| dc.description.abstract | Multi-robot task allocation is a general approach to coordinate a team of robots to complete a set of tasks collectively. The classical works adopt relevant theories from other disciplines (e.g., operations research, economics), but oftentimes they are not adequately rich to deal with the properties from the robotics domain such as perception that is local and communication which is limited. This dissertation reports the efforts on relaxing the assumptions, making problems simpler and developing new methods considering the constraints or uncertainties in robot problems.
We aim to solve variants of classical multi-robot task allocation problems where the team of robots operates in dynamic and uncertain environments. In some of these problems, it is adequate to have a precise model of nondeterministic costs (e.g., time, distance) subject to change at run-time. In some other problems, probabilistic or stochastic approaches are adequate to incorporate uncertainties into the problem formulation. For these settings, we propose algorithms that model dynamics owing to robot interactions, new cost representations incorporating uncertainty, algorithms specialized for the representations, and policies for tasks arriving in an online manner.
First, we consider multi-robot task assignment problems where costs for performing tasks are interrelated, and the overall team objective need not be a standard sum-of costs (or utilities) model, enabling straightforward treatment of the additional costs incurred by resource contention. In the model we introduce, a team may choose one of a set of shared resources to perform a task (e.g., several routes to reach a destination), and resource contention is modeled when multiple robots use the same resource. We propose efficient task assignment algorithms that model this contention with different forms of domain knowledge and compute an optimal assignment under such a model.
Second, we address the problem of finding the optimal assignment of tasks to a team of robots when the associated costs may vary, which arises when robots deal with uncertain situations. We propose a region-based cost representation incorporating the cost uncertainty and modeling interrelationships among costs. We detail how to compute a sensitivity analysis that characterizes how much costs may change before optimality is violated. Using this analysis, robots are able to avoid unnecessary re-assignment computations and reduce global communication when costs change.
Third, we consider multi-robot teams operating in probabilistic domains. We represent costs by distributions capturing the uncertainty in the environment. This representation also incorporates inter-robot couplings in planning the team’s coordination. We do not have the assumption that costs are independent, which is frequently used in probabilistic models. We propose algorithms that help in understanding the effects of different characterizations of cost distributions such as mean and Conditional Value-at-Risk (CVaR), in which the latter assesses the risk of the outcomes from distributions.
Last, we study multi-robot task allocation in a setting where tasks are revealed sequentially and where it is possible to execute bundles of tasks. Particularly, we are interested in tasks that have synergies so that the greater the number of tasks executed together, the larger the potential performance gain. We provide an analysis of bundling, giving an understanding of the important bundle size parameter. Based on the qualitative basis, we propose multiple simple bundling policies that determine how many tasks the robots bundle for a batched planning and execution. | en |
