Curiosity-Driven Development of Tool Use Precursors: a Computational Model

摘要

Studies of child development of tool use precursors show suc-cessive but overlapping phases of qualitatively different typesof behaviours. We hypothesize that two mechanisms in par-ticular play a role in the structuring of these phases: the in-trinsic motivation to explore and the representation used toencode sensorimotor experience. Previous models showedhow curiosity-driven learning mechanisms could allow theemergence of developmental trajectories. We build uponthose models and present the HACOB (Hierarchical ActiveCuriosity-driven mOdel Babbling) architecture that activelychooses which sensorimotor model to train in a hierarchy ofmodels representing the environmental structure. We studythis architecture using a simulated robotic arm interacting withobjects in a 2D environment. We show that overlapping phasesof behaviours are autonomously emerging in hierarchical mod-els using active model babbling. To our knowledge, this isthe first model of curiosity-driven development of simple tooluse and of the self-organization of overlapping phases of be-haviours. In particular, our model explains why and how in-trinsically motivated exploration of non-optimal methods tosolve certain sensorimotor problems can be useful to discoverhow to solve other sensorimotor problems, in accordance withSiegler’s overlapping waves theory, by scaffolding the learningof increasingly complex affordances in the environment.