Decentralised protocols for autonomous mobile robots in material handling: inductive learnings from centralised controller

Abstract

As customer demands for personalised items continue to grow, flexible material handling systems are becoming indispensable in manufacturing to accommodate the increasing product variety and demand fluctuations. Consequently, the adoption of autonomous mobile robots (AMRs) is on the rise, driven by their capability to autonomously respond to changes in demand. However, the myopic nature of AMRs’ autonomous decision-making may limit their ability to consider a system-wide perspective. To address this challenge, we propose an AI-based platform that extracts knowledge from a centralised agent, which makes system-wide decisions offline, and integrates this knowledge into decentralised AMRs operating online. We develop a mixed-integer linear programming model, a genetic algorithm, and a scheduling algorithm to generate centralised solutions. Additionally, we design features to convert these solutions into training instances and employ artificial neural networks to derive knowledge from them. This knowledge is then integrated into each AMR as a protocol. Through experiments, we evaluate the performance of the developed protocol, demonstrating its superiority over other dispatching rules in minimising the total weighted tardiness of transportation requests and the total waiting time of transportation requests at storages.