In the era of Industry 4.0, collaborative robots are one of the main pillars enabling flexible automation. In particular, multi-robot systems offer a lot of potential to the automation of assembly tasks due to increased flexibility and faster cycle times. However, because the robots have to be precisely synchronized with each other, the setup and maintenance efforts of such multi-robot systems are very high. In addition, necessary temporal and spatial safety margins negatively impact system efficiency. In this work, we present a deep reinforcement learning based multi-agent system for collision-free, minimum-time trajectory planning for multi-robot systems. We show that using the proposed system and control architecture and learning environments, we can successfully train deep reinforcement learning agents in simulation, which we validate on a dual-robot pick-and-place task.
