Content uploaded by Gergely György
Author content
All content in this area was uploaded by Gergely György on Jan 13, 2016
Content may be subject to copyright.
What should a robot learn from an infant?
Mechanisms of action interpretation and
observational learning in infancy
György Gergely
Institute for Psychology Hungarian Academy of Sciences
Victor Hugo u. 18-22
1132 BUDAPEST
gergelyg@mtapi.hu
Abstract
The paper provides a summary of our
recent research on preverbal infants (using
violation-of-expectation and observational
learning paradigms) demonstrating that one-
year-olds interpret and draw systematic
inferences about other’s goal-directed actions,
and can rely on such inferences when imitating
other’s actions or emulating their goals. To
account for these findings it is proposed that one-
year-olds apply a non-mentalistic action
interpretational system, the ’teleological stance’
that represents actions by relating relevant
aspects of reality (action, goal-state, and
situational constraints) through the principle of
rational action, which assumes that actions
function to realize goal-states by the most
efficient means available in the actor’s situation.
The relevance of these research findings and the
proposed theoretical model for how to realize the
goal of epigenetic robotics of building a ’socially
relevant’ humanoid robot is discussed.
1. Introduction: Can infancy research
inform robotics about how to
interpret and learn from intentional
actions of other agents?
Having been asked to talk about my infancy research
at this workshop on Epigenetic Robotics, I feel I must
start by making it clear that I can by no means be
considered to be an artificial intelligence researcher. In
fact, I am just an “old fashioned” cognitive
developmentalist using experimental behavioral
techniques such as “violation-of-expectation” or
imitative learning paradigms to study the interpretative
mechanisms and inferential capacities of infants’ early
understanding of intentional actions, and their ability to
learn novel means through observation and imitation of
the goal-directed actions of others. While in my work I
have been developing abstract conceptual models
intended to capture the remarkable complexities of
these earl y competences, I have never written a
connectionist program that would simulate these
abilities, not to speak about actually building a robot
that would do so.1
Given this background, you can imagine that I was
slightly (though admittedly, pleasantly) surprised (but
also somewhat incredulous at first) when I received the
invitation to give a keynote address to the 3rd
International Conference of Epigenetic Robotics. After
some checking to make certain that the invitation wasn’t
sent by chance to the wrong e-mail address, my initial
(and admittedly somewhat narcissistic) joyful reaction
soon turned into a slight but distinctly unpleasant
feeling of anxiety: Is there really anything instructive or
relevant that our own research programme can offer to
experts in AI and AL who are trying to build humanoid
robots capable of acting in a goal-directed manner and
of learning new ways of acting so by imitating other
agents? So I did a little bit of home-work to see what
kind of questions are being currently pursued by
researchers in epigenetic robotics and how these
questions flare with the type of problems that our
developmental and experimental approach focuses on. I
must admit that I was pleasantly surprised and
intellectually quite intrigued to find an unexpected
amount of significant convergence, but I also noted
some important differences in the dominant focus of
current research that characterizes the two approaches.
Briefly, my impression is that recent research in
epigenetic robotics has been strongly preoccupied with
and made significant advances towards modeling the
“lower level” mechanisms involved in action perception
and production and the ways in which these
mechanisms maybe inherently related. The basic issues
1 I did engage though in challenging arguments with ‘real’ AI
researchers who attempted to replace our conceptual models
with specific connectionist simulations that, they claimed,
were able to account for our results without making use of our
model’s abstract constructs of folk psychology such as ‘goals’
and ‘principles of rationality or efficiency’.
addressed include questions about how simple and more
complex motor actions can be generated, what is the
nature and role of “motor primitives” and forward
models in action planning and control, or how to design
engineering solutions to the thorny “correspondence
problem” of establishing the perceptual-motor mapping
or equivalence between the same or similar actions
when they are observed and when they are being
executed, to name just a few of the focal research
questions of current-day robotics (e. g., Breazeal and
Scassellati, 2002; Dautenhahn and Nehaniv, 2002;
Schaal, 1999).
In these areas there has clearly been a significant
amount of healthy interdisciplinary cross-fertilization
between AI and robotics, and important advances in
recent computational models (e. g., forward models) of
intentional action planning and control (e.g., Wolpert et
al., 2001), and the new discoveries made in cognitive
neuroscience such as the identification of specialized
neurons in the superior temporal sulcus (ST) of
macaques that are sensitive to highly specific actions of
body parts (e.g., Perrett et al., 1985) or of mirror
neurons in the F5 region of the macaque brain that are
sensitive to both the perception and the motor execution
of specific goal-directed object manipulative actions (e.
g., Rizolatti et al., 1988).
In contrast, when it comes to our own research on
infants’ understanding of intentional actions, the
questions addressed are at a qualitatively different level,
and the types of, admittedly extremely hard and
exciting, research issues listed above concerning the
basic mechanisms that mediate action perception,
generation, and matching, are not directly dealt with or
investigated, rather, they are mostly simply presupposed
by our approach. Instead, our research programme has
focused on “higher-order” cognitive processes in an
attempt to characterize the representational and
inferential mechanisms and built-in “top-down”
architectural constraints of the domain-specific action
interpretational system that, we believe, underlies
young infants’ demonstrated (and, as I hope to convince
you, truly remarkable) early capacity to identify, infer,
and attribute goals to observed actions of others, and
their ability to infer which (or which aspects of)
perceived intentional actions of other agents they should
imitate during observational learning. It is for this
reason (of qualitatively different levels of investigation)
that I first felt rather uncertain as to whether and to what
degree our demonstrations and models may prove to be
informative to the central concerns of researchers in
epigenetic robotics.
Eventually, however, I have become cautiously
confident that the time may, in fact, have come for our
infancy work on these “higher-level” cognitive
constraints and interpretive mechanisms to provide
useful information and possibly even to suggest new
directions for future research for epigenetic robotics.
This more optimistic diagnosis certainly seems
warranted by such reoccurring programmatic statements
that have recently been made by leading robotics
researchers about the anticipated new directions and
questions that they feel their research domain needs to
turn to and tackle in the near future to make significant
further progress possible. Just to give you a few
examples: Cynthia Breazeal and Brian Scassellati
(2002) in a recent review in Trends in Cognitive
Sciences raise as their focal issue the question: “How
does a robot know what to imitate?”. Furthermore, in
their “Questions for future research” they ask: “Just as
children develop the ability to imitate the goal of an
action rather than a specific act, can we construct robots
that are capable of making this inference? Today’s
robots – they say - respond only to the observable
behavior without any understanding of the intent of an
action.” (p. 486). Then they go on to ask: “Who should
the robot learn from, and when is imitative learning
appropriate? Robots that imitate humans today are
programmed to imitate any human within view” (p.
