A Demonstration of the Efficiency ofDevelopmental LearningMarek W. Doniec, Ganghua Sun, Brian ScassellatiDepartment of Computer ScienceYale UniversityNew Haven, Connecticut 06511, U.S.A.{marek.doniec, ganghua.sun, brian.scassellati}@yale.eduAbstract— Previous research has suggested that developmentallearning can make the learning of advanced sensorimotor andcognitive skills possible. In this paper, we demonstrate thatdevelopmental learning based on skill progression is also moreefficient than traditional divide-and-conquer methods. Using amodel based on the skills of reaching and pointing to visualtargets, we demonstrate an implementation for a humanoid robotthat is more efficient at learning joint attention skills than otherpublished methods. This efficiency results from (1) a structuredset of learning tasks that progresses from low-dimensional tohigh-dimensional problems and (2) a greater exploitation of thelearning environment that does not follow from the completelytask-based decomposition that divide-and-conquer provides.I. INTRODUCTIONThe word “development” has been used in computer sciencewith many different meanings. It can mean maturation ofsensory and motor capacities such as improvement of visualacuity and increase of muscle strength. It is often used as asubstitute for learning of specific skills such as reaching orwalking. In this paper, we reserve the word “development”to mean the acquisition of a progression of skills. Althoughthere is a great deal of variability among individual infants,typically skills in different domains are mastered in an orderlyfashion. Infants usually learn to sit and crawl before they startto walk. Single words are uttered before syntax and grammarare mastered.It has been suggested that following a developmental pro-gression of skills might enable a robot to achieve the intel-ligence or capabilities of an infant [1][2]. Most applicationsof developmental learning in robotics focus on the learningof individual skills. Metta et al. proposed a method to learnvisually-guided reaching by assuming that each arm con-figuration can be decomposed into a few motor primitivessuch that the total number of degrees of freedom to becontrolled are drastically reduced [3]. Schaal et al. introduced aprocedure to use imitation to jumpstart the learning of complexmovements such as a tennis swing [4][5]. Rosenstein andBarto demonstrated that learning to use tools can be facilitatedwith a structured form of reinforcement learning [6]. (Otherlearning examples include [7],[8] and [9].) All of these studieshave shown that diverse, sophisticated skills can be learned byexploiting recent findings from neurophysiology, psychologyand machine learning. However, the benefits of incrementalskill learning have not been sufficiently studied.Incremental development of skills allows for a structureddecomposition of a complex system. Constraints based onlimited perceptual or cognitive capabilities of an infant aid inthe acquisition of complex skills simply by allowing learningto occur first in lower-dimensional spaces. As infants masterbasic skills, the already acquired skills become useful tools toreduce the complexity of learning more complex skills. Whilethe efficiency of such a learning process is never questionedwithin developmental psychology, computational models ofdevelopment have yet to truly demonstrate this efficacy.While developmental learning at first glance seems toclosely resemble traditional divide-and-conquer approachesthat are well known in engineering, the subtleties of a de-velopmental process provide benefits that exceed those ofdivide-and-conquer strategies. In divide-and-conquer, a com-plex problem is broken into a set of simpler sub-problems.Once solutions to the sub-problems have been constructed,these complete components are connected together (typicallyin a sequential chain) to produce a solution to the largerproblem. Developmental learning differs from this process inthat (1) skills need not be learned or applied sequentially and(2) the decomposition of subproblems can be modified at eachstage through interaction with the environment, resulting in aset of components that on face value will not produce thedesired complex behavior but in fact will achieve that resultthrough interaction with the appropriate environment. (We willreturn to this point in Section III.)Empirical studies on how efficiency can be achieved througha developmental structure are lacking. Current work focusesprimarily on uncovering useful mechanisms, such as motorsynergies and imitation, for skill learning. Breazeal has studiedthe use of appropriate facial expressions to facilitate thelearning of social skills [10]. In this work, appropriate facialexpressions are considered to be innate skills and do notrequire learning. In one rare example, Metta and Fitzpatrickshow that by exploiting existing motor skills, the concept ofobject affordance can be learned [11]. They, however, alsohave not provided data on how the developmental methodimproves efficiency.In this paper, we provide an extended example of howdevelopmental learning can be more powerful and efficientthan traditional divide-and-conquer methods by exploitingalready acquired skills. The following section describes adevelopmental model for learning joint attention behaviors,which allow an individual to attend to the same object asanother individual. We demonstrate how this developmentalapproach allows for a faster, more efficient, and more accuratebehavior than those produced by the best divide-and-conquermethods available. We then conclude with a discussion ofwhere the advantages seen in developmental learning originate.II. AN EXTENDED EXAMPLE:LEARNING JOINT ATTENTIONIn this section, we demonstrate the implementation of asystem for joint attention which gives a humanoid robot calledNico the ability to attend to the same object of interest asa human caregiver. The system is constructed from a set ofbasic skill behaviors that include reaching to a visual targetand pointing to a visual target. We describe first the existingsystems that have achieved reasonable results on joint attentiontasks.A. Related WorkReaching can be defined as the arm movement that enablesthe hand or the end-effector to touch a desirable object.Pointing is t he gesture signaling ones interest in an object.Joint attention is the process of recognizing and attending tothe object another person is looking at. While reaching can beseen as a