A Three-Layer Reinforcement Learning-based Approach for Dynamic Task Allocation Under Multiple Task Resource Constraints

This paper addresses the Single-Task robot, Single-Robot task, Time-extended Assignment (ST-SR-TA) problem in Multi-Robot Task Allocation (MRTA), focusing on dynamic robot arrivals, diverse task requirements, and heterogeneous robot resources. However, achieving effective allocation in such dynamic environments to optimize cumulative rewards remains a challenge. Classical exact algorithms like dynamic programming are limited to small-scale tasks due to exponential complexity, while heuristic methods typically operate iteratively, requiring re-execution when the environment changes and thus struggling to adapt to dynamic scenarios. To overcome these limitations, we propose a three-layer reinforcement learning method based on attention mechanisms. Specifically, we decompose the task allocation problem with multiple task resource constraints into two sub-problems: capability matching and sequence optimization. We then exploit the attribute and spatial relationships between robots and tasks to design a hierarchical RL strategybyintroducingtwoattentionmechanisms. Thisstrategyaddresses the whole problem progressively across three layers: Global Allocation performs task–robot matching from a global perspective to generate a candidate task pool, Individual Selection determines executable tasks for each robot based on its real-time state, and Sequence Optimization refines the task execution order according to spatial relationships to minimize path cost. Experimental results show that the proposed method significantly improves the task reward compared to other approaches. Moreover, the method demonstrates few-shot and zero-shot generalization to new task allocation scenarios, providing an efficient and practical solution.

Authors

• Shuo Wang Zhengzhou University
• Yuzhen Zhang Zhengzhou University
• Hao Su Zhengzhou University
• Yazhou Hu Zhengzhou University
• Pei Lv Zhengzhou University

Keywords

• Multi-RobotTaskAllocation
• ReinforcementLearning
• DynamicEnvironments
• Multiple Task-Resource Constraints
• Attention Mechanism