Hierarchical reinforcement learning for cooperative payload-constrained task allocation and path planning of multi-airship in dynamic environments

The expansion of stratospheric airship applications in fields such as ocean monitoring has raised higher requirements for the cooperative operation capability of multiple airships. Most existing methods only optimize task allocation or path planning separately. They cannot handle the decision conflicts caused by the coupling of these two aspects in dynamic task scenarios. Other challenges include the high-dimensional action space and the matching constraints between different task types and payloads. Hierarchical reinforcement learning, with its ability to decompose tasks and perform hierarchical decision-making, provides an effective approach to solving such problems. Therefore, this paper proposes a hierarchical multi-agent dynamic task planning algorithm based on hierarchical reinforcement learning. The algorithm adopts a two-layer control architecture. The high-level controller is responsible for cooperative task allocation at the global level, while the low-level controller performs path planning and action control. The two controllers work together through iterative interaction to achieve the global task objectives. In addition, a priority-based conflict-resolution mechanism is introduced at the task-allocation layer to improve allocation efficiency and support algorithm convergence. Simulation results show that the proposed H-MADTP algorithm achieves a task success rate of 90% in dynamic task environments, compared to 79% for the auction-based algorithm (AA), 59% for QMIX, and 47% for MATD3.