Managing temporal uncertainty under limited communication : a formal model of tight and loose team coordination

Abstract

In the future, groups of autonomous robots will cooperate in large networks in order to achieve a common goal. These multi-agent systems will need to be able to execute cooperative temporal plans in the presence of temporal uncertainty and communication limitations. The duration of many planned activities will not be under direct control of the robots. In addition, robots will often not be able to communicate during plan execution. In order for the robots to robustly execute a cooperative plan, they will need to guarantee that a successful execution strategy exists, and provide a means to reactively compensate for the uncertainty in real-time. This thesis presents a multi-agent executive that enables groups of distributed autonomous robots to dynamically schedule temporally flexible plans that contain both temporal uncertainty under communication limitations. Previous work has presented controllability algorithms that compile the simple temporal networks with uncertainty, STNUs, into a form suitable for execution. This thesis extends the previous controllability algorithms to operate on two-layer plans that specify group level coordination at the highest level and agent level coordination at a lower level. We introduce a Hierarchical Reformulation (HR) algorithm that reformulates the two-layer plan in order to enable agents to dynamically adapt to uncertainty within each group plan and use a static execution strategy between groups in order to compensate for communication limitations. Formally, the HR algorithm ensures that the two-layer plan is strongly controllable at the highest level and dynamically controllable at the lower level. Furthermore, we introduce a new fast dynamic controllability algorithm that has been empirically shown to run in O(N³)

(cont.) The Hierarchical Reformulation algorithm has been validated on a set of hand coded examples. The speed of the new fast dynamic controllability algorithm has been tested using a set of randomly generated problems.