A Multi-USV-UAV Environmental Exploration Method Based on Cross-Modal Attention Fusion

Abstract

To address the issues of inefficiency and susceptibility to local optima caused by the limited local perception of multiple unmanned surface vehicles (USVs) during collaborative exploration in unknown waters, this paper proposes a multi-USV-UAV collaborative exploration framework. This framework combines the decision-making capability of deep reinforcement learning with trajectory optimization methods. The core lies in the design of a policy network based on a cross-modal attention mechanism, which effectively fuses the global prior images provided by unmanned aerial vehicles (UAVs) with the local observation information of the USVs. This fusion provides USVs with prior knowledge of the macro environment, enabling them to make more forward-looking waypoint planning decisions and avoid redundant exploration. On this basis, a trajectory optimization problem incorporating safety, smoothness, and dynamic feasibility is further formulated and solved using the MINCO framework, generating high-quality execution trajectories. Experimental results demonstrate that the proposed method significantly enhances the exploration efficiency of USV clusters in complex aquatic environments.