Multi-agent reinforcement learning for controlling gantry robot systems

Industry 4.0 forces a major transition in the field of production and logistics. On one hand this transition offers a great potential for increasing efficiency and productivity, but on the other hand, it requires a fundamental adaptation of processes and the associated software systems. One of the new requirements is that production control algorithms should be able to autonomously and dynamically adjust to changing conditions on the shop floor. Reinforcement learning is one approach to enable the required autonomy. This paper presents a new decentralized multi-agent reinforcement learning (MARL) algorithm for controlling complex gantry robot systems that meets these requirements. The algorithm is capable to train autonomous agents for gantry robot systems control utilizing more than one gantry or producing more than one product efficiently. Furthermore, the training of the MARL approach is much faster compared to a comparable single-agent approach. The training results are validated and presented for different setups of the gantry system.