and cognitive science.

(1) Reactive: a hard-wired releaser of ﬁxed action

patterns and an interrupt generator. This level has only

the most rudimentary affect.

(2) Routine: the locus of unconscious well-learned

automatized activity and primitive and unconscious

emotions.

(3) Reﬂective: the home of higher-order cognitive

functions, including metacognition, consciousness, self-

reﬂection, and full-ﬂedged emotions.

Ortony et al. address the design of emotions in

computational agents (these include ‘softbots’ as well as

embodied robots) that must perform unanticipated tasks

in unpredictable environments. They argue that such

agents, if they are to function effectively, must be endowed

with curiosity and expectations, and a ‘sense of self ’ that

reﬂects parameter settings that govern the agent’s

functioning.

Sloman [19] also offer a three-level view of central

processes:

(1) Reactive: producing immediate actions. When

inconsistent reactions are simultaneously activated one

may be selected by a competitive mechanism.

(2) Deliberative: using explicit hypothetical represen-

tations of alternative possible predictions or explanations,

comparing them and selecting a preferred option.

(3) Meta-management: allowing internal processes to

be monitored, categorized, evaluated and controlled or

modulated.

Sloman also notes the utility of an ‘alarm’ system – a

reactive component that gets inputs from and sends

outputs to all the other components and detects situations

where rapid global redirection of processing is required.

To reconcile the two schemes, we suggest using four

levels: reactive, routine, reﬂective–deliberative, and

reﬂective–meta-management.

Finally, recent work in multi-agent teamwork suggests

that in virtual organizations, agents that simulate

emotions would be more believable to humans and could

anticipate human needs by appropriate modeling of

others [15,20] .

Robot ethology

In the ‘ethological robots’ reviewed here, each drive,

perceptual releaser, motor-, and emotion-related process

is modeled as a different ‘specialist’ [7] or ‘schema’ [8] .

Each schema computes on the basis of its inputs to update

its ‘activation level’ and internal state, and sends on

appropriate outputs. A robot controller based on the

ethology of the preying mantis [9] provided four different

motivated visuomotor behaviors: prey acquisition, pre-

dator avoidance, mating and chantlitaxia (coined by

Rolando Lara, from the Nahuatl word chantli for ‘shelter’

and the Latin taxia for ‘attraction’). For prey acquisition,

hunger is the primary motivator; for predator avoidance,

fear serves similarly; for mating the sex drive dominates.

The behavioral controller was implemented on a small

hexapod robot.

In Bowlby’s [10] theory of attachment, infants view

certain individuals as sources of comfort, with the ‘comfort

zone’ depending on the circumstances. Likhachev and

Arkin [11] extrapolated these ideas to produce useful

behavior in autonomous robots, rather than to model the

human child. The ‘comfort zone’ ensures that the robot

does not stray from a given task or area; it can also provide

a basis for creating a robot pet that can relate to a

particular human being.

Studies of canine ethology support work on human–

robot bonding for Sony’s AIBO [12] , a speaking dog-like

entertainment robot. Ekman’s model [13] , with its six

basic emotional states, has been inﬂuential in work on

emotional expression in robots. The Kismet robot [14] , for

example, can communicate an emotive state and other

social cues to a human through facial expressions, gaze,

and quality of voice. The computations needed to commu-

nicate an ‘emotional state’ to a human might also improve

the way robots function in the human environment.

In order to interact with another agent, it is essential to

have a good conceptual model for how this agent operates

[15] . As complexity of environment and interactions

increases, the social sophistication of a robot interacting

with humans must be scaled accordingly. Some would

argue that this entails that the robot ‘has’ emotions, but

others would distinguish having a model of emotions of the

other agent from having emotions oneself. This in some

sense reverses the simulation theory of human empathy

[16,17] : in this theory, there is no question that the human

‘has’ emotion, and the proposal is that the system for

expressing one’s own emotions drives the ability to

recognize those of others.

The neurobiological roots of emotion

We turn now from outlining a functional view of robots

to review research linking brain and emotion [3,21–23] ,

before proposing a functional framework for synthesis. We

shall see that: (i) emotion is not computed by a centralized

neural system; (ii) emotions operate at many time scales

and at many behavioral levels; and (iii) there is no easy

separation between emotion and cognition.

