Within this project our focus is on teaching human-like interactions to humanoid robots. We aim at learning complex interactions and have a Nao robot repeat those with a person or possibly another humanoid robot. Since humans tend to have a low repetitive accuracy when repeating motions, generalizing capabilities must be integrated.
Therefore, we implemented a novel learning algorithm which is based on human motions reduced in dimensionality while preserving temporal features of a shown interaction. Fundamentally, the algorithm utilizes neural nets or echo state neural nets to learn a mapping between the motions of the two persons teaching the interaction. The learned model is then used to calculate suitable robot motions depending on observed human postures in an ongoing interaction.