EFFECTIVE TOOL USE IN A HABILE AGENT
Alexander B. Wood
Tlmmas E. Horton
Robert St. Amant
lnfomtion Sciences Institute
University of Southern California
Manna del Rey, CA 90292 USA
Department of Computer Science
North Carolina State Universi5
Raleigh, NC 27693 USA
ABSTRACT
The intelligent use of tools is a general and important human competence that AI research has not yet examined in
depth. Oilier fields have studied the topic. however. with
results we can compile into a broad characterization of
habile (tool-using) agents. In flus paper we give an overview of research on the use of physical tools, using this information to motivate a set of requirements for building
artificial habile agents. We describe the design of a lmbiie
robot, based on the Aibo platFonn. that can pick up a stick
and use it as a tool to reach objects otherwise out of its
range. We argue that analysis.of activities of such toolusing agents offers an informative waj- to evaluate intelligence.
1
INTRODUCTION
The use of tools is a hallmark of intelligent behavior. It
would be I d to describe modem human life without mentioning tools of one sort or another. For example, before
writing this paragraph, 1 brushed my teeth with a tool,
cooked and ate breakfast with tools, unlocked rm. office
door with a tool, and jotted down notes with a tool. At my
desk 1 am surrounded by tools that support cognitive as
well as physical activities. A calendar aids my memory, a
calculator improves my arithmetic, a whiteboard allows me
to apply my visualization abilities, and of course my computer facilitates an even wider range of cognitive activities.
In philosophical circles, some hold that tool use is central
to intelligent behavior and that it should rival language in
the study of cognitive phenomena (Preston. 1998).
Tool use can be thought of as a specialized case of
problem solving. but it can also be viewed in more general
terms. Nilsson (1995) writes, “[Intelligent programs] need
to be able to find out about what knowledge and.tools are
available to match the problenls they face and to learn how
to use them’’ arguing that qstems with such capabilities
may gain general human competence. Tlus challenge has
been taken up, in part. by researchers in robotics (Bluethmann et al.. 2003; Stoytchey, 2003) and intelligent user interfaces (Sl. Amant and Zettlemoyer, 2000). It also drives
the work presented in tlus paper; we believe that physical
tool use can act as a precursor to cognitive tool use and
thus more intelligent effective behavior.
The overall goal of our research is to develop a description of the class of habile (tool-using) agents in coinputational tenns and to build detailed siinulated and physical robotic systems that can act as csemplan of this class.
Broadly speaking. a sophisticated Ilabile agent can reason
about how to esploit objects and features in its emrironment to reach its goals. A physical lnbile agent can perform spatial and physical reasoning (both qualitative and
quantitative) about objects in relation to the agent’s own
location and dimensions. Most importantly. a habile agent
can reason about the relationships between its capabilities,
the problem at hand. and the tools that it can bring to bear
in order to transform the problem into one that is more easily solved. While specialized physical habile agents already exist (e.g., some factoy robots and most prominently Robonaut (Bluethmann et al.. 2003)). the goal of our
research is to develop agents with much more geneml
physical tool-using abilities - such agents may not even
explicitly incorporate the concept of tools, but can opportunistically esploit objects with appropriate pmperties.
Ths paper focuses on a particular type of tool use.
based on what we describe below as “effective“ tools. In
the following section: we first outline research on natural
tool use in fields outside AI. We use this brief s u n q to
motivate a set of desirable properties in an artificial habile
agent. In Section 3 we describe functional components in
the design of a habile agent for using effective tools and
discuss its implementation on a robot platfonn. A task that
the robot carries out involves obtaining an object that is out
of reach on a shelf by grasping a stick and pushng the object until it falls and can be retrieved. Tlus is one of the
simplest tool-using tasks possible, but it is illustrative of
the requirements on the agent’s design and suggestive of
more complex bsks that can be accomplished in the future.
Section 4 describes related work in robotics, and Section 5
concludes with a discussion of the possibility of defining
tests of agent intelligencebased on tool using ability.