Behavioral control columns

An animal comes with a set of basic ‘drives’ that provide

the ‘motor’ (motivation) for behavior. Most of the neural

circuitry underlying these drives involves speciﬁc nuclei of

the hypothalamus. Swanson [24] introduced the notion of

the ‘behavioral control column’ ( Figure 1 ), comprising

interconnected sets of nuclei in the hypothalamus under-

lying speciﬁc survival behaviors: spontaneous locomotion,

exploration, ingestive, defensive and reproductive behav-

iors. The hypothalamus sends this information to higher

centers such as the amygdala and the orbitofrontal cortex.

Cognitive architectures

We now turn to general ‘cognitive architectures’, in

which the role of emotion can be situated at several

levels. Ortony et al. [18] analyze the interplay of affect

(value), motivation (action tendencies), cognition

(meaning), and behavior at three levels of information

processing:

www.sciencedirect.com

Review

556

TRENDS in Cognitive Sciences Vol.8 No.12 December 2004

(a)

(b)

(c)

MPN

pro

AHN

suo

PVHd

VMH/

tub

PMv

PMd

MAM

mam

VTA

SNr

(-)

Figure 1. Major features of interactions of the behavioral control column and cerebral cortex in regulation of motivated behavior, as seen on ﬂatmaps of the rat central

nervous system. (a) The neuroendocrinemotor zone shown in black and three subgroups of hypothalamic nuclei: the periventricular region (PR) which contains visceromotor

pattern generators; the medial nuclei (MN); and the lateral zone (LZ). (b) Almost all nuclei in the behavioral control column generate a dual projection, descending to the

motor system and ascending to thalamocortical loops. (c) The embedding of the column in cortical computations. This prototypical circuit element consists of an excitatory

projection from cortex with a collateral to the striatum (the input system for the basal ganglia which play a key role in the sequencing and interleaving of actions). The striatum

then generates an inhibitory projection to the motor system with a collateral to the pallidum (the output system for the basal ganglia). Finally, the pallidum generates an

inhibitory projection to the brainstemmotor system, with a collateral to the dorsal thalamus (which projects back to cerebral cortex). This pallidal projection is disinhibitory

because it is inhibited by the striatal input. (Adapted from Swanson [24] , Figs 8,10,14, respectively, whose captions explain abbreviations not needed in this article).

Amygdala and cerebral cortex

The human ability to plan behaviors on the basis of future

possibilities rather than present contingencies alone has

been linked to the increased size and differentiation of

cerebral cortex [25,26] . Kelley stresses that feedforward

hypothalamic projections provide the motivational net-

work with access to associative and cognitive cortical

areas [27] . The amygdala can inﬂuence cortical areas via

feedback from proprioceptive, visceral or hormonal sig-

nals, via projections to various ‘arousal’ networks, and

through interaction with the medial prefrontal cortex [28]

( Figure 2 a). The prefrontal cortex, in turn, sends distinct

projections back to several regions of amygdala, allowing

elaborate cognitive functions to regulate the amygdala’s

roles in emotion. For example, our ability to create

heuristics and general rules from our everyday experi-

ences has been shown to depend on prefrontal cortices [29]

and their ability to bias activity in target structures [30] .

Because of the tight interactions between amygdala and

prefrontal cortex, it is likely that our ability to generalize

and abstract is directed by (and inﬂuences, in turn) some

aspects of our emotional state. How this is done, and how

robots could take advantage of it remains an open

question. Some functional connections between the amyg-

dala, thalamus and cortical areas allow for both fast

elicitation of emotion and more reﬁned emotion control

based on memory and high-level sensory representations

( Figure 2 a). The role of the amygdala in the experience

and expression of fear has received particular study

[21,31] . Stimuli that elicit fear reactions can be external

(e.g. a loud noise) or internal, from the behavioral control

columns or from memory structures such as hippocampus

or prefrontal cortex. Decision-making ability in emotional

situations is also impaired in humans with damage to the

medial prefrontal cortex, and abnormalities in prefrontal

cortex might predispose people to develop fear and anxiety

disorders [32] .

