On robots imitating movements through motor noise prediction

Abstract

Robot ego-noise, that is the noise produced while the robot is moving around, can carry useful information about the motor system and the embodiment of the agent. We present an experiment where a mobile robot acquires knowledge about its ego-noise. In particular, we adopt a learning strategy based on self-exploration behaviours and on an inverse model for encoding the mappings between ego-noise and the motor commands that produced it. A Convolutional Autoencoder is adopted for semi-supervised learning of auditory features. The inverse model maps both auditory features and perception of the robot speed to the motor commands that produced the ego-noise. We demonstrate how the trained models can be used for imitating movements from listening to the noise they produce.