486). In a similar vein, one of the “Outstanding
question” of Stefan Schaal’s (1999) excellent recent
review in Trends in Cognitive Sciences is this:
“Understanding task goals: How can the intention of a
demonstrated movement be recognized and converted to
the imitator’s goal?” (p. 240). These new questions
clearly signal a growing need within epigenetic robotics
to move towards the “higher order” cognitive issues that
are at the very center of our own research inquiries
about infants’ interpretative capacities within the
domain of action understanding.
Therefore, I hope that by summarizing our major
empirical findings of preverbal infants’ capacity to
interpret, reason about, and learn from the observed
intentional actions of other agents (Csibra et al., 1999;
2003; Gergely et al., 1995; 2001, 2002) and by outlining
our theoretical account of this early competence that we
call the one-year-old’s ‘naïve theory of rational action’
or the ‘teleological stance’ (Gergely and Csibra, 1997,
2003; Csibra and Gergely, 1998), I’ll be able to make a
useful contribution to the newly forming
interdisciplinary dialogue between cognitive infancy
research and epigenetic robotics.
2. What one-year-old infants
understand about the goal-directed
actions of other agents.
Before presenting supporting empirical evidence, let
me first provide you with a brief list of those capacities
that our research has demonstrated in one-year-old
infants in the domain of action interpretation and
observational learning and that seems to me to coincide
with (or even go beyond) the new and ambitious
research targets that the epigenetic robotic movement
has set for itself. In other words, I suggest that these are
the basic competencies that one-year-old human infants
possess in this domain and that humanoid robots (or, if
not them, then at least their designers) should learn from
them:
(1) ATTRIBUTING GOALS TO ACTIONS: One-
year-olds can interpret other agents’ actions as goal-
directed;
(2) EVALUATING THE RATIONALITY OR
EFFICIENCY OF ACTIONS AS MEANS TO GOALS:
One-year-olds can compare and evaluate which of the
alternative actions available to an agent within the
physical constraints of a given situation is the most
efficient means to the goal;
(3) PREDICTING NOVEL MEANS ACTIONS IN
NEW SITUATIONS: By one year of age infants can
form an active expectation that in order to realize her
goal an agent ‘ought to’ perform the most rational or
efficient means action available to her within the
particular situation;
(4) INFERRING (NON-VISIBLE) ASPECTS OF
GOAL-DIRECTED ACTIONS: Going “beyond the
information given” – as Bruner famously said in 1957
(Bruner, 1957) - , one-year-olds can draw systematic
and productive inferences to identify and
(representationally “fill in”) any one of the three basic
aspects (Goal, Means action, and relevant Situational
Constraints, see Figure 1) of the representation of an
intentional action when that aspect is perceptually
inaccessible to them, as long as they have direct access
to the other two relevant aspects of the goal-directed
action, on which they can base such an inference. In
particular:
(4a) INFERRING A (NEW) MEANS ACTION: Given
perceptual information about the Goal and the
(changed) Situational Constraints, infants can infer (and
predict) what novel Means action the agent ‘ought to’
perform to achieve its goal in the most efficient manner
given the changed constraints of the situation;
(4b): INFERRING AN (UNSEEN) GOAL: Given
perceptual information about the Situational Constrains
and about the initial part of an Action (whose end-state
is occluded from them), they can infer an (unseen) Goal
that would justify the action as a rational or efficient
means to the goal in the given situation; and
(4c): INFERRING (UNSEEN) SITUATIONAL
CONSTRAINTS: Given perceptual information about
the Action and about the Goal state, infants can infer a
(non-visible) physical Situational Constraint (such as an
obstacle that is occluded from their view) whose
presence would justify the action as a rational or
efficient means to the goal.
(5): TELEOLOGICAL EMULATION OF NEW
GOALS AND RATIONAL IMITATION OF NOVEL
MEANS IN OBSERVATIONAL LEARNING: When
learning new actions to achieve a novel goal from
observing an unfamiliar means action demonstrated by
another agent, 14-month-olds can evaluate the
rationality or efficiency of the observed means both in
relation to the situational constraints of the model and
in relation of their own situational constraints, and can
use this information to decide whether to imitate the
demonstrated novel means or to achieve the new goal
through emulation.
3. The one-year-old’s ‘teleological
stance’ and the inferential ‘principle
of rational action’.
What makes these remarkable inferential feats
possible for one-year-old infants who are, arguably, still
lacking the metarepresentational means to attribute
abstract and invisible causal intentional mental states
(such as intentions, desires, and beliefs) to the agent’s
mind? (For contrary views that assume the early
availability of at least some mentalistic representational
aspects of a theory of mind by the end of the first year,
see e.g., Tomasello, (1999); Kelemen, (1999).) To
answer this question, Gerg Csibra and myself have
proposed (Gergely and Csibra, 1997; 2003; Csibra and
Gergely, 1998) that one-year-olds possess a non-
mentalistic (reality-based) teleological action
interpretational system or strategy, that we call the
‘teleological stance’ (Figure 1) that establishes a
teleological (rather than causal) explanatory relation
among three relevant aspects of (current and future)
reality: the observed behavior, a future state of realit y
(future in relation to the behavior), and the relevant
aspects of physical reality that constrain possible
actions in the particular situation in which the observed
behavior
Figure 1: Teleological representation of goal-directed
actions
unfolds. This action interpretational schema provides a
well-formed (and thus acceptable) teleological
representation of the observed behavior as an efficient
goal-directed action only if the behavior can be
evaluated as an effective (rational) way to bring about
the future state given the physical constraints of the
situation. If this well-formedness condition (that is
articulated by the ‘principle of rational action’, see
below) is satisfied by the representation in question, the
future state will become encoded as the Goal, the
behavior as a Means to the goal, and the relevant
aspects of physical reality as the Situational Constraints
on action (Figure 1).
We propose that such teleological action
interpretations are driven by the core ‘principle of
rational action’ that captures our normative
assumptions about the essentially functional nature of
intentional actions (see Dennett, 1987; Gergely and
Csibra, 2003). The rationality principle serves both as a
criterion of well-formedness for teleological action
interpretations and as an inferential principle guiding
and constraining the construction of such action
interpretations. In particular, the principle of rational
action presupposes that a) actions function to bring
about future goal states, and b) goal states are realized
by the most rational (or efficient) action available to the
actor within the constraints of the situation. Thus, the
principle asserts that a teleological action explanation is
well-formed (and therefore acceptable) if, and only if,
the action realizes the goal state in a rational (or
efficient) manner within the particular situational
constraints (Figure 1).
