Staged Competence Learning in Developmental Robotics

Abstract

M.H. Lee, Q. Meng and F. Chao, 'Staged Competence Learning in Developmental Robotics', Adaptive Behavior, 15(3), pp 241-255, 2007. the full text will be available in September 2008Developmental psychology has long recognized the presence of stages in human cognitive development, although the underlying causes and processes are still an open question and subject to much debate. This article draws inspiration from psychology and describes an approach towards developmental growth for robotics that utilizes natural constraints in a general learning mechanism. The method, summarized as Lift-Constraint, Act, Saturate (LCAS), is applicable to all levels of control and behavior, and can be implemented in any robotic configuration. An implementation based on sensory-motor learning in early infancy is described and the results from experiments are presented and discussed.Peer reviewe

A developmental algorithm for ocular–motor coordination

“…Our results show three distinct stages of behaviour emerging from a single process; this shows how qualitative change in behaviour may occur without structural change, but by the consolidation of experience. This concept of staged growth under constraints may provides a valuable method for use in many sensory-motor learning applications [26].…”

“…[41,19,45,13] for a review see [15], but we believe that our developmental approach [26] is able to produce a method with a unique set of features, namely: (i) it is not pre-wired but learns how to saccade, (ii) it learns very rapidly -much faster than current neural network based approaches, (iii) it continuously adapts to correct errors and accommodate any changes in the ocular-motor system, (iv) it does not use or require any calibration process or prior knowledge, and (iv) the generated behaviour displays distinct and qualitatively different stages which emerge during learning.…”

The strategic student approach for life-long exploration and learning

“…Artificial development [57] will require particular structures that will guide exploration and learning beyond what can be addressed by pure measures of learning progress. These mechanisms include maturational constraints [5], [6], [58], the development of intrinsic rewards [24], [59], pre-dispositions to detect meaningful salient events, among many other aspects.…”

An evaluation of environmental constraints for biologically constrained development of gaze control on an iCub robot

“…Constraints can be released when suitable levels of competence have been achieved and we use thresholds on internal state variables to trigger their removal in a semi-structured manner [22]. Hence, a robot built using Type A constraints can be expected to follow the general infant timeline, with variances reflecting its own particular circumstances.…”

“…We have performed such a mapping using the iCub humanoid robot [20] and produced a comprehensive chart of the general developmental possibilities for the sensorimotor systems of the iCub [21]. Sensorimotor learning is conducted based on our mapping framework, and utilises the modulating influence of a dynamic constraint network to shape the developmental sequence following our approach towards constraint based learning [22]. There are various kinds and sources of constraints, but there are two main types and here we consider these as two alternate implementation strategies for robotic systems.…”