Real-Time Supervision for Human Robot Teams in Complex Task Domains

Ongoing research on multi-robot teams is focused on methods and systems to be utilized in dynamic and dangerous environments such as search and rescue missions, often with a human operator in the loop to supervise the system and make critical decisions. To increase the size of the team controlled by an operator, and to reduce the operator\u27s mental workload, the robots will have to be more autonomous and reliable so that tasks can be issued at a higher level. Typical in these domains, such high-level tasks are often composed of smaller tasks with dependencies and constraints. Assigning suitable robot platforms to execute these tasks is a combinatorial optimization problem. Operations Research and AI techniques can handle large numbers of robot allocations in real time, however most of these algorithms are opaque to humans; they provide no explanation or insight about how the solution is produced. Recent studies suggest that interaction between the human operator and robot team requires human-centric approaches for collaborative planning and task allocation, since black-box solutions are often too complex to examine under stressful conditions and are often discarded by experts. The main contribution of this thesis is a methodology to help operators make decisions about complex task allocation in real time for high stress missions. First a novel, human-centric graphical model, TAG, is described to analyze and predict the complexity of task assignment and scheduling problem instances, taking into account the spatial distribution of resources and tasks. Then, the TAG model is extended for dynamic environments to the MAP model. Two user studies were conducted, first in static and then in dynamic environments, in order to identify and empirically verify the key factors, derived from the graphical model, which affect the decision making of human supervisors during task assignment for a team of robots. In these user studies, participants used software tools developed for this work. One of these software tools allows for two different levels of autonomy for the interaction scheme: manual control and collaborative control, with an option to invoke an automated assignment tool. Findings relating to the impact of decision support functionality on the mental workload and the performance of the supervisor are presented. Finally, steering of the common algorithms utilized by decision support tools, using the strategies employed by user study participants, related to the TAG and MAP model parameters, are discussed