Heterogeneous multi-AUV cooperative area search with an LLM-assisted hierarchical planning framework

In heterogeneous multi-AUV cooperative area search, existing methods often decompose the task region before assigning sub-regions to individual AUVs. However, coverage costs associated with platform capability differences, time-energy coupling, and temporal coordination constraints are often insufficiently considered during region decomposition, which may lead to suboptimal partitions and capability-task mismatches. To better account for these factors, a Voronoi-based region parameterization method is developed to formulate region decomposition and task allocation as a continuous optimization problem under coupled constraints. An improved CMA-ES-based optimization method, incorporating weighted K-means initialization and cost-guided directional sampling, is proposed to solve the resulting non-convex and non-smooth black-box problem. Furthermore, motivated by advances in large language models (LLMs), this paper establishes an LLM-assisted hierarchical planning framework in which operators’ natural-language task intentions are mapped to the inputs and constraints of the optimization method, enabling flexible task-level planning and pre-deployment iterative plan refinement. Simulation results show that the proposed framework effectively balances task completion time, temporal coordination, and energy utilization, while supporting iterative refinement of plans prior to deployment. These results highlight the potential of integrating LLM-based task interpretation with optimization-based planning for cooperative underwater robotic systems.