A Roadmap for Cognitive Development in Humanoid Robots

This book addresses the central role played by development in cognition.We are interested in particular in applying our knowledge of development in natural cognitive systems, i.e. human infants, to the problem of creating artificial cognitive systems in the guise of humanoid robots. Thus, our subject matter is cognition, development, and humanoid robotics. These three threads are woven together to form a roadmap that when followed will enable the instantiation and development of an artificial cognitive system. However, to begin with, we must be clear what we mean by the term cognition so that, in turn, we can explain the pivotal role of development and the central relevance of humanoid embodiment.

In this chapter, we consider the phylogeny of human infants and, in particular, we look at the innate capabilities of pre-natal infants and how these develop before and just after birth. We begin by looking at the role of action in cognitive behaviour, noting that anticipatory goal-directed actions, initiated by the infant in response to internal motivations, are the key to development. This is consistent with what we said in the previous chapter regarding co-development being a self-generated process. We then proceed to consider the phylogeny of a neonate and the development that occurs prior to birth. We refer to this as pre-structuring and it occurs in several guises: in the morphology of the body, in the motor system, and in the perceptual system. The resultant capabilities that exist at birth are subject to accelerated development early on. These form functional systems to sustain life and to explore and adapt to the infant’s new environment. We then address the core abilities in more detail, looking at core knowledge with respect to capabilities concerning the perception of objects, numeric quantities, space, and people. This brings us to the issue of core social and explorative motives that are responsible for driving development. We conclude this chapter with a summary of the key points that enable the development of cognitive capabilities, the subject matter of the next chapter.

Although all our basic behaviours are deeply rooted in phylogeny, they would be of little use if they did not develop. Core abilities are not fixed and rigid mechanisms but are there to facilitate development and the flexible adaptation to many different environments. Development is the result of a process with two foci, one in the central nervous system and one in the subject’s dynamic interactions with the environment. The brain undoubtedly has its own dynamics that makes neurons proliferate, migrate and differentiate in certain ways and at certain times. However, the emerging action capabilities are also crucially shaped by the subject’s interactions with the environment. Without such interaction there would be no functional brain. Perception, cognition and motivation develop at the interface between neural processes and actions. They are a function of both these things and arise from the dynamic interaction between the brain, the body and the outside world. A further important developmental factor is the biomechanics of the body: perception, cognition and motivation are all embodied and subject to biomechanical constraints. Those constraints change dramatically with age, and both affect and are affected by the developing brain and by the way actions are performed. The nervous system develops in a most dramatic way over the first few months of postnatal life. During this period, there is a massive synaptogenesis of the cerebral cortex and the cerebellum [173, 174]. Once a critical mass of connections is established, a self-organizing process begins that results in new forms of perception, action and cognition. The emergence of new forms of action always relies on multiple developments [371]. The onset of functional reaching depends, for instance, on differentiated control of the arm and hand, the emergence of improved postural control, precise perception of depth through binocular disparity, perception of motion, control of smooth eye tracking, the development of muscles strong enough to control reaching movements, and a motivation to reach.

We now shift from developmental psychology to neurophysiology to focus on the relationship between perception and action. In particular, we are interested in discovering what we can learn about the way a primate brain handles the perception of space, the perception of objects upon which the primate can act, structured interaction, and selective visual attention. In doing so, we will be concerned in particular with teasing out the dependency of perception on actions, both actual and potential. What we learn from this exercise results from a shift in our understanding of the way different parts of the brain interoperate. This shift represents a move away from a prevalent view of a complete separation of function among the dorsal and ventral streams in the brain, the former supposedly dealing exclusively with issues of location and space, the latter supposedly dealing exclusively with issues of identity and meaning. Instead, what emerges is a picture in which the dorsal stream plays a very active role in the recognition of actions and in object discrimination due to their affordances.We will also see that perceptions are directly facilitated by the current state of the premotor cortex.

