Hierarchical task allocation for multi-agent systems encoded by stochastic reachability specifications

We consider the problem of satisfying a set of objectives over a collection of agents. For a single agent, the optimal solution can be obtained via a stochastic reachability framework where optimal control policies come along with a performance metric, defined as the probability of successfully achieving a specified objective. As the number of agents increases, the approach quickly becomes computationally expensive and often intractable. We propose a method which includes an advisory controller that allocates tasks among agents based on their ability of handling individual objectives. This ability is encoded by the stochastic reachability performance metrics. The proposed method is tailored to an autonomous surveillance system composed of pan-tilt-zoom (PTZ) cameras and verified experimentally.