USV pursuit–evasion using a complementary scientific machine learning with control barrier functions approach

Abstract

The maritime pursuit–evasion problem is increasingly relevant to autonomous robotics and naval operations, particularly for security, surveillance, search and rescue, and environmental monitoring. Effective pursuit requires accurate evader behavior prediction combined with robust obstacle avoidance in cluttered maritime environments. Traditional methods, including differential game theory and heuristic planning, often neglect realistic complexities and provide limited safety guarantees. Recent reinforcement learning approaches improve flexibility but struggle with generalization and formal safety assurance in complex scenarios. To bridge this gap, we propose a novel integration of scientific machine learning with control barrier functions, enabling provably safe pursuit and navigation under realistic vessel dynamics, partial observability, and nonconvex obstacle constraints. Simulations validate ability of the proposed methods to achieve safe and effective pursuit in challenging maritime environments.