A reinforcement learning-based optimization method for task allocation of agricultural multi-robots clusters

The Agricultural multi-robot task allocation (AMRTA) can allocate the optimal operation sequence for the cluster of agricultural robots and improve overall operational efficiency, which is an important research direction for the development of intelligent agriculture. In this paper, we first analyzed the practical requirements of multi-robot task allocation in agriculture and reformulate it as Node Workload-Constrained Multi Traveling Salesman Problem (NWC-MTSP), aiming to minimize the maximum operating time of sub-robots while ensuring a balanced distribution of workload as much as possible. Then, we implemented path planning algorithm required for task allocation and constructed an objective function based on it; we also constructed a graph structure containing workloads of nodes, used graph neural networks to obtain node feature information, and propose a Reinforcement Learning-based Attention Mechanism Policy Optimization Network (NWC-APONet) method to find the optimal allocation scheme. Finally, our model evaluated using real agricultural datasets, i.e., the TSPLIB public dataset and random datasets. Experiments results demonstrate that NWC-APONet achieves superior task allocation, which prove our model’s practical applicability and effectiveness in AMRTA.