Simulation aided, knowledge based routing for AGVs in a distribution warehouse

Traditional routing algorithms for real world AGV systems in warehouses compute static paths, which can only be adjusted to a limited degree in the event of unplanned disturbances. In our approach, we aim for a higher reactivity in such events and plan small steps of a path incrementally. The current traffic situation and also up to date time constraints for each AGV can then be considered. We compute each step in real time based on empirical data stored in a knowledge base. It contains information covering a broad temporal horizon of the system to prevent costly decisions that may occur when only considering short term consequences. The knowledge is gathered through machine learning from the results of multiple experiments in a discrete event simulation during preprocessing. We implemented and experimentally evaluated the algorithm in a test scenario and achieve a natural robustness against delays and failures.