Activation of the human amygdala can be produced

by observing facial expressions, and lesions of the

human amygdala can cause difﬁculty in the identiﬁ-

cation of some such expressions [33,34] . The amygdala

and prefrontal cortices are therefore involved in social

as well as internal aspects of emotion, and together

play a crucial role in the regulation of emotion, a key

www.sciencedirect.com

Review

TRENDS in Cognitive Sciences Vol.8 No.12 December 2004

557

(a)

mPFC dlPFC

Working memory

Sensory cortex

Arousal

Amygdala

Thalamus

Basal forebrain

Brainstem

Locus coeruleus

Hypothalamus

Hippocampus

Bodily feedback

External

stimulus

Hormones

Proprioception

Behavior

(b)

From somatosensory cortex

Hypothalamus

Posterior inferior

temporal visual

cortex

Inferior temporal

visual cortex

Amygdala

TRENDS in Cognitive Sciences

Figure 2. (a) Interaction of the amygdala with cortical areas in the mammalian brain (adapted from [32] ). (b) Lateral view of part of the macaque monkey brain emphasizing

how the orbitofrontal cortex (involved in social emotions) links to amygdala, and to sensory cortices. V4 is visual area 4 (adapted from [3] ).

component of affective style, affective disorders [35] and

social interactions [36] .

Figure 2 b provides a view of how orbitofrontal cortex

links to the amygdala in macaque monkey. Damage to

monkey caudal orbitofrontal cortex produces emotional

changes that include the tendency to respond inappropri-

ately. Orbitofrontal neurons also serve as part of a

mechanism that evaluates whether a reward is expected,

and different subregions of the prefrontal cortex are

selectively involved during positive rewards or punish-

ments [37] .Dolcoset al. have shown that different

subregions of the medial temporal lobe memory system

are selectively and differentially involved for emotional

and neutral stimuli in human, and that this area was

strongly correlated with amygdala activations during

emotional stimuli [38] . Many other brain areas have

been involved in the experience and expression of

emotions in humans, including the anterior cingulate

cortex, insula, hippocampus and fusiform gyrus [39,40] .

The mirror system, language and empathy

Going beyond the hypothalamo-amygdala-cortical inter-

actions, we note that language plays a unique role in

shading and reﬁning human emotions. It is therefore

interesting that recent research suggests that ‘mirror

neurons’ might provide the substrate for both ‘empathy’ –

the ability to recognize the emotional dispositions of

others – and communication through language.

In monkey, parietal area AIP [41] processes visual

information concerning objects to extract possibilities

for manual interaction and is reciprocally connected

with the so-called ‘canonical neurons’ of ventral premotor

area F5 [42] whose discharge correlates with more-or-less

speciﬁc hand actions. Certain F5 neurons, called mirror

www.sciencedirect.com

Review

558

TRENDS in Cognitive Sciences Vol.8 No.12 December 2004

neurons [43,44] , discharge when the monkey observes the

experimenter make a gesture similar to one that, when

actively performed by the monkey, involves activity of

that neuron. PET experiments in humans showed that

superior temporal sulcus (STS), the inferior parietal

lobule, and the inferior frontal gyrus (area 45) in the left

hemisphere were signiﬁcantly activated for both grasp

execution and grasp recognition [45,46] . Area 45 is part of

Broca’s area, a major component of the human brain’s

language mechanisms. F5 is considered to be the monkey

homologue of Broca’s area.

These ﬁndings grounded the hypothesis that the mirror

system provided the basis for the evolution of human

language via intermediate stages involving ‘complex’

imitation (acquiring novel sequences of abstract actions

in a few trials), protosign (manually-based communi-

cation, enabled by freeing action from praxis to be used in

pantomime and then conventionalized communication)

and protospeech (vocally-based communication exploiting

the brain mechanisms that support protosign) [47,48] .

However, mirror neurons have also been implicated in

empathy – but with the emphasis now on recognizing

facial expressions instead of manual actions. Adolphs [49]

and Ochsner [36] stress the important role of social

interaction in the forming of emotions. Clearly, human

emotions are greatly inﬂuenced by our ability to

empathize with the behavior of other people [50] . Indeed,

some have suggested that mirror neurons can contribute

not only to ‘simulating’ other people’s actions as the basis

for imitation [51] , but also ‘simulating’ other people’s

feelings as the basis for empathy [16,17,52] .