2
MOTIVATION
Accounts of tool use have appeared in the AI iiterature
(e.g.,Agre and Horswill, 1997: Bluethmann et al.. 2003;
Iriki et al. (1996) demonstrated that the body scheina of a
tool user extends beyond the boundaries of the physical
body to include the tool. On release of the tool. the bod?scheina re\.erts to its usual state. Stovtcliey (2003) gves
an extensive review of compamble neuroscience findings:
we discuss his work further in Section 4.
Given a rough understanding of the capabilities of a
habile agent. the issue arises of how agents apply their capabilities. in the contexT of actual tool use. Insight into tlus
issue can be found in taxonomies of tool-based action
There exist esteisive taxonomies of tools based on
application or engineering principles. but.these do not capture important aspects of tools in use. For example. the
U.S. Patent Office describes mrenches as "engaging a work
part and exerting or transmitting a twisting strain thereto,
or means for imparting or transmitting an actuating force to
such a tool" (as quoted in Holm 2005). Taxonomies of
human tools for subsistence (e.g.. spears and fishing implements) due to Oswalt (1 973) come closer, but are too
abstract. Baber gives an estensive overview of tool use
and cognition froin a lwman factors perspective (Baber.
2003): but no general taxonomies of action. Perhaps most
promising is a small taxonomy that describes sofhvare
tools as metaphoors for pliysica1,tools (St. Amant and Hort o n 2002):
Brad? et al., 1984). but the mosl estcnsi\,e analyses are
found elsctvhere. Due to space limitations. n e can give examples of work from onl. a few fields; we focus mainly on
results in animal cognition.
The most widely accepted definition of tool use, in
any field, comes from Beck (1980). "Thus tool use is the
estemal employment of an unattached environmental object to alter more efficiently the fornil position or condition
of another object, another organism or the user itself when.
the user hold5 or carries the tool during or just prior to use
and is responsible for proper and effective orientation of
the tool.'' Although tlis definition may appear unwieldy:
all of its conditions turn out to be needed to distinguish
tool use in the animal kingdom from other activities.
A few esamples illustrate what we can learn by study.ing aniinal tool use. Some of the simplest animals to use
tools are wasps that pound earth down into a nest with the
help of a pebble (Oswalt, 1973). Activities recognizable as
tool use can evolve in agents with limited manipulation no
capacih- for learning, and only the simplest processing capabilities. In summarizing esperiinents with capuchin
monkeys, Visalbergh and Limogelli (1996) observe that
their subjects often engage in rapid-fire execution of different strategies. but with poor judgment in selecting
strategies:.for example, to obtain food from a long narrow
tube, one monkey unwrapped a bundle of sticks but then
used the tape to try to push out the food with the piece of
tape. Such esamples are suggestive of an automated
search process driyen by weak heuristics. In analogous
experiments with chimpanzees. Povinelli (2000) finds that
his subjects base their tool selection on the graspability
rather than the overall effectiveness of the tool. Povinelli
5
Eflective tools produce a persistent effect on materials or the environment. e.g.. harruners, screwdrilTers, and sans.
In.srriv"it.s pmvide information about materials
or the environment, e.g., calipers and magIufving
glasses.
Constraining tools constrain or stabilize nlaterials
for the application of effective tools, e.g., ciamps
and straight edges.
Deiiiarcating tools impose perceivable structure
on the environment or materials, e.g., a carpenter's pencil or pushpins.
further speculates that orangutans and chiinpanzees are
much more capable tool users than gorilIas because the arboreal environments of the former (gorillas are ground
dwellers) have resulted in a flesible kinematic body schema that can be adapted to tool-using activities.
This research is significant in it helps us generally
characterize habile agents in terms of their capabilities and
constraints on their behavior:
'
To shift perspective to the actions of a habile agent,
the agent uses effective tools to transform its actions on the
environment and uses instruments to transform its perception of the environment. The agent uses constraining tools
to stabilize the environment for its effective actions and
uses demarcating tools to add information to the emironment in aid of perception, bg creating external representations. These categories of action guide the selection of
tasks by which we can evaluate the competence of a given
habile agent and judge whether its design is functionally
adequate.
The actions of habile agents are constrained by
the reiationshps (geometrical. kinematic, dynamic. etc.) between their physical architecture
and the tools they use.
In acting. habile agents apply strategies that range
from scripted sequences to search-based problemsolving.
