Motivations Intrinsèques et Contraintes Maturationnelles pour l'Apprentissage Sensorimoteur

Adrien Baranes 1
1 Flowers - Flowing Epigenetic Robots and Systems
Inria Bordeaux - Sud-Ouest, U2IS - Unité d'Informatique et d'Ingénierie des Systèmes
Abstract : Learning new sensorimotor knowledge and know-how in real environments leads to an important number of challenges for today's robots. In order to learn new skills, they need to explore sensorimotor spaces which are generally high-dimensional, high-volume, redundant, and possess areas of heterogenous levels of complexity. In this thesis, introduced within the developmental robotics domain, we propose different processes in order to guide and constrain the autonomous acquisition of new sensorimotor skills in such spaces. We propose an unified approach in order to resolve these problems which takes inspiration from phenomenon of developmental constraints introduced in biology and psychology, and more particularly intrinsic motivations and maturational constraints. After formalizing a computational framework based on these notions, we present three different algorithmic architectures, each one reused in an integrated manner within the next one : The first one, called RIAC, for Robust-Intelligent Adaptive Curiosity, corresponds to the implementa- tion of an active learning algorithm which orients the exploration in bounded spaces whose dimensionality is known and which possess regions of different levels of complexity. This system, which uses heuristics taking inspiration from knowledge based intrinsic motivations mechanisms, effectively directs a progres- sive exploration of new sensorimotor knowledge, which corresponds to the learning of forward models. It also leads to the emergence of self-organized developmental trajectories related to the orientation of the sensorimotor exploration toward activities of intermediate complexity. Then, we propose the SAGG-RIAC algorithm for Self-Adaptive Goal Generation - RIAC, as a com- petence based intrinsic motivations exploration mechanism, which allows highly-redundant robots whose sensorimotor spaces are high-dimensional to learn effectively and actively new motor skills in their task spaces. The main idea of this algorithm is to guide the robot to do active babbling in a low-dimensional task space, in contrast with a motor babbling carried out in a higher-dimensional control space, by acti- vely and adaptively self-generating goals in regions of the task space which bring the highest improvement of competences for reaching previously attempted goals. Finally, we introduce the McSAGG-RIAC algorithm for Maturationally-Constrained SAGG-RIAC, which is based on a coupling of computational models of intrinsic motivation and physiological matura- tional constraints. We argue that these mechanisms may have complex bidirectional interactions allowing the active control of the increase of complexity in the sensorimotor development, in order to direct efficient learning and exploration processes. We introduce more particularly a functional model of maturational constraints inspired by the biological process of myelination, and show how this can be coupled with the SAGG-RIAC algorithm. We show qualitatively and quantitatively that this integrated approach of the three architectures introduced in this thesis answers some problematics raised by real environments, by controlling the complexity, volume, dimensionality and redundancy of skills explored in a manner intrin- sic to the robot, thus decreasing in an important extent the necessity of constraining and preparing the environment in en external manner.
Document type :
Other publications
Liste complète des métadonnées

https://hal.inria.fr/hal-00653308
Contributor : Adrien Baranes <>
Submitted on : Monday, December 19, 2011 - 12:12:08 PM
Last modification on : Thursday, November 16, 2017 - 5:12:03 PM

Identifiers

  • HAL Id : hal-00653308, version 1

Citation

Adrien Baranes. Motivations Intrinsèques et Contraintes Maturationnelles pour l'Apprentissage Sensorimoteur. 2011. ⟨hal-00653308⟩

Share

Metrics

Record views

302