Using Artificial Intelligence for Smart Conveyor Belt Management in Warehouse Environments

Abstract

This thesis investigates whether multi agent reinforcement learning systems can be seamlessly integrated into a production environment, using a Unity based simulation that captures real world constraints as the test bed. The environment is a cooperative benchmark in which three agents coordinate diverter gates on a conveyor belt, and the work compares several cooperative and baseline algorithms against a hand coded heuristic upper bound. Two empirical questions drive the study: whether decentralised cooperative methods can solve a sparse reward coordination task with shuffled destination mappings, and whether adding a learned inter agent communication channel helps or hurts performance under partial observability. The thesis further explores the nature of the communication channel and its benefits in a cooperative setting, and grounds its observations in formal statistical analysis. The contribution lies in the diagnostic methodology and its mechanistic synthesis, which together deliver a quantified and interpretable failure pattern in cooperative sparse reward MARL. On the strength of this analysis, the thesis identifies automatic entropy tuning, discrete message representations, and scaling to larger agent populations as the most promising directions for future work.