Scaling Up Sensorimotor Systems: Constraints from Human Infancy

Abstract

Work in human infancy and behavior-based robotics that grounds intelligent abilities in sensorimotor exchanges between a system and its environment shares recurrent problems of when, whether, and how scaling up from basic to supposedly higher abilities is possible. An action-based model of the infant is introduced that converges with features of independently motivated animat models exploiting emergent functionality and challenges alternatives that invoke conceptual representations. Adaptive change routinely exhibited in infants' everyday activities outstrips the scaling-up potential of current robotic systems and clarifies effective principles obeyed by naturally intelligent systems. A general form is outlined to subject-environment interaction that "engineers" restructuring of early abilities in the direction of greater anticipation (considered an upper boundary for the competence of concept-free human and animat systems); and an action-based account of the phenomena is provided. This emphasizes the relationship between representation and situated inference and the role of reciprocal constraints between cognitive and physical-motor mechanisms. Finally, this article questions how far typical self organizing connectionist networks take us toward understanding a system that is capable of mapping recurrent viable patterns of activity into more permanent adaptive changes. , .~, Traditionally, sensorimotor abilities have been of relatively little concern to the mainstream of either artificial intelligence (AI) or developmental psychology. AI's predominantly centralized view of mind has (ab)used the in-principle separation of virtual and physical machines to license preoccupation with disembodied and disembedded programmed models that rely on the manipulation of concepts and of explicit, exhaustive representations to know the world and preplan behavior within it. Developmental psychology has generally allocated sensorimotor abilities to the

Staged Development of Robot Skills: Behavior Formation, Affordance Learning and Imitation with Motionese

“…• In our robotic system, we showed that the use of simple-tocomplex perceptual skills, i.e., first tactile, then visual and finally social cues, was necessary and sufficient for the progressive development of the targeted sensorimotor skills. This is in accordance with the characteristics of infant development [62], and thus we may speculate that the mechanisms we used in our implementation can be considered as a possible model for those mechanisms in the human infant. • We showed that while the necessary mechanisms for goal emulation can be acquired by self-interaction, in order to bootstrap learning of more complex behaviors in a feasible time, the robot needed to leave self-exploration strategy and engage in observational learning by interacting with tutors.…”

“…The sensorimotor system is constrained in the early ages of human development, and these constraints are gradually lifted during development [62], [63]. As discussed in [64] and [65] constrained sensing is very useful to deal with the complexity of input stimuli in absence of necessary perceptual processes, and to reduce the task space for more effective learning.…”

Integration of Active Vision and Reaching From a Developmental Robotics Perspective

“…Initially, these constraints reduce the perceived complexity of the environment and limit interaction, providing a scaffold which helps the infant to make sense of the world [19], [20]. These constraints are then gradually eased, or lifted, allowing the infant to advance into a new stage of development [20].…”

Intrinsic Activitity: from motor babbling to play

“…Learning proceeds under a given set of constraints and eventually a satisfactory level of competence at some tasks will be reached. At this point a new level of task or difficulty may be exposed by the lifting of a constraint [27]. In this way, by building on the previous accumulated experience, the properties of the newly scoped task are discovered.…”

“…These in clude sensory characteristics, motor effects, mechanical properties from anatomical and morphological consid erations, as well as general maturational limitations and environmental effects. These constraints can reduce reso lution, accuracy, bandwidth and/or motor dimensionality and thus restrict the task space and effectively reduce the perceived complexity of the environment [26], [27]. Learning proceeds under a given set of constraints and eventually a satisfactory level of competence at some tasks will be reached.…”