Data-Driven Distributionally Robust Coverage Control by Mobile Robots

This paper provides a data-driven solution to the problem of coverage control by which a team of robots aims to optimally deploy in a spatial region where certain event of interest may occur. This event is random and described by a probability density function, which is unknown and can only be learned by collecting data. In this work, we hedge against this uncertainty by designing a distributionally robust algorithm that optimizes the locations of the robots against the worst-case probability density from an ambiguity set. This ambiguity set is constructed from data initially collected by the agents, and contains the true density function with prescribed confidence. However, the objective function that the robots seek to minimize is non-smooth. To address this issue, we employ the so-called gradient sampling algorithm, which approximates the Clarke generalized gradient by sampling the derivative of the objective function at nearby locations and stabilizes the choice of descent directions around points where the function may fail to be differentiable. This enables us to prove that the algorithm converges to a stationary point from any initial location of the robots, in analogy to the well-known Lloyd algorithm for differentiable costs when the spatial density is known.