Having looked at the development of cognitive abilities from the perspective of psychology and neuroscience, we now shift our sights to recent attempts at building artificial cognitive systems. In the following, we will focus on cognitive architectures rather than on specific cognitive systems. We do this because cognitive architectures are normally taken as the point of departure for the construction of a cognitive system and they encapsulate the various assumptions we make when designing a cognitive system. Consequently, there are many different types of cognitive architecture. To provide a framework for our survey of cognitive architectures, we must first address these assumptions and we will do this by beginning our discussion with an overview of the different paradigms of cognition. We have already discussed one approach to cognition in Chap. 1 — enaction — and we will take the opportunity here to position enaction within the overall scheme of cognition paradigms. With the differences between the different paradigms of cognition established, we can then move on to discuss the importance of cognitive architectures and survey the cognitive architecture literature, classifying each architecture according to its paradigm and highlighting the extent to which each architecture addresses the different characteristics of cognition. We will make this survey as complete as possible but we will emphasize those architectures that are intended to be used with physical robots and those that provide some developmental capability. On the basis of this review of cognitive architectures — both general characteristics and specific instances — we will conclude with a summary of the essential and desirable features that a cognitive architecture should exhibit if it is to be capable of forming the basis of a system that can autonomously develop cognitive abilities.

In Chapter 1 we discussed the principles of developmental cognitive systems in general, and of enactive systems in particular. Chapters 2, 3, and 4 identified the constraints arising from the developmental psychology and neurophysiology of neonates, while Chap. 5 revealed a number of insights derived from several computational models of cognition. Now we weave all of these constraints, requirements, and insights together to produce a comprehensive list of functional, organizational, and developmental guidelines for an artificial system that is capable of developing cognitive abilities. These guidelines provide the basis for the design of an enactive cognitive architecture and its practical deployment. In other words, they define a roadmap for the development of cognitive abilities in a humanoid robot, a roadmap which embraces both phylogeny and ontogeny. In the next chapter, we describe the current status of a project to implement these guidelines in a cognitive architecture for the iCub humanoid robot. This cognitive architecture, together with the physical robot, provides the platform for the development of cognitive abilities. The developmental process — or ontogenesis — must proceed in a structured manner. Consequently, we will draw heavily on the material in Chap. 3 on the development of human INFANTS to inform this structure and present a roadmap for ontogenesis. Thus, our roadmap has two sides: the phylogenetic side, informed by enaction, developmental psychology, neurophysiology, and computational modelling, and the ontogenetic side, informed by developmental psychology (see Fig. 6.1). We begin by addressing the phylogeny of the system in Sect. 6.1 and then turn to its ontogeny in Sect. 6.2.

In this chapter, we examine how the roadmap guidelines set out in Chap. 6 have influenced the design of a cognitive architecture for the iCub humanoid robot. We begin with a overview of the iCub and we discuss briefly the iCub mechatronics and software infrastructure. We then describe the iCub cognitive architecture, focussing the selection of a minimal set of phylogenetic capabilities derived from the seven groups of roadmap guidelines. Since the iCub cognitive architecture is a work-in-progress, it represents only a partial implementation of the roadmap guidelines. Consequently, we close the chapter by examining the exact extent to which each guideline has been followed. The next chapter, which concludes the book, addresses some of the research challenges posed by a complete implementation of the roadmap guidelines.

Drawing on the insights from Chaps. 1 to 5, Chap. 6 presented the core of this book: a comprehensive list of forty-three guidelines for the design of an enactive cognitive architecture and its practical deployment as a roadmap of cognitive development in a humanoid robot. Chap. 7 discussed in detail how these guidelines were used to influence the design and implementation of a cognitive architecture for the iCub humanoid robot.We saw that, although many of the guidelines were followed, several were either only partly followed and some have not yet been followed at all (see Table 7.4 in the previous chapter). We emphasize here that these omissions are not because these guidelines are not important — quite the opposite — but because the iCub cognitive architecture, like all cognitive architectures, is a work-in-progress and future versions will reflect more complete implementation of all guidelines. In accomplishing this, we will inevitably face some significant challenges and, in this chapter, we wish to bring the book to a close by re-visiting some of the issues that are particularly pivotal to cognitive development, in general, and the complete implementation of the forty-three guidelines in the iCub cognitive architecture, in particular.