Integrering av uppmärksamhetsmekanismer i Graph Neural Networks för flerrobotplanering : En utvärdering av deras inverkan på koordination och prestanda i ett decentraliserat system

This study investigates the integration of attention mechanisms into Graph Neural Networks (GNNs) for decentralised multi-robot path planning (MRPP). Specifically, it assesses whether the use of Graph Attention Networks (GATs) improves robot coordination and overall system performance in an environment where robots have a limited field of view and communicate locally with close neighbours. Two model variants (one based on standard GNNs and the other one on GATs) are implemented and evaluated under the same simulation conditions. Both models are trained on a dataset of 30,000 grid-based scenarios and assessed using success rate and path efficiency metrics. Results show that while GATs perform similarly to standard GNNs, they do not outperform them in terms of success rate or path efficiency. These findings suggest that, under the tested conditions, attention mechanisms do not provide a performance advantage in decentralised MRPP. However, their potential might emerge in more complex or varied scenarios.