Attention Scheduler Based on Reinforcement Learning for Multi-robot System

In a manufacturing job shop, many machines demand the assistance of auxiliary robots, such as supplement raw materials. In order to balance energy saving and effective scheduling, auxiliary robots have to understand the urgency of tasks and plan a safe and stable working path in the job shop. However, previous works based on exact methods and approximation methods suffer from many realistic constraints, such as complex factory environments and non-deterministic polynomial (NP) characteristics. To address the shortcomings of those works, we propose an attention scheduling (Att-Sched) module for Multi-Agent Deep Deterministic Policy Gradient (MADDPG) framework. Instead of hand-crafted function-based algorithms, we leverage MADDPG to tackle nonlinear and NP characteristics between robots and machines with the job shops. To capture the spatial relationships between robots and learn prioritization dispatching rules respectively, we employ the attention mechanism for distinguishing the urgency of tasks. Through experiments on several simulation environments of job shops, we demonstrate our approach can achieve socially acceptable scheduling and fulfill the demands of machines.