A Multi-Robot Task Allocation Method Based on Graph Attention Network and Unsupervised Learning

Abstract

The task allocation for multiple robots is a critical component in the coordination of unmanned clusters. The existing heuristic algorithms are hard to achieve satisfactory results in large-scale problems, and reinforcement learning-based methods face challenges in ensuring the rationality of reward design. This paper introduces a model based on multi-head attention and graph neural networks to address the schedule-dependent multi-robot task allocation problem, trained using unsupervised learning techniques. This model can be trained with varying numbers of robots and tasks without necessitating changes to its structure or parameters. In the experiment of this paper, the model is trained under two different conditions, and the performance is evaluated across six different problem scales. Comparing the proposed model against greedy algorithms and genetic algorithms, the results demonstrate that the proposed model has sianiflcant advantages in overall performance.