Scholarlink

With the increasing scale and complexity of global maritime traffic, ensuring the safety of autonomous vessel navigation has become a critical challenge. This paper presents a deep reinforcement learning (DRL) approach for autonomous maritime collision avoidance, with a focus on ensuring safety under both nominal and adversarial conditions. A policy is trained using local observations of surrounding vessels to generate COLREGs-compliant maneuvers in decentralized multi-agent scenarios. The method is evaluated in diverse encounter geometries inspired by the Imazu problem set, demonstrating the agent’s ability to generalize to unseen head-on, crossing, and overtaking situations. To enhance robustness against positioning interference, we introduce an anomaly detection mechanism based on Inertial Navigation System estimation. During GPS spoofing attacks, the system compares GPS and INS position estimates, and penalizes discrepancies in the reward function, enabling the agent to identify and mitigate spoofed signals without relying on external supervision. Experimental results across multiple scenarios confirm the agent’s ability to preserve safe trajectories and avoid collisions, even under sensor-level adversarial attacks.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Deep Reinforcement Learning Framework for Robust Maritime Collision Avoidance Under GPS Spoofing

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ying Ding,
  • Weizhi Meng,
  • Shaoming He,
  • Wenjuan Li

摘要

With the increasing scale and complexity of global maritime traffic, ensuring the safety of autonomous vessel navigation has become a critical challenge. This paper presents a deep reinforcement learning (DRL) approach for autonomous maritime collision avoidance, with a focus on ensuring safety under both nominal and adversarial conditions. A policy is trained using local observations of surrounding vessels to generate COLREGs-compliant maneuvers in decentralized multi-agent scenarios. The method is evaluated in diverse encounter geometries inspired by the Imazu problem set, demonstrating the agent’s ability to generalize to unseen head-on, crossing, and overtaking situations. To enhance robustness against positioning interference, we introduce an anomaly detection mechanism based on Inertial Navigation System estimation. During GPS spoofing attacks, the system compares GPS and INS position estimates, and penalizes discrepancies in the reward function, enabling the agent to identify and mitigate spoofed signals without relying on external supervision. Experimental results across multiple scenarios confirm the agent’s ability to preserve safe trajectories and avoid collisions, even under sensor-level adversarial attacks.