An Experimental Evaluation of Bayesian Soft Human Sensor Fusion in Robotic Systems

Abstract

As we move forward into the twenty-rst century, we are seeing new horizons opening to us through the use of autonomous robots as explorers, going places we can’t or won’t. The interaction between the human operator and robotic agent is of paramount importance in exploring these new horizons. Both humans and robots have their strengths and weaknesses, and this paper explores how they compliment each other and looks at new, more ecient ways to facilitate communication between the robot and its operator. This paper investigates conditions to the use of Bayesian information fusion algorithms with Gaussian mixture models to fuse soft human input with robotic sensor data to complete cooperative mission objectives. The condition investigated here are the benets of training individual models for the human input, or if a generic model trained on many people is sucient to complete the mission; we are also seeking to determine if it is useful to allow the human operator to assign a condence value to the information sent to the robot. These two conditions are evaluated with sizable number of human subjects working with a robotic platform to complete a multi-target search mission. Along with the results of these two conditions, reactions to the limited interaction with a robotic partner of a variety of people will be presented based on a survey that was performed once the trials were completed.