An intemal representation of the agent's unaided
action capabilities adapts (in some cases in a hard
coded fashion, in others more flexibly) to the
agent's acquisition and application of a tool.
3
This last point is bolstered by research in cognitive
neuroscience. In experiments with a macaque monkey,
EFFECTIVE TOOL USE
Consider a variant of the monkey and bananas problem:
76
An agent. has the goal of obtaining a ball, which is
visible on a shelf that is slightly out of Each. The
agent can see various other objects nearby, including
sticks and boxes. How can it retrieve the ball?
ties are automatically accounted for in the behavior of the
agent. For example. if I pick up wooden pointer in the
classroom. I adapt to its dinensions iinmediatel!.. and I am
able to point to different objects as If I were using my index finger: though with sliglitltl!.lower accuracy and speed.
The MMC net (Cruse, 2003. Steinkulder and Cruse.
1998) supporls adaptation of tlus type. The MMC net is a
model for the control of multi-segmented nlanipulators
with redundant degrees of freedom. An MMC net uses a
recurrent network representing various geometrical relationships behveen the segment lengths and joint angles in
the manipulator. The calculations involved in the MMC
rely on tlie notion that several geometric relationships can
be used to calculate the same joint value. Multiple calculations are performed in parallel to find values for all variables. The final value is obtained by calculating the mean
of these multiple computations (MMC). The network is
recurrent, which means that each new iteration uses the
previous iteration's output as input. As the network relaxes, it adopts a stable state corresponding to a geometrically correct solution even if the problem does not fully
constrain the solution.
MMC nets can be applied to both 2D and 3D spaces,
and fall into two basic catagories: linear and nonlinear. A
linear MMC net is somewhat simpler to implement and requires fewer calculations, but allows the lengths of the
segments in the manipulator to van.. Nonlinear nets,
which hold the segment lengths constant, are thus better
models for real-world applications. including the control of
our habile agent. In simulation our implementation of the
MMC net provides for conventional contml in path planning for a two-joint robot ann, adapting almost instantaneously to new requirements on positioning. Furthermore,
we present a hybrid of the linear and nonlinear models for
internal evaluation of morphological constraints against the
environment.
One solution involves pushng a box under the shelf.
climbing on it. and reaclung the ball. Another solution,
one that our work pursues, is for tlie agent to pick up a
stick of appropriate length and push the ball off the shelf.
where it can then be retrieved. The latter solution is an unambiguous example of physical tool use, Though the two
solutions are similar, in that both involve altering the environment to facilitate goal achievement. the latter emphasizes manipulation a key aspect of tool use, rather than
navigation. To accomplish this task, the agent must understand or learn the linlitatioils on its body in relation to the
goal. and how these limitations can be overcome by making use of available objects, specfically by manipulating
these objects as tools.
We begin our work toward building such agents by
identifiing functional capabilities that act as necessarc.
building blocks for tool use. In nature. tool-using abilities
are layered on more basic functions, including object recognition ob-ject manipulation navigation and whateiier
problem-solving capabilities the creature may have. We
base the design of a habile agent for using effective tools
on a comparable set of basic functions, which ma?-vary;
we can imagine habile capabilities built on top of reactive
agents.. planning agents. learning agents. and so forth, with
a variety of constraints on their reasoning and their interaction with the environment. Based on the work surveyed
above, our design efforts focus on two key capabilities that
support effective tool use (instruments and constraining
and demarcating tools remain for future work):
.
Recognizing opjwrhinifiesfor tool use: The general ability to recognize and exploit relationships
behveen the action capabilities of the agent
(sometimes called its effectivities (Tunley and
Slmw, 1979)), its goals, and features of the objects
with 1~1ichit can interact in pursuit of its goals.
-4dupring to new e.$ectii+ties:
The general abilih
to adapt the internal representation of the agent to
match the altered effectivities provided b5- the acquisition and application of a tool.
3.2 Implementation on a Robot Platform
As a test platform for our work, we use a Son? Aibo
robot. Despite its limitations, the Aibo offers the advantages of navigation and inanipulation Capabilities integrated into a single widely available and inexpensive platform. Our development relies on CMU's Tekkotsu. an
application framework for robotic platforms. Tekkotsu is
an event driven object oriented architecture that uses Cc+
to interact with various Aibo control objects. The framework contains modules for vision processing and low level
motor behaviors, as well as a hierarchical state machine.
