Predictive Path Coordination of Collaborative Transportation Multirobot System in a Smart Factory

Abstract

Smart factories employ intelligent transportaton systems such as autonomous mobile robots (AMRs) to support real-time adjusted production flows for agile and flexible production. While decentralized transportation task execution provides a scalable multirobot system (MRS) for a smart factory, new coordination challenges arise in implementing such a system. Transportation-MRS collaborates with production-MRS to accommodate just-in-time (JIT) production, leading to nonstationary transportation tasks that transportation-MRS must learn and adapt to. Also, decentralized operation on a shared shop floor means that one robot cannot factor in peer robots’ task execution planning, leading to competitive collisions. Meanwhile, predictive coordination with communication among multiple learning and adapting intelligent robots is still an open problem. On top of identifying the aforementioned challenges, this article first proposes a multifloor transportation graph model to discretize transportation task execution and allow real-time adjustment of transportation paths toward collision-free. We introduce a unique collaborative multi-intelligent robot system approach taking each robot as a cyber–physical agent with automated artificial intelligence (AI) workflow. First, it includes a novel multiagent reinforcement learning (MARL) algorithm, where each robot predictively plans collision-avoidant paths. Second, we introduce a token-passing mechanism to resolve inevitable competitive collisions due to nonstationary tasks. The proposed approach innovatively uses the multifloor model as a domain model for planning. By allowing competitive collision to occur and resolve, a robot only needs to learn and adapt to uncertain parts of the environment—nonstationary tasks and peer robots’ paths. Computational experiments show that our approach is both sample-efficient and computationally efficient. The transportation-MRS quickly reaches near-optimal performance levels, which are empirically shown to scale with the number of robots involved.