4. Empirical evidence supporting the
inferential productivity of the
rationality principle in teleological
action interpretations of one-year-
olds.
Early understanding of goal-directed actions has been
demonstrated using a variety of paradigms such as
imitation (Carpenter et al., 1998; Gergely et al., 2001,
2002; Meltzoff, 1988; 1995), joint attention (Carpenter
et al., 1998; Tomasello, 1999), and violation-of-
expectation looking time studies (Csibra et al., 1999,
2003; Gergely et al., 1995; Király et al., 2003;
Woodward, 1998; Woodward and Sommerville, 2000).
Let me illustrate the complex nature of this
understanding by one of our violation-of expectation
studies (Gergely et al., 1995) (Figure 2).
Figure. 2. Experimental and Control conditions
of Gergely et al., (1995)
Twelve-month-olds were habituated to a computer-
animated goal-directed action (Figure 2A) in which a
small circle repeatedly approached and contacted a
large circle (goal) by jumping over (means act) an
obstacle separating them (situational constraint). Even
though these 2D shapes had no human-like features,
when we asked adults to describe what they see, they
immediately interpreted this visual event as depicting an
efficient means action to achieve a goal state (that of
contacting the large circle), because they could justify
the jumping approach as the most rational action
available to realize that goal given the physical
constraints of the situation (i. e., the presence of the
‘obstacle’ separating the two circles) (cf. Heider and
Simmel, 1944). To test whether or not one-year-olds
would also interpret this event in the same manner, we
presented them, following habituation, with two types
of test events (with their order of presentation being
randomized across subjects) in which the obstacle was
no longer present. In one of the test displays (Figure
2C) they saw again the already familiar ‘jumping
approach’, which, however, could no longer be justified
(by the presence of an obstacle) as a rational action to
achieve the goal (the small circle jumped over empty
space during its approach of the large circle). In
contrast, in the other test event (Figure 2D) infants were
presented with a perceptually novel, but sensible
‘straight-line goal-approach’ (that has become an
available action alternative to get to the goal after the
removal of the obstacle).
In spite of the fact that the old ‘jumping approach’
(2C) was perceptually similar to the means action
presented during the habituation event (2A), subjects
looked at it significantly longer (indicating violation-of-
expectation) than at the novel ‘straight-line approach’
(2D) to which (even though it was perceptually novel)
they showed no dishabituation at all. This suggests that
the infants found the old ‘jumping action’ (2C) test
event unexpected, because it seemed to them an
inefficient way to reach the goal in the new situation
where there was no obstacle to justify the jumping
action as a rational means. In contrast, the fact that they
did not dishabituate to the novel ‘straight-line goal-
approach’ (2D) indicates that this action was expectable
for them (in spite of its perceptual novelty) as it
appeared to be the most efficient means to the goal that
has become available after the disappearance of the
obstacle.
Appropriate control conditions (see Csibra et al.,
1999; Gergely et al., 1995) ruled out obvious alternative
explanations. In the control study, during the
habituation phase the rectangular object appeared
behind the small circle (Figure 2B) and so it did not
form an obstacle towards the goal object (making the
more efficient straight-line goal-approach available
already during the habituation event). In spite of this,
the small circle approached the large circle through the
same jumping action as in the experimental condition.
Note, however, that in the Control condition this
behavior could not be represented in a well-formed
teleological representation as a goal-directed action, as
there was an obviously more rational alternative means
to the goal available, but not realized (the straight-line
approach). Therefore, the infants could not generate any
specific expectation about what type of goal-approach
the small circle would follow in a changed situation. As
a result, when the very same two test events that were
shown in the experimental condition (see Figures 2C
and 2D) were presented to the infants in the control
study, the differential looking pattern found in the
experimental condition has disappeared: the subjects
looked equally at the old ‘jumping approach’ and the
new ‘straight-line approach’ test events.
These results indicate that by 12 months infants can
(a) interpret an other agent’s action as goal-directed, (b)
evaluate which one of the alternative actions available
within the constraints of the situation is the most
efficient means to the goal and (c) expect the agent to
perform the most efficient means available in the given
situation to realize the goal.
Above I argued that by applying the inferential
principle of rational action one-year-olds can “go
beyond the information perceptually given” and can
productively infer any one of the three representational
aspects of the teleological representation of a goal-
directed action (Means act, Goal state, or relevant
Situational constraints) if that aspect is perceptually
inaccessible to them, as long as they have direct
perceptual information about the contents of the other
two representational elements. To demonstrate this
property of inferential systematicity and productivity of
the rationality principle, we habituated 12-month-old
infants to computer-animated goal-directed actions in
three types of situations (Gergely et al., 1995; Csibra et
al., 1999; Csibra et al., 2003) (Figure 3). The different
event displays were designed so that in each case one of
the three basic elements necessary for a well-formed
teleological action interpretation was made visually
inaccessible. To interpret the action as an efficient and
justifiable goal-approach, the infants had to use the
rationality principle to infer and “fill in” the content of
the relevant missing element of the representation.
Figure 3: Three types of inference drawn from the
teleological stance
Figure 3A (which depicts the Gergely et al. (1995)
study that was discussed above) exemplifies the first
type of teleological inference where infants have to
infer the particular means action that is congruent with
(i.e., can be seen as an efficient goal-approach in
relation to) the visually specified goal state and
situational constraints. As described above, the finding
that infants looked significantly longer at the
incongruent test display (old jumping approach) than at
the congruent one (novel straight-line goal-approach) is
evidence that they could draw the type of inference in
question.
Figure 3B illustrates the second type of teleological
inference where the infants had to infer a (non-visible)
goal state to rationalize the incomplete action whose
end state was occluded from them, as an efficient
‘chasing’ action (Csibra et al., 2003). During
habituation a large ball was approaching a moving small
ball until the latter passed through a small aperture
between two obstacles and left the screen. The large
ball, being too big to get through the aperture, had to
make a detour around the obstacles before it also
disappeared from view. In the two test events the upper
part of the screen was opened up revealing one of two
different end states: one congruent with the inferred
goal state of an efficient ‘chasing’ action (the small
circle stopped, at which point the large circle changed
its course so that it ‘caught up with’ the small circle and
contacted it), and one that was incongruent with the
inferred goal (when the small circle stopped, the large
one, without modifying its direction, passed by it
leaving the screen without ever ‘catching’ the small
circle). Twelve-month-olds looked significantly longer
at the incongruent than at the congruent test display,
suggesting that the incongruent outcome violated their
expectation about the goal state that they had inferred to
rationalize the incomplete action as an efficient
‘chasing’ event (for appropriate controls, see Csibra et
al., 2003).