For convenience, we refer to the system we have developed on the Aibo as Canis Habilisl or simply Canis.
The behavior of Canis as described in the remainder of
this section therefore relies on functional placeholders in
several areas of visual processing: objects are distinguished
by color. rather than shape; a long object is grasped at a
3.1 Schema Adaptation.
Adaptation to the acquisition and application of a tool. another critical component of tool use. is provided by an
adapme internal model. A farmliar propem of tools is
that once in hand. they feel as if the! become a part of the
user. rather than remaining an independent object being
manipulated. When an agent picks up a tool, its effectivities are changed. and for famdiar tools. the new effectivi77
'
particular place on its surface identifiable by a \isual inark:
dimensions of objects are known a priori. The impetus for
providing a siinplified problem was to overcome 1:arious
hardware linutatioiis and provide more reliable perception.
pending further integmion work.
For this task. the actions carried out by Canis are governed, by a state machine, as shown in Figure 1. (In future
work this state machine will be replaced by a more general
motion planning system.) The state machine runs through
tlie sequence of actions shown to achime the goal of reaching and touching the ball directly with its end effector.
Find Goal
Find Tool
\wLical rotational components of segments L1. L2, L3: and
L4. wllich represents the body. neck. head, and tool respectively. Using this simplified model of the body allows the
internal body lo be controlled in a fashion similar to a robotic ann. The mapping between the MMC net model and
the robot body is shown in the overlay in Figure 3 .
Y
kpproath
X
Figure 2: A Four Segment Model of the Robot Body R e p
resented as Segments L1,L2. L3. and L4.
Figure 1: State Machine Controller for Actions.
Multiple observations and studies in animal cognition
have supported these varied levels of sophstication con-
’
cerning action selection. Richard Byrne (2004) hypothesized that non-primate tool use functions are generally limited in a highly specific way. The distinction Byme
suggests between primate tool use and tool use found in
simpler organisms is that “tool use in primates is learnt
from experience, whle tool use in other animals is innate.
coded on species genes” (page 32). The reliance on a high-
level state machine for contml over the tool using behavior
set in image) Body, Neck, Head: and Optionalli a Tool.
is along the similar levels of complexity as tlie scripted
control of tool-using found in insects, but serves as a good
illustration of tool-using capabilities possible on the Aibo.
At a lower level of abstraction, Canis’s manipulation
actions (tool acquisition and application) are controlled by
the MMC net, which acts as its internal body schema.
Canis‘s body schema consists of four connected segments:
body. neck, head, and an optional tool segment. These
map onto the topology of the MMC network. The nehvork
calculates a complete set of state variables relative to the
, body model. The network for Canis takes as input each
segment shown in Figure 2, in addition to horizontal and
The nonlinear MMC net comes into play for two of
the actions in the state machine, Pickup Tool and Reach
Ball. These alter the inputs to the MMC net such that it
converges to a specific posture; once a solution has been
produced, Canis moves into the resulting joint configuration. The Reach Ball action is carried out by internally
computing the conect posture with the M M C net, using a
variable tool length. Taking advantage of MMC net’s
flexibiiity. the tool length can be changed at any time
without the need for recalculating the network. Thus to
78
control motion when no tool is present, a tool length of
zero can be used.
One drawback to the nonlinear MMC network is that it
must remain at a fixed location. Although this model is effective for posture calculations in near space, it would be
useful to use the body schema for testing if an object is
reachable. To support a body centered approach to evaluating the capabilities of the agent and its ability to reach
the ball. we used a niodlfied version of the MMC net that
combines both linear and non-linear approaches.
Recall that in a linear MMC network, segment lengths
are not preserved: as the network converges. the segments
stretch to rcach tlie target location. The idea of body centered reasoning for Canis im-olves translating the body
schema in a two dimensional plane. and performing the
reach task in the world il knows internally. Under this notion the application of the linear nehvork becomes clear;
we can modifi the base segment of the nonlinear network
(Ll) into a linear component to allow for translation. As
the nettvork converges, this translational segment stretches
to the target location.
