Probabilistic Human-Robot Information Fusion

Abstract

This thesis is concerned with combining the perceptual abilities of mobile robots and human operators to execute tasks cooperatively. It is generally agreed that a synergy of human and robotic skills offers an opportunity to enhance the capabilities of today’s robotic systems, while also increasing their robustness and reliability. Systems which incorporate both human and robotic information sources have the potential to build complex world models, essential for both automated and human decision making. In this work, humans and robots are regarded as equal team members who interact and communicate on a peer-to-peer basis. Human-robot communication is addressed using probabilistic representations common in robotics. While communication can in general be bidirectional, this work focuses primarily on human-to-robot information flow. More specifically, the approach advocated in this thesis is to let robots fuse their sensor observations with observations obtained from human operators. While robotic perception is well-suited for lower level world descriptions such as geometric properties, humans are able to contribute perceptual information on higher abstraction levels. Human input is translated into the machine representation via Human Sensor Models. A common mathematical framework for humans and robots reinforces the notion of true peer-to-peer interaction. Human-robot information fusion is demonstrated in two application domains: (1) scalable information gathering, and (2) cooperative decision making. Scalable information gathering is experimentally demonstrated on a system comprised of a ground vehicle, an unmanned air vehicle, and two human operators in a natural environment. Information from humans and robots was fused in a fully decentralised manner to build a shared environment representation on multiple abstraction levels. Results are presented in the form of information exchange patterns, qualitatively demonstrating the benefits of human-robot information fusion. The second application domain adds decision making to the human-robot task. Rational decisions are made based on the robots’ current beliefs which are generated by fusing human and robotic observations. Since humans are considered a valuable resource in this context, operators are only queried for input when the expected benefit of an observation exceeds the cost of obtaining it. The system can be seen as adjusting its autonomy at run-time based on the uncertainty in the robots’ beliefs. A navigation task is used to demonstrate the adjustable autonomy system experimentally. Results from two experiments are reported: a quantitative evaluation of human-robot team effectiveness, and a user study to compare the system to classical teleoperation. Results show the superiority of the system with respect to performance, operator workload, and usability.