Finally, Figure 3C provides an example of the third
kind of teleological inference to specify the particular
situational constraints (occluded from view by a
screen) in order to rationalize the small circle’s visible
action (jumping approach) as an efficient means to
realize the visible goal state (contacting the large circle)
(Csibra et al., 2003). In the two test displays the screen
was lifted either revealing an obstacle whose presence
justified the jumping approach (congruent display) or
revealing no such obstacle (incongruent display).
Twelve-month-olds again looked significantly longer at
the incongruent than at the congruent display, indicating
that they inferred the presence of the occluded obstacle
to justify the jumping approach as an efficient means to
the goal (again, for appropriate controls see the original
study reported in Csibra et al., 2003).
In sum, these results provide converging evidence
indicating that by 12 months infants can take the
teleological stance to interpret actions as means to
goals, can evaluate the relative efficiency of means by
applying the principle of rational action, and can
generate systematic inferences to identify relevant
aspects of the situation to justify the action as an
efficient means even when these aspects are not directly
visible to them.
5. Beyond the shortest pathway: The
generality of the rationality principle.
At this point I anticipate strong resistance (see e.g.,
the controversy between Premack and Premack (1997)
and Gergely and Csibra (1997)) and at least one specific
objection against our – maybe, at first, radically
sounding - theoretical proposal that one-year-olds, who
may still lack the mentalistic competence to infer,
represent and attribute intentional mental states (such as
beliefs and desires) to other agents, nevertheless,
already possess and apply inferentially productively the
general and abstract principle of rationality that
philosophers consider to be the central inferential
component of mature theory of mind (Dennett, 1987;
Fodor, 1987, 1992). The concrete objection is a
straightforward one (communicated to me first by Paul
Harris, pers. com.): looking at our habituation studies
summarized above, one could justifiably suggest that in
each case the action that, according to our theory,
infants evaluate as the most rational means available to
the goal, in fact, always coincides with the shortest
approach route to the target object. It may be objected,
therefore, that instead of relying on the general principle
of rationality, infants may apply a simpler and more
specific criterion of expecting the actor to always take
the shortest pathway available to reach the target
location.
In contrast, if infants, similarly to adults, employ the
more general principle of rational action, they should be
able to apply other kinds of criteria as well that could,
under some circumstances, override the ‘shortest
pathway’ criterion when interpreting behavior as an
efficient goal-directed action. To demonstrate that this
is, indeed, the case, we designed a study (Csibra et al.,
1998) that pitted against each other ‘shortest pathway’
versus ‘least effort’ to see if the latter could also be
applied as one of the criteria for evaluating the
rationality of a goal-approach.
In a violation-of-expectation paradigm we habituated
two groups of 12-month-olds to one or the other of two
different versions of a 2D computer-animated event in
which a rectangle approached a circle performing a
worm-like motion pattern (see Figures 4A and 4B). The
rectangle passed through a gap on a wall that separated
it
Figure 4A: The „Squeeze” study: Experimental group
from the stationary circle on the other side. The gap was
either wide, allowing for an ‘effortless goal-approach’,
or it was narrow, requiring the rectangle to squeeze
through it exhibiting effortful movements (‘effortful
goal-approach’). The gap was positioned in such a way
that passing through it corresponded to the shortest,
straight-line pathway in between the rectangle and the
circle. The two types of displays were randomly varied
during habituation.
In the habituation events presented to the
Experimental group (Figure 4A) the rectangle didn’t
have a choice of alternative routes to get to the circle on
the other side of the wall: the only pathway it could take
was through the one single gap on the wall. In contrast,
in the habituation events presented to the Control group,
the upper part of the wall was absent and so an
alternative route to the goal
Figure 4B: The „Squeeze” study: Control group
(apart from the pathway through the gap) was also
available which, though longer and involving a spatial
detour, would not have made it necessary for the
rectangle to engage in effortful squeezing (Figure 4B).
In spite of this, similarly to the Experimental condition,
the rectangle in the Control condition always took the
‘shortest pathway’ even when it had to squeeze through
the narrow gap (‘effortful goal-approach’).
After habituation, the Experimental and Control
groups were presented with the same two test events
(Figure 5). The rectangle was again separated from the
circle by a wall that had two gaps in it this time. The
narrow gap, similarly to the habituation events, allowed
for the shortest approach route to the goal, but it
required effortful squeezing to get through (‘shortest
pathway/more effort’). In contrast, the position of the
new (but wider) gap that was opened in the wall some
distance below the
Figure 5: The “Squeeze” study: Looking times for the two
types of test events in the Experimental vs. the Control
conditions
narrow gap required a longer approach route involving a
spatial detour, however, without making effortful
squeezing necessary to get through it (‘longer
pathway/less effort’). Both test events were presented to
each subject with their order of presentation being
randomized across subjects. In the ‘shortest
pathway/more effort’ test event the rectangle
approached the target circle through the narrow gap,
which allowed for the shortest approach route to the
goal, but required effortful squeezing to get through it.
In the ‘longer pathway/less effort’ test event the
rectangle approached the target object through the wider
gap that required a longer route through a spatial detour
to get to the goal, but without the need to engage in
effortful squeezing. The duration of the two test events
was exactly the same (7.5 sec).
The results (see Figure 5) showed that subjects in the
Experimental group looked longer at the ‘shortest
pathway/more effort’ test event than the ‘longer
pathway/less effort’ test event, while the Control group
exhibited the opposite looking pattern. This crossover
was significant as evidenced by a significant Condition
X Test event interaction in a two-way ANOVA (F1,
38=8.35, P<.01) of looking times. Non-parametric
statistics also confirmed this result (see Csibra et al.,
1998).
We can, therefore, conclude that one-year-olds do not
always expect an agent to approach its goal through the
shortest path available. This suggests that the simpler
‘shortest pathway’ criterion is not a viable alternative to
the more general rationality principle as the basis for
judging what the most efficient goal-approach is within
the constraints of a given situation. As the results for the
Experimental group suggest, subjects expected the
agent to take the longer pathway that required less
effort when such an alternative to the goal became
available during the test event. This indicates that one-
year-olds are not restricted to the single criterion of
expecting the ‘shortest pathway’ to the goal when
evaluating the rationality of a goal-directed means
action: in fact, they can clearly rely on other criteria as
well (in particular, the criterion of ‘least effort’) when
making such a judgment.
