Building Specific Contexts for On-line Learning of Dynamical Tasks through Non-verbal Interaction

Antoine de Rengervé (1) , Souheïl Hanoune (1) , Pierre Andry (1) , Mathias Quoy (1) , Philippe Gaussier (1)

1 ETIS - UMR 8051 - Equipes Traitement de l'Information et Systèmes

Trajectories can be encoded as attraction basin resulting from recruited associations between visually based localization and orientations to follow (low level behaviors). Navigation to different places according to some other multimodal information needs a particular learning. We propose a minimal model explaining such a behavior adaptation from non-verbal interaction with a teacher. Specific contexts can be recruited to prevent the behaviors to activate in cases the interaction showed they were inadequate. Still, the model is compatible with the recruitment of new low level behaviors. The tests done in simulation show the capabilities of the architecture, the limitations regarding the generalization and the learning speed. We also discuss the possible evolutions towards more bio-inspired models.

Domaines: Intelligence artificielle [cs.AI]; Robotique [cs.RO]; Réseau de neurones [cs.NE]; Neurosciences

Antoine de Rengervé, Souheïl Hanoune, Pierre Andry, Mathias Quoy, Philippe Gaussier. Building Specific Contexts for On-line Learning of Dynamical Tasks through Non-verbal Interaction. ICDL-Epirob 2013, Oct 2013, Osaka, Japan. pp.6. ⟨hal-01063958⟩