Proposal of a cloud-based agent for social human-robot interaction that learns from the human experimenters

Abstract

The field of human-robot interaction usually uses an experimental technique called Wizard of Oz where a human operator (the experimenter or a confederate) remotely controls the behavior of the system. Per contra, if robots are autonomous during the interaction, they have a limited pre-programmed set of behaviors. We propose to use reinforcement learning for adaptation of autonomous robotic behavior during the interaction and to benefit from the advantages that brings the field of cloud computing. The overall goal is to design robotic behaviors less boring and more effective and thus, to prepare robots for a long-term human-robot interaction.