Background
With the rapid development of smart manufacturing and automated logistics, Autonomous Mobile Robots (AMRs) have gradually become essential tools in modern factories for material handling, component delivery, and production-line support. However, real-world factory environments are typically highly dynamic. Factors such as the real-time arrival of new tasks, changes in production-line demand, equipment status changes, aisle congestion, and inconsistent AMR execution progress all increase the complexity of task assignment and fleet scheduling.
In a dynamic factory environment, when a new task arrives, the system must determine whether to reassign tasks immediately. If every task change triggers rescheduling, task responsiveness may be improved; however, this also incurs substantial computational overhead. In particular, as the number of AMRs and the scale of tasks increase, the computation time required to resolve the assignment problem rises rapidly, making it difficult for the system to satisfy the real-time requirements of practical deployment. Conversely, if the rescheduling frequency is too low, AMRs may fail to respond promptly to current task demands, leading to longer waiting times, idle resources, or reduced overall throughput.
Traditional task assignment methods often use fixed-frequency scheduling, such as reassigning tasks at regular intervals. However, fixed-period scheduling may not effectively reflect real-time changes in factory conditions. When the environment changes rapidly, fixed scheduling may respond too slowly; when the environment is relatively stable, continuously performing rescheduling may lead to unnecessary computational waste. Therefore, conventional fixed-frequency scheduling is not necessarily the most suitable solution for dynamic factory environments.
Given the above challenges, this subproject aims to develop a high-efficiency scheduling mechanism that can respond to dynamic changes in factory environments. Our scheduling algorithm will dynamically determine the optimal rescheduling timing based on real-time factory information, including task arrival status, AMR locations and load conditions, task waiting times, site congestion levels, and system throughput. It will also adjust the matching relationships between AMRs and tasks in real time. By avoiding unnecessary, frequent rescheduling while ensuring the system responds rapidly to critical state changes, this project aims to strike a balance between computational cost and scheduling performance, thereby further improving the overall operational efficiency and throughput of factory logistics systems.
Research Objectives
This subproject aims to develop a real-time AMR fleet scheduling mechanism suitable for dynamic factory environments. By leveraging real-time site-state perception, adaptive rescheduling triggering, and efficient task matching, the proposed mechanism addresses the computational waste and response delays that may arise from conventional fixed-period scheduling in dynamic job-arrival scenarios.
The proposed method will dynamically determine the optimal rescheduling timing based on information such as task arrival status, AMR locations and load conditions, task waiting times, site congestion levels, and system throughput. It will also adjust the matching relationships between AMRs and tasks in real time. By reducing computational costs, shortening task waiting times, improving fleet utilization, and increasing overall operational throughput, this subproject ultimately aims to strengthen the real-time dispatching capability and operational efficiency of factory logistics systems in dynamic environments.
Methods
This subproject aims to develop a dynamic scheduling framework for flexible manufacturing environments. The framework integrates real-time state perception, adaptive scheduling triggering, and efficient task matching, while coordinating AMR fleets through two-stage path planning to complete the operational cycle of material preparation, material transportation, and production.
First, a GNN will be used to extract relational features from dynamic information related to AMRs, tasks, workstations, and path congestion. Second, we use DDQN to learn the rescheduling triggering strategy, enabling the system to determine the optimal scheduling timing based on real-time site conditions and avoid the computational waste and response delays caused by fixed-period scheduling.
During the task assignment stage, we use a genetic algorithm to rapidly solve the combinatorial optimization problem between AMRs and tasks, thereby reducing idle and waiting times while improving fleet utilization. Finally, using a Coarse-to-Fine two-stage path-planning method, we first generate reference paths and then incorporate spatiotemporal constraints for path refinement. This approach aims to improve the efficiency of real-time planning and the feasibility of paths in dynamic environments.
Innovation
The innovation of this subproject lies in proposing a state-driven dynamic scheduling framework for AMR fleets that integrates real-time site-state perception, learning-based rescheduling triggering, efficient task matching, and Coarse-to-Fine path refinement into a unified decision-making process. Compared with conventional fixed-period scheduling methods, this study not only focuses on “how to assign tasks” but also addresses the fundamental question of “when to reschedule.” By using a GNN to extract dynamic relational features among AMRs, tasks, workstations, and path congestion, and by combining it with a DDQN to learn the optimal scheduling trigger policy, the system can adaptively adjust scheduling timing based on real-time conditions. For task assignment and path planning, this study combines a genetic algorithm with two-stage path optimization, enabling efficient task-matching results while also satisfying practical path feasibility. Overall, this subproject achieves a better balance among computational cost, real-time responsiveness, and system throughput, thereby enhancing the scheduling efficiency and resilience of AMR fleets in dynamic factory environments.
Expected Outcomes
The expected outcomes and benefits of this subproject are as follows:
Real-Time Site-State Perception - We use a GNN-based factory site state representation method to extract relational features among AMRs, tasks, workstations, paths, and congestion states, thereby enhancing the system’s understanding of dynamic environments.
Rescheduling Triggering -
We use DDQN to learn the optimal timing for rescheduling, enabling the system to dynamically determine whether rescheduling is required based on real-time site conditions. Such a design helps avoid computational waste and response delays caused by fixed-period scheduling.
AMR–Task Matching -
We employ a genetic algorithm to rapidly solve the combinatorial optimization problem involving multiple AMRs and multiple tasks. This approach will shorten task assignment time, reduce AMR idle and waiting ratios, and improve fleet utilization.
The proposed approach can reduce scheduling computation costs, shorten task waiting times, increase AMR operational throughput, and strengthen the system's real-time responsiveness under dynamic job arrivals and changing site conditions.
We can apply the outcomes of this subproject to help factories improve resource utilization, material supply stability, and flexible manufacturing efficiency.