Rolf Pfeifer and Josh Bongard, How the Body Shapes the Way We Think:
A New View of Intelligence. Cambridge, MA: The MIT Press, 2007, xxiv +
394 pages. ISBN 10: 0–262–16239–3
Reviewed by Raymond W. Gibbs, Jr. (University of California,
Santa Cruz)
Published in:
Mechanicism and Autonomy: What Can Robotics Teach Us About Human Cognition and Action?,
Special issue of Pragmatics & Cognition 15:3 (2007), pp. 610--614.
Edited by Maria Eunice Quilici Gonzalez, Willem F.G. Haselager and Itiel E. Dror
UNESP Marília / Radboud University / Southampton University
One of the curious features of contemporary research in many parts of
the cognitive sciences is that scholars appear to be more interested
in studying and understanding parts of people than they are in
figuring out how the parts work together to create a whole,
intelligent human being. For instance, many cognitive scientists
distinguish themselves less in terms of their disciplinary
orientation, than in terms of which part of human behavior they work
on such as perception, attention, motor control, language, memory,
decision-making, emotion, consciousness and so forth. Cognitive
scientists certainly believe that the different parts function to-
gether somehow, and at least a few scholars aim to describe the
overall organization of cognition. Yet the essence of cognitive
science, like many scientific enterprises, embraces a “divide and
conquer” strategy, where the issue of how the pieces of mind fit
together is a topic for latter discussion and study.
The “divide and conquer” strategy has been applied notably to
the topic of intelligence. Traditional efforts to understand the
nature of human intelligence sought to define what is special about
intelligence, as typically measured on various sorts of IQ tests. But
in the 1980s, starting with the work of Howard Gardner (1983), psy-
chologists and others began exploring the idea that there are
“multiple intelligences”composed of different abilities (e.g.,
linguistic, logical-mathematical, spatial, bodykinesthetic, musical,
intrapersonal and interpersonal). Once more, psychologists rushed off
to study these different “faculties of mind” with the typical belief
that these varying kinds of intelligence likely had little to do with
one another. Thus, our verbal or linguistic abilities seem very
different from our abilities to move about in the world, an idea
rooted in the traditional belief of a separation of mind and body.
Yet the last 15 years has also witnessed a deep questioning of the
existence of a “disembodied mind” that is perhaps composed of
isolated faculties or “modules”. Many cognitive scientists aim to
study the “embodied mind” to show how people’s bodily experiences
provide much of the fundamental grounding for human cognition and
behavior. Cognition is now seen as what occurs when the body engages
the physical and cultural world, and should be studied in terms of the
dynamical interactions between people and the environment (Gibbs
2006). One implication of this view is that cognition is no longer
assumed to be purely internal, symbolic, computational, and
disembodied, but inextricably shaped by embodied action. Cognitive
scientists typically still study parts of mind (e.g., language,
perception, memory), but are far more interested now in seeing how the
parts not only fit together, but emerge from similar processes of
bodily action. In this way, understanding the nature of intelligent
thought and action are not seen as separate tasks, but as being
fundamentally intertwined.
Rolf Pfeifer and Josh Bongard’s book represents another of a
growing list of monographs that describe thinking and intelligent
action as arising from our specific morphology and the material
properties of our bodies. The distinguishing feature of Pfeifer and
Bongard’s book and work is their embrace of the basic methodology of
artificial intelligence — “understanding by building” — in terms of
which they describe recent advances in creating intelligent
robots. Many decades of work in AI have demonstrated the tremendous
difficulty of getting machines to perform complex cognitive tasks
such as understanding language and solving problems. Moreover, work in
robotics has had limited success in building machines that move about
the world in intelligent ways, especially when these machines were
“programmed” to plan and execute specific behaviors given real-
world contingencies. Part of the problem was that scholars assumed too
great of a separation of mind and body with the belief that internal
thoughts (i.e., symbolical programs) needed to direct a machine to
move in very specific ways. Under this view, a machine’s “body”
passively obeyed the commands of some “inner mind”. Yet the
promising recent trend in AI (and work in the closely allied field of
“artificial life”) has been to acknowledge that thought is not
independent of the body, but enabled by embodied activity, enough so
that intelligent thought and action is understood as emerging from the
interaction of brains, bodies, and world.
Consider as one example the task of building a robot that cleans
up a room littered with Styrofoam cubes. There are certain ways of
approaching this problem. The robot has to first find a cube, then
locate the nearest heap or cluster of cubes, and then pick up the cube
just encountered and bring it to the desired location and deposit it
there. Although this task sounds easy, it actually requires complex,
sophisticated visual processing and planning that could take
considerable computational effort. But one robot, the “Swiss robot”
accomplishes the clean-up task by taking advantage of its own
morphology and the properties of the environment around it. In fact,
the Swiss robot is only programmed with simple reflexes for avoiding
obstacles. To clean up the room, the robot exploits information about
the size of the cubes, their weight, and the fact that the environment
is a room enclosed by walls, and has a flat floor that is easy for
traverse. If the cubes were too big or heavy, and if the room did not
have walls, the robot would be unable to complete the task. But
given the constraints of its own morphology, the laws of physics, and
the properties of the environmental niche, the robot can go about
picking up and depositing the cubes in a computationally cheap
manner. There is no need for the robot to have as part of its
“programmed knowledge” all the various rules that one could assume
underlie planning and carrying out these various actions. By
utilizing its own “body” the Swiss robot can act intelligently
without having a mind full of intelligent programs.
