Learning to react to abstractions: Accumulating adaptations in a humanoid embodiment

Abstract

Human beings and several other living organisms are capable of acquiring a diverse repertoire of adaptive behaviours or skills, through interaction with an appropriate environment. Based on observations of human embodiment and existing cybernetic theory, an operational description of this form of ‘scalable’ adaptive behaviour is derivied. An articulated mechanism using the principles identified is implemented and used to control the humanoid robot iCub. The experimental physical embodiment is tested with a number of environments. Preliminary results demonstrate a limited form of emergent behavioural growth and corresponding ‘task’-non-specificity: the iCub is able to cumulatively learn multiple ‘tasks’, by chaining together sequences of primative ‘reactions’, to ‘abstractions’.