Observation of Human Response to a Robotic Guide Using a Variational Autoencoder

Abstract

This paper proposes a robotic-guide system equipped with a haptic device that can deliver kinesthetic feedback to and receive kinesthetic reaction from a follower. In addition, a feature-extraction method from a depth image of a user following the robotic guide based on a variational autoencoder (VAE) model is presented. One of the major roles of a sensory assistive robot is to help visually impaired people to walk through unknown spaces while avoiding obstacles. Haptic sensory information can be used as a directional cue for these people in recognizing the correct direction. We focus on how people react to haptic guidance from the assistive robot because an accurate prediction for human response enables robots to perform a more active role in not interfering with the human movement. In an indoor experiment, we observed the user reaction following our robotic guide in terms of the kinesthetic force that the user received and the depth image taken from the robot. Using the VAE model, the latent variable well represented the feature of the depth image, e.g., brief position information of a user torso. Furthermore, we tracked the precise trajectory of both the user and robotic guide using a motion-capture system.