In high-end aerospace mission scenarios such as distributed sensing and tactical reconnaissance, path planning for multi-UAV cooperative formation flight has become an indispensable core component of mission systems. However, a key unresolved technical bottleneck remains: how to achieve an optimal balance between strictly maintaining the UAV formation’s preset geometric configuration and realizing collision-free, flexible flight in dynamically changing complex environments. Multi-Agent Reinforcement Learning (MARL) can theoretically solve this coupled optimization problem, but the control instructions generated by the existing mainstream MARL methods often have high-frequency fluctuation issues. Direct deployment of these fluctuating commands on physical UAV hardware will increase alternating stress on mechanical structures, raise the risk of motor overheating, and significantly boost battery energy consumption. These problems directly restrict the engineering application of such methods.
To address the above challenges, this paper proposes a robust control framework that accounts for UAV physical characteristics to solve the model mismatch between simulation environments and real flight scenarios during carrier-based UAV deployment. Firstly, to improve the learning efficiency of the algorithm and avoid the cold start problem in the initial training stage, this paper introduces the Adaptive Curriculum Learning (ACL) method. Unlike the fixed-difficulty training mode, ACL can dynamically adjust the environmental complexity based on the real-time task performance of the UAVs. It first allows agents to master basic formation maintenance skills, before training is gradually transitioned to more complex tasks such as navigation in intricate environments. Additionally, considering the computational limitations of the onboard embedded processors of UAVs, this paper only integrates the Graph Attention Network (GAT) into the centralized Critic network of the centralized training and decentralized execution (CTDE) framework. The core advantage of this design is that the system no longer solely relies on distance as the judgment basis but assesses the influence of adjacent UAVs based on the collision risk level. This not only improves the accuracy of value estimation, but also does not compromise the response speed of local control instructions- a feature that is crucial for the real-time requirements of UAV flight control.
The proposed method was tested and validated in a high-fidelity simulation environment built in strict compliance with UAV kinematic constraints. The experimental results fully demonstrate its superiority: in dense obstacle areas, the success rate of the task reached 80.4%; compared with the aggressive baseline method, the jerk was reduced by 10.9%. Most importantly, the flight trajectories generated by this framework are smooth and adaptable to actual flight conditions, and the corresponding control commands can be directly deployed on real flight controllers without additional parameter adjustments. This result demonstrates that the proposed framework achieves substantial technical progress in enhancing the autonomous response capabilities of UAV clusters to complex scenarios.