The evaluation of a reachable object results from the
body schema internally reaching for the ball. Upon the
convergence of the network to a stable posture, the reachable condition is ea? to evaluate by comparing the distance between the end effector and the ball's location.
This representation allows Canis to "imagine" itself doing
the task in order to properly evaluate if an object is reachable.
In order to achieve this translational body schema on a
four segment MMC network, me must adjust the segment's
representations. In Figure 1. L I must now be linear, as it is
tlie translational component of bo* schema. Segments
L2, L3, and L1 correspond to the body; neck, and head respectfully. The representation for a held tool differs from
the nodinear model. Iriki et. a l (1996) suggests that the
tool becomes an extension of the hand during intentional
tool use. Under this premise, we represent a tool by estending the bod? segment that is holding the tool. Although this representation is more approximate than assigning the tool its own segment in the body schema, it
yields an effective approach to tool representation during
the evaluation phase of tool use.
The MMC net surpassed our expectations in its role as
an internal body schema; different networks are intended to
map to the robot body to match different task requirements,
and often this kind of generality is associated with increased variability. For tlus task some actions, such as
picking up a tool, allow little room for emr in calculating
the position of the body. In order to produce controlled
movements to aclueve a desired position joint values must
be known. The internal kinematics performed on the body
model are camed out with the MMC net, and its computations allow Canis to carry out these actions correctly. For
reasons of complexily management however. the MMC net
is not used for all of Canis's actions. The current implementation does not incorporate a segment for tlie mouth of
the agent, nrlich is utilized for,grasping. The MMC net
positions the agent's mouthup to the point of grasping. and
the robot subsequently runs through a scripted motion sequence to coinplete the grasping process.
The system is versatile, even when tliere is no geometrically plausible solution is ayailable. In cases where the
ball is out of reach (with or without the presence of a tool),
the segments converge to a posture that minimizes the distance of the end effector and tlie ball. Thus. the body
model will. still converge to an approximate solution. The
use of the MMC net has been successful in both tool acquisition and ball dislodgement, which offers useful evidence
for the basic utili@ of the adaptive sclmna approach.
1
RELATEDWORKS
Work most closel? related to ours is due to Stoytcliev
(2003): who describes an extended robot body schema for
controlling robot actions based on self-organizing body
schemnas. This approach is coinparable to our work. but
has a stronger emphasis on sensory input and learning.
The work is based on a cognitive neuroscience perspective
on tool use, whch results in design requirements v e q
similar to our own. The main difference between
Stooytcliev's approach and our own is that of scope: we are
interested in a more general conception of tool use, as is
represented by the activities in the task emironment described in the previous section, which include navigation
of the mobile platform. tool identlfication and acquisition,
and eventually tools with other properties. Furthennore.
the model proposed by Stoptchev must remain at a fised
location thus disabling the possibiliQ- of using the body
schema for constraint evaluation.
5
CONCLUSIONS
The current implementation of a habile agent on a robotic
platform Ius expanded on kinematics research related to
advances in MMC application and derivation. Deviating
from the traditional three segment body model common the
MMC net literature, we derived a four segment model in
order to cast a better representation of Canis's body structure. Furthermore. a new form of MMC nehvorks is presented, the linearhonlinear hybrid. which can be used to
support body centered reasoning in the environment. By
using this representation we offer a biologically plausible
solution to evaluating a habile agent's constraints against
the achevement of a goal.
Conceptually, this work supports the basic idea that
we can develop tool-using agents bised on body schemas
that are built into a robotic architecture. We have successfully constructed a habile agent to operate in an obtain ball
contesk The impetus for this study was to establish a
79
'
gain from posing such problems is the opportunity
to obsente an agent's esploraton- strategies.
Learning loof me b!! obsenwtion. TIUS can be
seen as an extension of non-tlansparent tool use.
First, iinagine a task where a correct procedure
may be possible to discover via blind search but
only after a veri- long time. Enter a demonstmor.
who carries out appropriate actions or sequences
.to accomplish the task. If the habile agent can
sohe comparable problems after having observed
tlie demonstrator's actions. we count tlus as an
example of learning. a key facet of intelligence.