Pfeifer and Bongard describe many other, even more complex, robots
that can accomplish various intelligent real-world behaviors such as
an artificial mouse that has a sophisticated whisker system enabling
it to navigate a maze with walls of different textures and the
infamous humanoid robot Cog that was able to manipulate the
environment by poking objects in its path to see how they move — an
example of vision through sensory-motor coordination. More generally,
Pfeifer and Bongard assert that there are five essential properties of
the “complete agent”that reflect their embodied nature (p. 95):
1. They are subject to the laws of physics.
2. They generate sensory stimulation through their interaction with the real
world.
3. They affect the environment through their actions.
4. They are complex dynamical systems that have the tendency to settle
into different attractor states.
5. They perform morphological computation such that the body performs cer-
tain processes that otherwise would have to be controlled by the brain.
At the same time, Pfeifer and Bongard provide various “design
principles” that should guide work on building intelligent agents,
including:
1. The three-constituents principle: Designing an intelligent agent
involves defining the ecological niche, defining the desired
behaviors and tasks, and designing the agent.
2. The complete agent principle: In designing agents, we must think about the
complete agent behaving in the real world.
3. The cheap design principle: If agents are built to exploit the
properties of the ecological niche and the characteristics of the
interaction with the environment, their design and construction
will be much easier or “cheaper”.
4. The redundancy principle: Intelligent agents must be designed so
that their different subsystems function on the basis of different
physical processes, and there is partial overlap of functionality
between different subsystems.
5. The sensory-motor coordination principle: Sensory stimulation is induced
through sensory-motor coordination.
6. The ecological balance principle: Given a certain task environment,
there has to be a match between the complexities of the agents’
sensory, motor, and neural systems. Furthermore, there must be a
certain balance or task distribution between morphology,
materials, and environment.
7. The parallel, loosely coupled processes principle: Intelligence is
emergent from a large number of parallel processes that are
coordinated through embodiment, or embodied interaction with the
environment.
These design principles are extremely important not only for building
intelligent robots, but for understanding the essence of human
cognition and action. Although people outside the AI and AL community
may first question the relevance of machines that pick up Styrofoam
cups and use artificial whiskers to move through a maze for
understanding higher-order intelligence, there are significant lessons
about the nature of intelligence that cognitive scientists and
scholars from the humanities can learn by adopting these design
principles. First, there are major limits to the typical “divide and
conquer” strategy where human beings are assumed to be a mere
compilation of independent faculties. The main drawback is that
scholars end up assuming too much cognition “in the head” is needed
to cause intelligent behavior to occur in the real-world. This
criticism is just as applicable to understanding higher-order
cognitive abilities such as language use and memory as it is to
simpler human actions such as adaptive movement in the environment.
Second, Pfeifer and Bongard properly emphasize the importance of
adopting a “dynamical systems” approach to intelligence in which an
individual’s behavior emerges from nonlinear interactions of brains,
body and environment rather than from simple directional
brain-body-world causality. Dynamical systems are extremely
sensitive to contextual variations in a way that standard linear
systems are not because of the great difficulty of reassembling
individual modules together to create adaptive intelligent
action. Overall, the design principles outlined in this book offer
concrete ways of thinking about intelligence as arising from a
“complete agent” or “whole person” rather than from separate parts
of machines or people.
Later chapters in the book extend these different design principles
to address important research on the development and evolution of
cognition, collective intelligence (i.e., multiple agents cooperating
in the pursuit of some goal), how embodied intelligence can be
exploited in the design of “ubiquitous computing technology” (e.g.,
the use of both autonomous and non-autonomous technologies to augment
human abilities, such as navigation, and information retrieval), and
the development of robotic technology in everyday life (e.g., humanoid
companion robots capable of social communication).Unlike many
earlier writings in AI, where scholars made fantastic claims about the
future of computing and robotics and predicted that robots would soon
(in the 1970s and 80s) be capable of doing all sorts of human-like
behaviors, Pfeifer and Bongard stick to what has been already
accomplished in AI and AL, and correctly note that simply building ro-
bots that replicate human or animal behaviors will be inadequate. What
is always needed, both for building intelligent robots and
understanding intelligent human behaviors, is complete recognition of
an agent’s own embodiment, precisely because this provides the
grounding for the ways we think and act.
I admit a bias as a reviewer of this book because of my own work
on embodied cognition and language, and my enthusiasm for dynamical
approaches to intelligent action (Gibbs 2006; see also Juarrero
1999). Nonetheless, I can’t help but say that “How the Body Shapes
the Way we Think” is a marvelous book that offers profound insights
into the fundamental nature of natural and artificial intelligence.
Pfeifer and Bongard are to be congratulated for writing such an
inspiring book, one that is well-written, tutorial in all the right
ways (e.g., their explanations of dynamical systems theory, artificial
evolution, and symbol grounding) and deeply appreciative of both the
beauty and intelligence of bodies and bodily experience.
References
Gardner, H. 1983. Frames of Mind: The Theory of Multiple
Intelligences. New York: Basic Books. Gibbs, R. 2006. Embodiment and
Cognitive Science. New York: Cambridge University Press. Juarrero,
A. 1999. Dynamics in Action: Intentional Behavior as a Complex
System. Cambridge, MA: The MIT Press.
Author’s address
Raymond W. Gibbs, Jr.
Dept. of Psychology
University of California, Santa Cruz
Santa Cruz, CA 95064
USA
gibbs@ucsc.edu
About the author
Raymond W. Gibbs, Jr. is Professor of Psychology at the University of
California, Santa Cruz. He is author of The Poetics of Mind:
Figurative Thought, Language, and Understanding (1994), Intentions in
the Experience of Meaning (1999), and Embodiment and Cognitive Science
(2006). He is co-editor with Gerard Steen of Metaphor in Cognitive
Linguistics (1999), with Herbert Colston of Irony in Language and
Thought: A Cognitive Science Reader (in press), and editor of the
journal Metaphor and Symbol. His research interests include
psycholinguistics, figurative language, and pragmatics.