Communication is fundamental to successful coordination in teams of robots. Indeed, the promise that robot teams can complete tasks faster and more efficiently than any single agent depends on their ability to share information and effectively coordinate their actions. Today, teams of mobile robots are increasingly being deployed to tackle challenging tasks in environments without existing network infrastructure, relying instead on peer-to-peer communication. While there exists a considerable body of research dedicated to maintaining network connectivity, we still lack methods that are efficient, scalable, and practical. In this thesis we take a number of steps to address these challenges. First, we formalize the problem of Mobile Infrastructure on Demand (MID), wherein a team of mobile robots are deployed to create and sustain a wireless network that supports the communication requirements of a different set of task-oriented robots collaborating to accomplish an objective. This approach decouples the task planning and network maintenance problems and allows us to focus on developing algorithms for the communication/MID team that are task agnostic, enabling a large class of multi-robot algorithms to function without existing network infrastructure. Second, we demonstrate a data driven approach to the MID agent placement problem using convolutional neural networks (CNNs) that achieves comparable performance to an optimization based expert in a fraction of the time and scales to large teams. Finally, we introduce a complete solution to the MID problem that optimizes the position of mobile network relay nodes to directly improve the performance of the underlying routing protocol. We demonstrate our system in a set of experiments with ground robots equipped with 802.11 WiFi radios performing situational awareness and multi-robot exploration.