Colloborrrtive tool use. One aspect of intelligence
that is untested so far is the ability to communicate. This is addressed in a stage that involves a
collaborator as a part of all tasks. The collaborator attempts to solve a task but cannot succeed,
lacking some crucial physical aid or tool. For esample. acquiring a target object might involve
p i n g up a lid that is too heavy or large for one to
lift alone, tuming a nut on a bolt that rotates freely
unless held in place, and so forth. The habile
agent passes this test by observing the collaborator's actions. recognizing implicit goals, and contributing appropriately to reach the goals.
model. which can evolve across implementations into a
more general tool using agent.
We consider the development of Canis to be a significant engineering feat. not least because it was accomplished on an off-the-shelf. relatively inexpensive platform,
More important. howeyer, are the concepts underlying the
design of the system, and the enonnous potential future
such system haw as flesible agents that can act opportunistically to reach goals in natural environments.
Another direction for future work has a more plulosoplical flavor. Consider that esperimcntal research on tool
use in animals, with the goal of emhating intelligence, has
been carried olit for almost a centuq. Obsenations and
experiments deal with animals ranging from birds to elephants and primates. The evaluation methods developed
by animal cognition researchers sliould provide strong
guidance for the evaluation of artrficial habile agents. It
should be possible to adapt laboraton: procedures to support the evaluation of physical and simulated agents over a
wide range of tool-using capabilities, from simple hardwired insect-level behavior to sophisticated. creative, even
collaborative tool use.
This line of thought leads to a general approach to
evaluating intelligence, which philosopher Ronald Endicott
has called "the tooling test.'' The tooling test is a multistage test for tool-using intelligence, in which each stage
that is passed gives evidence for greater intelligence than
needed in tlie previous stage. The test involves the subject
manipulating physical tools in order to accomplish speclfic
tasks; obstacles are set in the way that can be overconie by
the use of these tools. We can imagine a sequence of increasingly difficult tests:
The tooling test overcomes several limitations of the
imitation game (Turing. 1950). and should be appropriate
for habile agent evaluation. Ficst. the tests have an eqlicit
success criterion: has a predetermined goal been achieved?
Second. the result of a sequence of tests is not a binary intelligentlnon-intelligentjudgment. Rather. just as in research on animal tool use, we can say that a tool-using
agent is more or less intelligent. based on the sophistication
of their actions. Third, the tests are independent of ~fian?.
of the subtleties of l ~ u m
verbal communication, whle
still being able to incorporate basic principles of cooperation.
Trumporent tool use. The task described in Section 3 is an esample of the use of objects whose
behavior is transparent to the tool user. These
sorts of tests are a staple in cognitive research
w~thnon-human primates. Results have helped
researchers to better understand the extent to
which animals can reawn about spatial relationships. kinematics and dynamics. qualitative physics, and so forth in comparison with humans.
Tool construction and modiJication. Consmction
of novel tools and refinement of existing tools is a
relatively recent capacity for h u m . on an evol u t i o n q scale. In the end. it is not clear that
there is a conceptual difference between tool use
directly on materials and tool use for malung
other tools. but this turns out to be a soft b o u n w
between human and animal tool use.
Nontransparent toul use. Sometimes the mechanism of a tool in its operation or its effect is hidden. For example. imagine picking a simple lock
or fishing for a wire inside a wall by feeding an
electrician's snake through a small hole. What we
REFERENCES
Agre. P. and Horswill, I. 1997. Lifewodd analysis. Joitmul
of..lrti$cial Iiirelligence Research 6:llI-145.
Baber, C . 2003. Cognition and Tool Use: F o r m of Engagemenr in Hitman and Snimaf Use of Tools. Taylor
and Francis.
Beck, B. B. 1980. Aniinal Tool Behavior: The Use and
Manifactwe of Tools. Garland Press.
Byme, R. W. 2004. The manual skills and cognition that
lie behind hominid tool use. In Russon, A. E.. and
Begun D. R. (Eds.), The Evolutions of Thought (pp.
3 144). Cambridge, England: Cambridge University
Press.
