A study of two complementary encoding strategies based on learning by demonstration for autonomous navigation task

Antoine de Rengervé (1) , Florent d'Halluin (2) , Pierre Andry (1) , Philippe Gaussier (1) , Aude Billard (2)

1 ETIS - UMR 8051 - Equipes Traitement de l'Information et Systèmes
2 EPFL - Ecole Polytechnique Fédérale de Lausanne

Learning by demonstration is a natural and interactive way of learning which can be used by non-experts to teach behaviors to robots. In this paper we study two learning by demonstration strategies which give different answers about how to encode information and when to learn. The first strategy is based on artificial Neural Networks and focuses on reactive on-line learning. The second one uses Gaussian Mixture Models built on statistical features extracted off-line from several training datasets. A simple navigation experiment is used to compare the developmental possibilities of each strategy. Finally, they appear to be complementary and we will highlight that each one can be related to a specific memory structure in brain.

Mots clés: Gaussian Mixture Models; Learning from demonstration; Navigation; neural networks

Domaines: Apprentissage [cs.LG]; Informatique

Antoine de Rengervé, Florent d'Halluin, Pierre Andry, Philippe Gaussier, Aude Billard. A study of two complementary encoding strategies based on learning by demonstration for autonomous navigation task. Tenth international conference on Epigenetic Robotics, Nov 2010, Örenäs Slott, Sweden. pp.105--112. ⟨hal-00551340⟩