Finally, the results of the Control group, where
already during the habituation event the agent did not
follow the available alternative route to the goal that –
at least, under some criteria such as ‘least effort’ – may
have seemed more rational to the infant, seem to allow
for two alternative interpretations. First, it is possible
that the one-year-olds inferred and attributed a specific
disposition to the agent (to always take the shortest
path, or to squeeze whenever possible), and so they
expected her to act according to this disposition even
under the changed situational constraints of the test
events. Second, it seems also possible that the infants
reasoned that there may have been some further
condition or aspect of the habituation situation (that
they did not notice or were ignorant about) that must
have justified the agent’s choice to take the shortest
pathway even though it apparently required more effort.
Therefore, on this ground they may have simply
assumed the agent’s going through the shortest pathway
must have been rational after all, and so they expected
her to take the same approach route (that they have
come to consider to be rational) in the changed situation
of the test events as well.
To sum up: the results of the squeezing study clearly
indicates that the principle of rationality that the one-
year-olds rely on when evaluating the relative efficiency
of alternative means to a goal is a general principle that
allows for the application of multiple criteria and cannot
be reduced to a single and more simple spatial criterion
of always preferring the ‘shortest path’ to the goal.
6. Imitative learning of novel means
in infants and robots: The problem of
selective imitation.
Up till now I have only provided evidence for the
teleological stance and its core inferential principle of
rational action as a mechanism specialized for
interpreting the goal-directed actions of other agents as
those are perceived by the infant. Clearly, however, one
of the most significant evolutionary advantage that the
ability to interpret other’s actions as goal-directed
provides for humans has to do with the vital functional
role it plays in the social transmission of culturally
relevant new goals and new ways of acting to efficiently
achieve such goals from observing other agents’ novel
intentional actions. The – possibly human-specific
(Tomasello et al., 1993; Tomasello, 1999) and innate
(Meltzoff and Moore, 1977, 1989) – ability to imitate
human actions has been proposed by many as the basic
mechanism that makes observational social learning of
novel means from the action demonstrations of other
human agents possible for our species. No wonder that
one of the most cherished ambitions of epigenetic
robotics has become to equip humanoid robots with the
basic competence to imitate the observed actions of
other agents (be it humans or other robots) and to use
this ability to imitate for acquiring novel goal-directed
actions (see Breazeal and Scassellati, 2002; Dautenhahn
and Nehaniv, 2002; Schaal, 1999). Naturally, one of the
major engineering hurdle towards achieving this aim
was (and still is) to find efficient and generative
solutions to the “correspondence problem” of mapping
perceived movements of others onto the robot’s
corresponding motor programs whose execution
produces equivalent actions either by pre-wiring such a
mapping or by designing learning solutions employing
different sophisticated versions of supervised learning
methods, forward modeling, “motor primitives” and
connectionist learning nets (see Schaal, 1999, for a
review). There has been clear progress in this area that
was reinforced and informed by the recent discoveries
of biological analogue mechanisms in the nervous
system such as the mirror neurons (e.g., Fadiga et al.,
1995; Rizolatti et al., 1996) and by psychological
models such as Meltzoff and Moore’s (1997) ‘Active
Intermodal Mapping’ mechanism to account for the
phenomenon of neonatal imitation (Meltzoff and
Moore, 1977, 1989).
However, it should be realized that no matter how far
we advance in discovering and understanding the neural
mechanisms that mediate the perceptual-motor mapping
of actions or in finding engineering solutions to equip
robots with analogous mapping mechanisms, such
progress will at best provide us with some of the
necessary, but never the sufficient conditions to fully
understand or model the functionally more significant
aspects of the human competence for imitative and
observational learning. This is so because by
exclusively relying on such automatic mechanisms that
allow for action imitation, we would remain stuck
forever with the pervasive problem of how to avoid
indiscriminate and automatic imitation of anything (or
at least any human or robot) that moves. Adults (and, as
we shall see, even 14-month-old infants) are rather
selective in what human action they imitate and under
what conditions they do so. In fact, automatically
imitating every human action that one is perceptually
exposed to is a seriously dysfunctional pathological
condition observable in patients with prefrontal lesions
who cannot inhibit the tendency to compulsively imitate
gestures or even complex actions performed in front of
them by an experimenter (Lhermitte et al., 1986).
Clearly, what is needed is an additional account of the
inferential capacities that constrain and guide the
imitative mechanism to be functionally selective, a most
significant problem that Breazeal and Scassellati (2002)
have clearly put their fingers on when they raised as
outstanding future problems for epigenetic robotics such
questions as: “How does a robot know what to imitate?”
or “Just as children develop the ability to imitate the
goal of an action rather than a specific act, can we
construct robots that are capable of making this
inference? Today’s robots respond only to the
observable behavior without any understanding of the
intent of an action.” (p. 486).
In fact, until quite recently (see Bekkering et al.,
2000; Gergely et al., 2001, 2002) the problem of
selective imitation has not been fully recognized in
developmental approaches to imitative learning either.
Let us take as an example one of the ingenious and
highly influential imitation studies by Meltzoff (1988,
1995) that has demonstrated that infants as young as 14
months of age can learn a novel means by imitation
from observing an adult model’s demonstration. The
infants observed the model illuminate a box by leaning
forward from waist and touching its top panel with her
forehead. After a week, 67% of infants re-enacted this
novel ‘head-action’, while no infant performed it
spontaneously in a base-line control group that had not
seen the action demonstrated. According to Meltzoff’s
(1995) own interpretation „infants do more than retrieve
general goal or end state information („the panel can be
lit”), which would not necessarily mandate use of the
head [emphasis added]. They can remember the specific
way something was done; they imitate the means used,
not solely the general ends achieved.” (p. 509).
Tomasello (1999) proposed that such imitative
learning is human-specific as primates have been
shown not to be able to imitatively copy specific novel
means acts demonstrated to them. Instead, apes try to
bring about the observed new outcome by performing
motor actions already in their repertoire in a ‘trial-and-
error’ manner (that actually often leads to the eventual
(re)discovery of the demonstrated means or some other
functional action with which they succeed in achieving
the observed outcome). Tomasello has named this kind
of observational learning “emulation” to distinguish it
from true “imitative learning” that involves the faithful
and automatic copying of the observed means. He
argued that if infants used emulation in the Meltzoff
study, one would have expected them to simply touch
the box with their hand, instead of imitating the
unfamiliar ‘head-action’. Meltzoff (1988, 1995),
however, did not report such ‘hand-actions’.
