Probabilistic Modeling And Estimation With Human Inputs In Semi-Autonomous Systems

Abstract

This thesis addresses three important issues that arise in the analysis and design of joint human-robot teams. Each issue deals with a different aspect of the following question: how to best combine human and robot capabilities to accomplish some set of tasks? The first issue addressed here is that of predicting human supervisory control performance in large-scale networked teams of robots. It is shown that models based on individual operator characteristics such as working memory capacity can be used to probabilistically predict human supervisory control metrics under different operating conditions via linear regression, Bayesian network, and Gaussian process models. The second issue addressed here is that of modeling human supervisors of multi-robot teams as discrete strategic decision makers. A probabilistic discriminative modeling approach is presented here, and novel fully Bayesian learning techniques are presented and validated for identifying appropriate discriminative model parameters and model structures from experimental data. The third issue addressed here is that of combining useful information from human observations with information obtained from traditional robot sensors. A novel recursive Bayesian estimation framework is presented for fusing imprecise soft categorical human observations with robot sensor data via Gaussian and Gaussian mixture approximations. The proposed data fusion approach is validated in hardware with a real human-robot team on a cooperative multi-target search experiment.