This paper presents the dynamic hierarchical attention graph neural network (DHA-GNN), an innovative framework designed for real-time task allocation in UAV swarms navigating complex, dynamic environments, such as urban cityscapes and rugged mountainous terrains. DHA-GNN integrates hierarchical feature aggregation with adaptive multi-level attention mechanisms to model evolving graph structures, capturing real-time interactions among heterogeneous UAVs, tasks, and environmental dynamics. The methodology employs self-learning graph representations to optimize task prioritization and resource allocation under constraints like dynamic obstacles, communication disruptions, and shifting mission priorities. Extensive simulations across diverse scenarios, including industrial facilities, forested regions, and counter-UAV operations, demonstrate that DHA-GNN achieves a task allocation accuracy of 98.7%, outperforming traditional methods by 15% in efficiency and reducing decision latency to 0.0023 milliseconds. It excels in urban (98.2% accuracy) and rugged terrains (97.6% accuracy), ensuring robust adaptability. These results establish DHA-GNN as a leading solution for UAV swarm intelligence, significantly enhancing operational efficiency in disaster response, surveillance, and military applications.

Task allocation in complex multiagent systems involves assigning tasks to agents with varying capabilities to optimize overall performance. The challenge lies in selecting the most suitable agent for each task, considering the agents’ heterogeneity and the intricate relationships between tasks. Traditional methods often fail to capture this complexity. To address these limitations, we propose the graph multiagent task allocation neural network (GMATANN), a novel approach utilizing a graph attention mechanism. GMATANN models the interactions between agents and tasks through a task-agent graph, where both agents and tasks are represented as nodes, and their associations are depicted as edges. The graph attention mechanism is crucial for capturing the key relationships and ensuring effective information flow between nodes. By learning attention weights, the network automatically identifies which agents are best suited for specific tasks. We employ a neural network framework based on this attention mechanism to train and evaluate the method. Simulation experiments demonstrate the effectiveness of GMATANN, achieving a task allocation accuracy of 92.3% and a reliability of 94.2%, outperforming traditional approaches. This innovative method offers a new strategy for complex task allocation in multiagent systems, providing an adaptive solution that selects suitable agents for diverse tasks, thereby enhancing system efficiency.

Efficient mission planning, including task allocation and path planning, is crucial for the successful operation of unmanned aerial vehicles (UAVs) and connected autonomous vehicles (CAVs) in complex scenarios. This article introduces an innovative mission planning approach that employs a collaborative model combining graph neural networks (GNNs) and Transformers to meet the intricate requirements of coordinating UAVs and CAVs. Our model excels in dynamic task allocation and accurate path planning, thereby boosting operational efficiency and reducing computational demands. We outline the shortcomings of current methods, notably their limited adaptability to dynamic changes and their substantial computational costs. By utilizing GNNs to capture complex interrelations and Transformers for effective information processing, our approach achieves greater adaptability and scalability. Experimental results demonstrate that our model surpasses leading methods, showing a 12% improvement in task allocation accuracy for UAVs and 10% for CAVs. Furthermore, we assess the model’s performance under various conditions, confirming its robustness and adaptability. This research provides a holistic solution for mission planning in UAV and CAV systems, setting the stage for future enhancements in autonomous vehicle coordination across logistics, surveillance, and disaster management sectors.

Multi-Agent Path Finding (MAPF) is essential to large-scale robotic systems. Recent methods have applied reinforcement learning (RL) to learn decentralized polices in partially observable environments. A fundamental challenge of obtaining collision-free policy is that agents need to learn co-operation to handle congested situations. This paper combines communication with deep Q-learning to provide a novel learning based method for MAPF, where agents achieve cooperation via graph convolution. To guide RL algorithm on long-horizon goal-oriented tasks, we embed the potential choices of shortest paths from single source as heuristic guidance instead of using a specific path as in most existing works. Our method treats each agent independently and trains the model from a single agent’s perspective. The final trained policy is applied to each agent for decentralized execution. The whole system is distributed during training and is trained under a curriculum learning strategy. Empirical evaluation in obstacle-rich environment indicates the high success rate with low average step of our method.