Tomasello (1999) argued that „imitative learning of
this type thus relies fundamentally on infants’ tendency
to identify with adults...” (p. 82) who are perceived by
them (through Meltzoff and Moore’s (1997) proposed
innate mechanism of ‘Active Intermodal Mapping) as
“just-like-me”. Tomasello further proposed that
identification is a human-specific innate capacity that is
lacking in apes as shown by the fact that apes can only
emulate rather than being able to engage in imitative
copying of observed means actions.
In sum: currently dominant models of imitative
learning of novel means in developmental psychology
(represented by the work of such influential researchers
as Meltzoff (1988, 1995) or Tomasello (1999)) are
characterized by two basic assumptions: 1. Re-
enactment of novel means is due to an automatic
tendency to copy the goal-directed action of a human
model, and 2. This tendency is due to a human-specific
drive for identification with human actors who are
perceived through an innate perceptual-motor action
mapping mechanism as similar (“just-like-me”) by the
infant.
As I suggested above, however, I think that this type
of theory suffers from a serious shortcoming in so far as
it cannot account for the functionally selective nature of
human imitative learning that is arguably a significant
adaptive feature of this possibly human-specific
capacity. One piece of highly suggestive evidence
indicating that young children’s imitation of novel goal-
directed actions does not necessarily and automatically
involve the re-enactment of the specific means action
modeled comes from a simple but clever set of studies
designed by Harold Bekkering and Andi Wholschlager
(e. g., Bekkering et al., 2000). In one condition they
asked children between 3 and
Figure 6: Ipsi-lateral error in imitation contra-lateral
action in the Bekkering et al., (2000)
6 years of age to imitate an adult model’s goal-directed
target actions that involved touching either their left or
their right ear with either an ipsi-lateral or a contra-
lateral hand movement (see Figure 6). They found that
while the children were practically errorless in
reproducing the goal of the demonstrated actions
(always touching their correct ear that corresponded to
that of the adult’s demonstration), they were,
nevertheless, rather susceptible to commit a specific
type of error: when contra-lateral hand actions were
demonstrated to them (for example, when the adult
touched his left ear with his right hand reaching across
his body), they very often touched their corresponding
ear (correct goal imitation) with an ipsi-lateral rather
than a contra-lateral hand movement (failing to imitate
the specific means action). In short, in their attempt to
realize the same goal state as the adult, they tended to
substitute for the modeled contra-lateral means action a
simpler, more familiar, and thus more rational
alternative means (the ipsi-lateral hand action) when
such an alternative action was available to them.
7. Inferential constraints on
observational learning in preverbal
infants: Teleological emulation of
new goals versus rational imitation of
new means.
Let me also point out that the intriguing results of
Meltzoff’s (1988, 1995) “magic box” study, showing a
rather automatic readiness on the part of 14-month-olds
to faithfully imitate the unfamiliar ‘head-action’ to
illuminate the box, is also hard to reconcile with the
findings of the series violation-of-expectation studies
(Csibra et al., 1999, 2003; Gergely et al., 1995; Gergely
and Csibra, 2003) with 12-month-olds that I have
reviewed above. This is so because those studies
provided converging evidence that a) by taking the
teleological stance one-year-olds – relying on the
rationality principle - can evaluate which of the
alternative actions available to an agent is the most
efficient means to the goal within the constraints of the
given situation and b) expect the agent to perform that
particular means action that was judged to be the most
rational alternative to realize its goal. Based on the
teleological stance one would therefore predict that
infants should re-enact the demonstrated action only if it
seemed to them to be the most efficient alternative
available to achieve the goal within the situation of the
actor. One may then ask: why did Meltzoff’s subjects
re-enact the novel ‘head-action’ so faithfully, when they
could have simply touched the box with their hands, an
alternative action available to them that is simpler, more
familiar, easier to perform, and so overall clearly a more
rational means to the goal than the novel ‘head-action’?
To solve this riddle, we speculated that Meltzoff’s
situation must have contained some situational features
that actually allowed infants to ‘rationalize’ the ‘head-
action’: in particular, we hypothesized that they may
have noticed and interpreted the fact that even though
the model’s hands were free, she nevertheless did not
use them, but touched the box with her forehead
instead. This observation may have led the infants to
hypothesize that there may be some aspect of the
situation that they haven’t noticed or are ignorant about,
but due to which the novel ’head-action’ must have
some advantage in comparison to the – seemingly more
rational – ‘hand-action’ in achieving the goal. Maybe
then it was in order to figure out (and learn about) the
nature of this assumed advantage, that, since their hands
were also free (and so their situational constraints were
identical to those of the adult), they decided to re-enact
the novel ‘head-action’ themselves.2
To test this hypothesis, we replicated Meltzoff’s study
(Gergely et al., 2001, 2002) with one single
modification using two conditions: in the ‘Hands-
occupied’ condition we changed the situational
constraints of Meltzoff’s original situation by arranging
that the model’s hands were visibly occupied when
performing the ‘head-action’ (she, pretending to be
chilly, wrapped a blanket around her shoulders holding
it tightly with both hands, see Figure 7A). In contrast,
the situational constraints remained the same as in
Meltzoff’s study in the ‘Hands-free’ condition (where
the model also pretended to be chilly and wrapped a
blanket around her shoulders, but then put both of her
hands on the table next to the box so that they were
visibly free to be used, see Figure 7B).
Figure 7A: Hands-occupied Figure 7B: Hands-free
In the ‘Hands-free’ condition (which, as pointed out
in footnote 2 above, was structurally analogous to the
Control condition of the “Squeeze” study), 69% of
infants re-enacted the ‘head-action’, replicating
Meltzoff’s (1988) original finding. By contrast, in the
‘Hands-occupied’ condition, imitation of the novel
2 Note the interesting analogy between this situation and the
stimulus event of the Control condition of the “Squeeze” study
discussed earlier (Figure 4B). In the latter the agent also had a
potential choice of an apparently more efficient alternative
route to the goal (involving a longer pathway which, however,
required no effortful squeezing) that it, nevertheless, did not
take, but approached the goal through the narrow gap that
required more effort. In that case, during the test events
(Figure 5) infants’ relative looking times indicated that they
formed an expectation that the agent would continue to
perform the effortful squeezing action even when a different
alternative route also requiring less effort has again become
available to it. This finding seems analogous to the faithful
imitation of the ‘head-action’ by Meltzoff’s subjects in a
situation where – just as their model – they also could have
opted to touch the box with their free hand, but they still
decided to re-enact the ‘head-action’ that originally must have
appeared to them as the less rational alternative means.
