Description
The dominant method of science is analysis and simplification. This was clearly articulated by Descartes in 1628: In studying any phenomenon, simplify it to its essential components, dissecting away everything else. This approach is motivated by the belief that complicated systems will be best understood at the lowest possible level. By reducing explanations to the smallest possible entities, the hope is that we will find entities that are simple enough to fully analyze and explain. The spectacular success of this methodology in modern science is undeniable. Unfortunately, it has not given us an understanding of how systems made up of simple elements can operate with sufficient complexity to be autonomous agents. Building artificial agents who can act and adapt in complex and varying environments requires a different kind of science, one that is principally about integration and complexity rather than analysis and simplification. The theoretical task of understanding of developmental process in biological systems also requires a science of integration.
Developmental robotics is based on the premise that principles of developmental process are the key to engineering adaptive and fluid intelligence. Although the promise of this idea is not yet fully realized, remarkable progress has been made over the last decade and half. This book presents the current state of the art. In so doing, the authors also make a case for deeper collaborations between developmental roboticists and developmental psychologists. At present the ties are weak. We are working on related problems, reading the same literatures, sometimes participating in joint conferences, but only rarely actually collaborating in a sustained way. I firmly believe that remarkable gains could be made in both fields through programmatic research by teams of researchers in human development and robotics. For developmental psychology, the promise is both better theory and new ways to test theories by manipulating the pathways and experiences using artificial developing intelligent systems. Accordingly, in this foreword, I highlight seven fundamental aspects of the human developmental process that might be better understood through developmental robotics.
1. Extended immaturity. Development, like evolution and culture, is a process that creates complexity by accumulating change. At any moment, the developing agent is a product of all previous developments, and any new change begins with and must build on those previous developments. Biological systems that are flexibly smart have relatively long periods of immaturity. Why is this? Why and how does “slow accumulative” intelligence yield higher and more abstract forms of cognition? One possibility is that a slow accumulative system—one that does not settle too fast—can acquire the massive amounts of experience that yield multiple layers of knowledge at multiple granularities. A second related possibility concerns what developmentalists sometimes call “readiness” and what recent research in robotics has called “learning progression.”1 As learning progresses, new structures and new ways of learning emerge so that that the same experiences later in development have different effects on the learning system than those experiences earlier in development. If these ideas are correct, then the developmental pathway itself may be part of the explanation as to why human intelligence has the properties that it does. It simply may not be possible to shortcut development—to try to build just the adult system—and achieve fluid and adaptive intelligence that characterizes biologically developing systems.
2. Activity. Learning experiences do not passively “happen” to infants. Piaget2 described a pattern of infant activity that is highly illustrative of this point. He placed a rattle in a four-month-old infant’s hands. As the infant moved the rattle, it would both come into sight and also make a noise, arousing and agitating the infant and causing more body motions, and thus causing the rattle to move into and out of sight and to make more noise. The infant has no prior knowledge of the rattle but discovers—through activity—the task and goal of rattle shaking. As the infant accidentally moves the rattle, and sees and hears the consequences, the infant will become captured by the activity—moving and shaking, looking and listening—and incrementally through this repeated action gain intentional control over the shaking of the rattle and the goal of making noise. Action and exploration creates opportunities for learning and new tasks to be conquered. This role of action is well covered in this book and is an area in which developmental robotics is clearly demonstrating its relevance to theories of development.
3. Overlapping tasks. Developing organisms do not solve just one task; they solve many overlapping tasks. Consider again the rattle example. The infant’s shaking of the rattle couples auditory, motor, and visual systems creating and changing the specialized regions in the brain and the connections between them.4 But these same systems and functional connections enter into many other behaviors and so achievements in shaking rattles may extend to influence means-end reasoning and or the processing of multimodal synchronicities. Developmental theory deeply needs a way to explore how multimodal and multitask experiences create an abstract, general purpose, and inventive intelligence. This is also an area in which developmental robotics is ready to make big contributions.
4. Degeneracy. Degeneracy as it is used in computational neuroscience3 refers to complex systems in which individual components may contribute to many different functions and in which there is more than one route to the same functional end. Degeneracy is believed to promote robustness in developmental outcomes. Because functionally redundant pathways can compensate for one another, they provide a kind of insurance against pathway failure. Robotic models may exploit these principles to build systems that are robust and that can succeed—over the long term and in multiple tasks—even given breakdowns in some components. Such robotic models also offer a rigorous way to test the implications of multicausality and complex systems of causes may constrain developmental outcome.
5. Cascades. Developmental theorists often refer to the far reach of early developments on later ones in terms of the “developmental cascade.” These cascades often evident in the perturbed patterns of atypical development also characterize typical development and such seemingly distinct domains of intelligence as sitting and visual object representation and walking and language input.4 Here is the deeper theoretical question: Are the facts of these cascades—the way earlier developments start the pathway for quite different later developments—relevant to how and why human intelligence has the properties that it does? Developmental robotics may not only advance the engineering of robots by taking on this question, but also provide a platform for understanding how the integrative nature the complex pathways that characterize human cognitive development are essential to human intelligence.
6. Ordered tasks. Biologically developing systems typically confront classes of experiences and tasks in a particular sequence and there is a large theoretical and experimental literature on the cascading developmental consequences of altering that natural order of sensorimotor development in animals.5 Human infants travel through a systematic of set of changing environments in the first two years of life as they proceed to rolling over, reaching, sitting steadily, crawling, and walking. The series of changes in motor skills in the first two years of human life provide strong and most likely evolutionarily selected gates on experience. The consequences and importance of ordered experiences and the significance of perturbations in that ordering have not been theoretically well specified in humans nor systematically pursued in developmental robotics; this is an important next frontier.
7. Individualism. It is the individual that develops. The history of the species may be in the intrinsic biology and environment may contain conspecifics who scaffold development but each developing organism has to travel the path. Because developmental pathways are degenerate, because development builds on itself, because intrinsic biologies and environments are inherently unique, different developing agents may come to comparable functional skills through different paths. This is a theoretically important idea to understanding both the robustness and variability in human intelligence and perhaps also a foundational idea for building multifunctional adaptive robots that can be intelligent in whatever environment they find themselves in.
This book is an excellent steppingstone to future advances in developmental science.