Multirobot collaborative task dynamic scheduling based on multiagent reinforcement learning with heuristic graph convolution considering robot service performance

Abstract

To address the problem of multirobot collaborative task scheduling considering the degradation of industrial robot performance and the recovery of robot performance through intervention of compensation measures, a robot collaborative task scheduling method based on multiagent reinforcement learning with heuristic graph convolution is proposed in this paper. Five types of constraints between tasks and robots from the temporal and spatial dimensions are designed, and a graph structure with different connection forms is utilized to represent the tasks, robots, and their mutual constraints. The method creates serviceability evaluation (SE) and error compensation (EC) agents to assess the robot's serviceability and compensate for positioning errors. Moreover, it combines heuristic rules and graph convolutional neural networks to create a heuristic graph convolution scheduling (HGCS) agent for task allocation and scheduling. The three agents are trained in a decentralized manner while executing in a centralized manner to achieve dynamic task scheduling that considers the robot's serviceability and error compensation intervention in practical scenarios. Through three case studies involving the evaluation of the robot core drivetrain component degradation status, compensation of industrial robot positioning errors, and multirobot collaborative task scheduling, the proposed method effectively addresses the performance changes in robots during the task allocation and scheduling process. Comparative evaluations with four relevant methods in various task scheduling scenarios demonstrate that the proposed method reduces the average task completion time by 17.99% and improves the resource utilization balance degree by 16.85%.