‘head-action’ dropped significantly to only 21% (p<.02)
(Figure 8). Thus, while it must have seemed sensible to
the infants that the model whose hands were occupied
performed the ‘head-action’ to illuminate the box
(goal), 79% of the 14-month-olds decided not to imitate
the ‘head-action’, because for them, whose hands were
free (acting under different situational constraints than
the model), the ‘head-action’ did not appear to be the
most rational means available. In fact, all of these
subjects illuminated the box by touching it with their
hands, a non-imitative means action that was clearly the
most rational alternative under their situational
constraints.3
Finally, (and admittedly unexpectedly) we found that,
whether the subjects re-enacted the ’head-action’ or not,
all infants in both conditions performed the ’hand-
action’ at least once (but often more than once: Mean=
2.1) within the 20 sec time-window of testing. This
suggests that 14-month-olds are subject to an automatic
emulation-like process whereby the memory of the
effect (illumination-upon-contact) activates the response
most strongly associated with establishing contact
(hand-action). These emulative ‘hand-actions’, in fact,
always preceded the imitative ‘head-action’ response
Figure 8: Number of subjects performing the ‘head-action’ vs.
the ‘hand-action’ in the two conditions
3 Note again the structural analogy, this time between the
“Hands-occupied” condition and the Experimental condition
(Figure 4A) of the “Squeeze” study. In the habituation event
of the latter, the agent had no choice, but to squeeze through
the only gap available on the wall to achieve its goal. This is
analogous to the demonstrator’s situation in the “Hands-
occupied” condition, who had no choice but to use her head to
touch and illuminate the box. In the test events of the
“Squeeze” study (Figure 5) when the situational constraints
have changed and an apparently more rational alternative
means action (requiring no effortful squeezing) became
available, infants, as indicated by their relative looking times,
formed an expectation for the agent to take the more rational
alternative route to the goal that has become available. This is
analogous to the finding that in the “ Hands-occupied”
condition infants chose not to imitate the model’s
demonstrated ‘head-action’, but rather chose to emulate the
goal in a rational manner by performing the ‘ hand-action’ that
in their own situation (whose situational constraints were
different from that of the model) seemed the most rational
means available to achieve the goal.
(where there was one) and were always successful in
achieving the goal (illuminating the light-box). This
makes it even more remarkable that the novel ‘head-
action’ was imitated, even though only selectively (and
therefore clearly not automatically) and only in the
“Hands-free” condition. In that condition infants
seemed to have interpreted the demonstrator’s choice to
perform the ‘head-action’ rather than the also available
– and, at least, apparently more rational – ‘hand action’,
to indicate that there must have been some aspect of the
situation (that the infants didn’t notice or understand)
that, after all, justified the demonstrator’s choice of the
‘head-action’ as more rational in achieving the goal
suggesting to the infants that the head-action’ must have
some advantage over the ‘hand-action’ after all. It may
be hypothesized that infants selectively imitated the
‘head-action’ in this condition driven by their
‘epistemic hunger’ to discover and learn about the
nature of this advantage by comparing it to the
alternative ‘hand-action’ (which they also performed).
In conclusion: these results strongly indicate that
early imitation of goal-directed actions is not an
automatic response evoked by identification with the
observed agent, rather, it is the result of a selective
inferential process guided and constrained by the
evaluation of the rationality of alternative means
available both in relation to the situational constraints of
the model and in relation to the situational constraints
of the infant herself.
8. Conclusions
In conclusion let us ask: What is there to be learned
from the infancy studies and our theory of the
teleological stance summarized above from the point of
view of the specific concerns and aims of epigenetic
robotics? In brief, I think that the main message is that
in order to build an even remotely ‘socially relevant’
humanoid robot it will not suffice to construct a
machine that can produce actions, can perceive and
imitate the actions of others, or can even learn to
produce new actions from observing and imitating
actions of other agents. To be able to equip robots with
these capacities is, of course, a highly relevant (and
obviously hard-won) achievement towards the
realization of the ambitious aims of epigenetic robotics,
but, in themselves, they amount to no more than
fulfilling (some of) the necessary preconditions for
constructing a ‘socially relevant’ humanoid robot. In
order to even approximate the competence of preverbal
human infants in the domain of action interpretation and
production, epigenetic robotics will have to turn to the
hard questions of how to construct mechanisms that
implement “top-down” constraints that can make
decisions guiding the action perception and production
system about what action to produce and when (as well
as what action not to generate under specific
conditions), or what action to imitate to reach a goal,
and what goal should be emulated rather that imitated
under certain situational constraints.
9
3
13 14
0
5
10
15
Hands Free Hands Occupied
Head Action Hand Action
I believe the most important lesson that can be
derived from our research on the one-year-old’s
competence of action understanding is that such a “top-
down” system should be conceived to be inferential in
nature practically all the way down and even at the level
when representing the causal mental states of other
agents may not yet be present.
The typical first reaction of connectionist researchers
in AI to our proposal that a teleological action
interpretational system involving such abstract
constructs as ‘goals’ and ‘rationality’ are present (and
likely to be hard-wired) already in preverbal infants is
to try to construct connectionist learning nets that, given
certain input conditions, will be able to simulate the
performance of our infants in our experimental
situations, but will do so without building the abstract
categories (such as goals and the principle of rationality
or efficiency) into the connectionist net in any form.
While I must admit that I am doubtful that such
attempts would eventually succeed in eliminating the
abstract representational concepts in question (certainly,
the specific simulations proposed up till now did not
manage to do so), I think running such simulations is
certainly a worthwhile exercise as they will show us
how far one can get with a purely “bottom-up”
approach: an empirical issue one should not prejudge.