Bluethmann. W.; Ambrose, R.; Diftler, M.; Askew; S.;
Huber. E.; Goza, M.: Rehnmark, F.; Lovchik, C.;and
80
Magruder. D. 2003. Robonaut: A robot designed to
work with humans in space. ~4tibonoiiious Robots
13(2-3): 179-197
Brady. J. M.: Agre, P.E.: Braunegg. D.J.: and Connell. J.H.
1984. The mechanic’s mate. ilrri$ciu/ Jnkdligence
72( 1-2): 173-2 15.
Cruse, H. 2003. The evolution of cognition: A Ii!.pothesis.
Cognitive Science 27: 135-155.
Holui. D. 2003. A romance of rust: Nostalgia. progress.
and the meaning of took Harper:\. Magazine
3 10(1836):1343.
Iriki. A. Tanaka. M., and Iwaiura, Y. 1996. Coding of
modfied body schema during tool use by macaque
postcentral neurons. Neuroreport 7: 2325 - 2330.
Nilssoa N. 1995. Eye on the prize. 4lMagazine.
Oswalt, W.H. 1973. Habitat and Teclznologv.:The Eiwlu. tion of Himling. Holt, Rinehart. & Winston.
Povinelli. D. Folk Physics fur ilpes, Osford Unkersity
Press. NY.2000.
Preston. B. 1998. Cognition and tool use. Mind and Longunge 13 (4): 513-547
St. Amdnt, R. and Horton, T.E. 2002. Characlerizing tool
use in an interactive drawing em4ronment. In Second
Intemntiunai Sviirposhiiw 0 1 7 Smart Graphics. 86-93.
St. Amant. R. and Zettlemoyer, L. S. 2000. The user interface as an agent environment. In Proceedngs of the
Fourlh internatIonaf Confereiice on 4 iitonotriotls
Apenu. 483490
Steinkuhler U., & Cruse, H. 1998. A holistic model for an
internal representation to control the movement of a
manipulator with redundant degrees of freedom. Biological Cybernetics. 79.157-466.
Stoychev. A, 2003. Computional model for an eiqendabie
robot body schema. Techrucal Reort GIT-CC-03-14,
Georgia Institute of Technology, College of Computing.
Turing, A. 1950. Computing machinery and intelligence.
lus research was under the direction of Dr. Frank E. Ritter
in the Applied Cognitive Sciences Lab. After the compIetion of Ius undergraduate degree, He pursued a Masters of
Science in North Carolina Stale Universip ’s Department
of Computer Science. His research placed an emphasis on
Artificial Intelligence, ivliere lie completed lis studies under the direction of Dr. Robert St. Amant. Wood can be
contacted by email at :?.hb<m.2S~-!;:~i1.e3u.
ROBERT ST. AMANT is Associate Professor in the Department of Coinputer Science at North Carolina State
University. He is currentl?, on sabbatical at the Informtion Sciences Institute. University of Southern California.
His research focuses on intelligent user interfaces. drawing
on interests in artificial intelligence planning. cognitive
modeling. and human-computer interaction. . In 2005 he
chaired the International Conference on Intelligent User
Interfaces. St. Arnant can be reached via email at staa~:a~ti?<:;i
“+??j;:
or ~.‘;zy;ar;t;l.-s
~:<v-,.~i;,+.~;;;,
THOMAS E. HORTON is a P1i.D. candidate in the Department of Computer Science at North Carolina State
University. His acadeinic interests include humancomputer interaction. artificial intelligence, and anthropology.
He may be reached by email at [email protected].
Mind, 59:433360. 1950.
Turvey. M. T. and Slmw, R. E. 1979. The primacy of perceiving: An ecological reformulation of perception for
understanding memory. In Nilsson L. G., ed., Perspecrives on /nemoyv research. Lawrence Erlbaum.
167-222
Visalberglu, E. and Limongelli, L. 1996. Acting and understanding: Tool use revisited through the minds of
capuchin monkeys. In Russon A. E.: Bard,K. A.: and
Parker, S. T., eds., Reaching into Thoughr; The minds
uJrhe grear apes, Cambridge University Press. Chapter 3, 57-79.
AUTHOR BIOGRAPHIES
ALEXANDER E. WOOD received his BS in Infonnation
Sciences and Technology from The Pennsylyania State
University in 2003. During his tenure as an undergraduate,
81