Adaptive collision avoidance strategy for USVs in perception-limited environments using dynamic priority guidance

Abstract

This paper proposes a dynamic adaptive priority guidance (DAPG) strategy for unmanned surface vehicles (USVs) to improve collision avoidance in dynamic maritime environments, particularly in unpredictable moving obstacles. Traditional local navigation methods often depend on fixed parameters within their cost functions, limiting adaptability. In contrast, the DAPG strategy integrates the strengths of multi-agent reinforcement learning (MARL) and multi-source information fusion strategy (MIFS). At a high level, the MARL-based algorithm dynamically adjusts fusion weights using neural networks, enabling the system to adapt flexibly to changing environments. At the low level, the MIFS algorithm processes these prioritized observations to generate the optimal navigation commands, ensuring safe and efficient navigation for each USV. The network is trained in a simulated dynamic environment using the parameter-sharing soft actor-critic (PSSAC) algorithm, enhanced with prioritized experience replay (PER) to accelerate learning. Experimental results show that the PSSAC-PER-MIFS algorithm significantly outperforms traditional reinforcement learning methods in convergence speed, reward stability, and navigation efficiency. Moreover, the DAPG strategy ensures compliance with COLREGs (International Regulations for Preventing Collisions at Sea), facilitating smooth and cooperative navigation in multi-USV scenarios. The source code is available at https://github.com/Shihong-Yin/PSSAC-PER-MIFS.