However, if the goal is to construct ‘socially relevant’
humanoid robots, I see no reason why researchers in
robotics and AL should not pursue this goal also by
designing forward engineering solutions (cf. Dennett,
1994) that would equip robots with representational
and inferential mechanisms (and the relevant
knowledge structures that these mechanisms could
access) of the kind that are formalized in our
teleological model and that could implement the “top-
down” constraints necessary to guide the action
production and perception systems to ‘socially relevant’
“choices” about when and what kind of action is
adaptive to execute, imitate, or emulate. I can only hope
that our experimental demonstrations of the essentially
inferential nature of early action understanding and our
formalization of the teleological interpretational system
guiding such inferences may succeed in specifying
useful directions for future research to be pursued in
epigenetic robotics and AL providing “a little help to
our friends” in these neighboring disciplines to realize
their ambitious goal of creating ‘socially relevant’
humanoid robots that they have so bravely set for
themselves.
References
Bekkering, H. et al. (2000). Imitation is goal-directed.
QuarterlyJournal of Experimental Psychology,
53A, 153-64.
Breazeal, C. and Scassellati, B. (2002). Robots that
imitate humans. Trends in Cognitive Sciences,
Vol.16, No.11, 481-487.
Bruner, J. (1957). Beyond the information given. In
Bruner, J. (1973) Going Beyond the Information
Given, New York: Norton.
Carpenter, M.,et al. (1998). Social cognition, joint
attention, and communicative competence from 9
to 15 months of age. Monographs of the Society
for Research in Child Development, Serial No.
255, Vol. 63, No. 4.
Csibra G., and Gergely, G. (1998). The teleological
origins of mentalistic action explanations: A
developmental hypothesis. Developmental
Science, 1:2, 255-259.
Csibra, G. et al. (1998). Beyond least effort: The
principle of rationality in teleological
interpretation of action in one-year-olds. Poster
presented at the Swansong Conference of the
Medical Research Council Cognitive
Development Unit, London, September 1998.
Csibra G. et al., (1999). Goal-attribution without
agency cues: The perception of 'pure reason' in
infancy. Cognition, 72, 237-267.
Csibra, G. et al. (2003). One-year-old infants use
teleological representations of actions
productively. Cognitive Science, vol. 27(1),111-
133.
Dautenhauhn, K. and Nehaniv, C. L. (Eds.) (2002).
Imitation in Animals and Artifacts. Cambridge
MA: MIT Press.
Dennett, D. C. (1987). The Intentional Stance.
Cambridge MA: Bradford Books, MIT Press.
Dennett, D. C. (1994). Cognitive science as reverse
engineering: Several meanings of „top-down“ and
„bottom-up“ (pp. 249-259) In Dennett, D. C.
Brainchildren: Essays on Designing Minds.
Cambridge MA: Bradford Books, MIT Press.
Fadiga, L. et al. (1995). Motor facilitation during
action observation: a magnetic stimulation study.
Journal of Neurophysiology, 73, 2608-2611.
Fodor, J. (1987). Psychosemantics. Cambridge MA:
MIT Press.
Fodor, J. (1992). A theory of the child's theory of
mind. Cognition 44:283–296.
Gergely, G., & Csibra, G. (1997). Teleological
reasoning in infancy: The infant's naive theory of
rational action. A reply to Premack and Premack.
Cognition, 63, 227-233.
Gergely, G., & Csibra, G. (2003). Teleological
reasoning about actions: The naïve theory of
rational action. Trends in Cognitive Science, July,
(in press).
Gergely, G. et al. (1995). Taking the intentional
stance at 12 months of age. Cognition, 56:2., 165-
193.
Gergely, G.,et al. (2001). Rational imitation of goal-
directed actions in 14-month-olds. (pp. 309-315),
In J. D. Moore & K. Stenning (Eds.), Proceedings
of Cogsci 2001, Edinburgh, August 1-5. LEA:
London.
Gergely, G. et al. (2002). Rational imitation in
preverbal infants. Nature, Vol. 415, p. 755.
Heider, F. & Simmel, S. (1944). An experimental
study of apparent behavior. American Journal of
Psychology, 57, 243-259.
Kelemen, D. (1999) Function, goals, and intention:
children’s teleological reasoning about objects.
Trends in Cognitive Sciences, 3, 461-468.
Király, I. et al. (2003). Generality and perceptual
constraints in understanding goal-directed actions
in young infants. Consciousness and Cognition,
(submitted).
Lhermitte, F. et al. (1986). Human autonomy and the
frontal lobes: I. Imitation and utilization behavior:
a neuropsychological study of 75 patients. Annual
Review of Neurology, 19, 326-334.
Meltzoff, A.N. (1988) Infant imitation after a 1-week
delay: Long-term memory for novel acts and
multiple stimuli. Developmental Psychology, 24,
470-476.
Meltzoff, A.N. (1995). Understanding the intentions
of others: Re-enactment of intended acts by 18-
month-old children. Developmental Psychology,
31, 838-850.
Meltzoff, A. N., and Moore, M. K. (1977). Imitation
of facial and manual gestures by human neonates.
Science, 198, 75-8.
Meltzoff, A. N., and Moore, M. K. (1989). Imitation
in newborn infants: Exploring the range of
gestures imitated and the underlying mechanisms.
Developmental Psychology, 25, 954-62.
Meltzoff, A. N. and Moore, M. K. (1997). Explaining
facial imitation: theoretical model. Early
Development and Parenting, 6, 179-92.
Perrett, D. I. et al. (1985). Visual analysis of body
movements by neurones in the temporal cortex of
the macaque monkey: a preliminary report.
Behavioral and Brain Sciences, 16,153-170.
Premack, D. and Premack, A. J. (1997). Motor
competence as integral to attribution of goal.
Cognition. Vol 63(2) May 1997, 235-242.
Rizzolatti, G. et al. (1988). Functional organization of
inferior area 6 in the macaque monkey: II. Area
F5 and the control of distal movements.
Experimental Brain Research, 71, 491-507.
Rizzolatti, G. et al. (1996). Localization of grasp
representations in humans by PET: 1. Observation
versus execution. Experimental Brain Research,
111, 246-252.
Schaal, S. (1999). Is imitation learning the route to
humanoid robots? Trends in Cognitive Sciences,
3:6, 233-242.
Tomasello, M. (1999). The Cultural Origins of
Human Cognition. Cambridge, MA: Harvard
University Press.
Tomasello et al. (1993). Cultural learning. Behavioral
and Brain Sciences, 16, 495-552.
Wolpert, D. M. et al. (2001). Perspectives and
problems in motor learning. Trends in Cognitive
Sciences, 5, 487-494.
Woodward, A. (1998). Infants selectively encode the
goal object of an actor’s reach. Cognition, 69, 1-
34.
Woodward, A. L., and Sommerville, J. A. (2000).
Twelve-month-old infants interpret action in
context. Psychological Science, 11, 73-77.
