Robot swarms have been proposed as a way to take advantage of the scalability, robustness, and adaptability of natural large-scale multiagent systems in order to solve engineering challenges. However, accomplishing complex tasks while remaining flexible and decentralized has proven elusive. Our prior work on planner-guided robot swarms successfully combined a distributed swarm algorithm implementing low-level behaviors with automated parallel planners and executives selecting high-level actions for the swarm to perform as a whole, but had only been tested in simplistic grid-world simulations. Here we demonstrate our approach on physical robots augmented with experiments in continuous-space simulation, showing that it is an effective and efficient mechanism for achieving difficult task objectives to which swarms are rarely applied. We also use a Large Language Model prompted with the planning domain definition and a natural language goal statement to generate the formal problem definition, enabling non-expert users to control the swarm.