Box 1. Three main neuromodulatory systems involved in

emotion

Dopamine

In the mammalian brain, dopamine appears to play a major role in

motor activation, appetitive motivation, reward processing and

cellular plasticity, and might be important in emotion. Dopamine is

contained in two main pathways that ascend from the midbrain to

innervate many cortical regions. Dopamine neurons in the monkey

have been observed to ﬁre to predicted rewards [67,68] . Moreover,

dopamine receptors are essential for the ability of prefrontal

networks to hold neural representations in memory and use them

to guide adaptive behavior. Therefore, dopamine plays essential

roles all the way from ‘basic’ motivational systems to working

memory systems essential for linking emotion, cognition and

consciousness.

Serotonin

Serotonin has been implicated in behavioral state regulation and

arousal, motor pattern generation, sleep, learning and plasticity,

food intake, mood and social behavior [69] . The cell bodies of

serotonergic systems are found in midbrain and pontine regions in

the mammalian brain and have extensive descending and ascending

projections. Serotonin plays a crucial role in the modulation of

aggression and in agonistic social interactions in many animals. In

crustaceans, serotonin plays a speciﬁc role in social status and

aggression; in primates, with the system’s expansive development

and innervation of the cerebral cortex, serotonin has come to play a

much broader role in cognitive and emotional regulation, particu-

larly control of negative mood or affect. The serotonin system is the

target of many widely used anti-depressant drugs.

Opioids

The opioids, which include endorphins, enkephalins and dynor-

phins, are found particularly within regions involved in emotional

regulation, responses to pain and stress, endocrine regulation and

food intake. Increased opioid function is associated with positive

affective states such as relief of pain, and feelings of euphoria, well-

being or relaxation. Activation of opioid receptors promotes

maternal behavior in mothers and attachment behavior and social

play in juveniles. Separation distress, exhibited by archetypal

behaviors and calls in most mammals and birds, is reduced by

opiate agonists and increased by opiate antagonists in many species

[70] . Opiates can also reduce or eliminate the physical sensation

induced by a painful stimulus, as well as the negative emotional

state it induces. Opioids and dopamine receptors are two major

systems affected by common drugs of abuse.

Neuromodulation

We now switch structural levels, turning from speciﬁc

brain structures to systems of neuromodulation. Neuro-

modulation refers to the action on nerve cells of endo-

genous substances called neuromodulators. These are

released by a few specialized brain nuclei that have

somewhat diffuse projections throughout the brain and

receive inputs from brain areas that are involved at all

levels of behavior from reﬂexes to cognition. Each

neuromodulator typically activates speciﬁc families of

receptors in neuronal membranes. The receipt of its own

neuromodulator by a receptor has very speciﬁc effects on

the neuron at various time scales, from a few milliseconds

to minutes and hours [53] . Each neuron has its own

mixture of receptors, depending on where it is located

in the brain.

Kelley’s [27] analysis of motivation and emotion

emphasizes three neuromodulatory systems – those for

dopamine, serotonin and opioids ( Box 1 ). Strikingly,

although these three neuromodulatory systems seem to

be distinct from each other in their overall functionalities,

they each exhibit immense diversity and synergism of

behavioral consequences. The different effects depend on

both molecular details (the receptors) and global arrange-

ments (the circuitry within the modulated brain region,

and the connections of that region within the brain).

Neuromodulation thus provides a simple but high-

impact signal that can fundamentally change the way

single neurons and synapses ‘compute’ and in this sense is

akin to the alarm system of Sloman [19] . Earlier, we said

that reconciling the cognitive architectures of Ortony et al.

and Sloman led us to consider four architectural levels.

Fellous [54] has also produced a four-level hierarchy, but

this time on the basis of a review of data on hypothalamus,

amygdala and orbitofrontal cortex, and the suggestion

that the neural basis for emotion involves both compu-

tations in such structures and their current state of

neuromodulation (see Figure 3 , and [55,56] ). Others have

suggested that neuromodulation might be a key to meta-

learning [57] .

Towards a functional view of emotions

Emotions are, of course, far more complex than a few brain

structures and three ascending modulatory systems that

interact with them. We can only outline the lessons that

neurobiology offers for the study of robotic emotions, not

provide all the details. However, we stress that the

richness of human emotion is in part due to the linkage

www.sciencedirect.com

Plik z chomika:

Inne pliki z tego folderu:

Inne foldery tego chomika: