Persistent and Robust Execution of MAPF Schedules in Warehouses

Abstract

Multi-agent path finding (MAPF) is a well-studied problem in artificial intelligence that can be solved quickly in practice when using simplified agent assumptions. However, real-world applications, such as warehouse automation, require physical robots to function over long time horizons without collisions. We present an execution framework that can use existing single-shot MAPF planners and ensures robust execution in the presence of unknown or time-varying higher-order dynamic limits, unforeseen robot slow-downs, and unpredictable obstacle appearances. Our framework also naturally enables the overlap of re-planning and execution for persistent operation and requires little communication between robots and the centralized planner. We demonstrate our approach in warehouse simulations and in a mixed reality experiment using differential drive robots. We believe that our solution closes the gap between recent research in the artificial intelligence community and real-world applications.