Learning from human-robot interactions in modeled scenes

Abstract

There is increasing interest in using robots in simulation to understand and improve human-robot interaction (HRI). At the same time, the use of simulated settings to gather training data promises to help address a major data bottleneck in allowing robots to take advantage of powerful machine learning approaches. In this paper, we describe a prototype system that combines the robot operating system (ROS), the simulator Gazebo, and the Unity game engine to create human-robot interaction scenarios. A person can engage with the scenario using a monitor wall, allowing simultaneous collection of realistic sensor data and traces of human actions.