A comparison of DRL with APF and A* with PI Control for Trajectory Planning with Obstacle Avoidance for Sailboat Robots

• André Araújo UFF
• Vishali Mankina Univ. of Salento
• Carlo Cernicchiaro CMMI Cyprus Marine and Maritime Institute
• Philippe Preux Centre Inria de l‘Université de Lille
• Cosimo Distante CNR, Inst. of Applied Sciences and Intelligent Systems
• Charles Vasconcellos UFF
• Diego Brandão CEFET-RJ
• Luiz Gonçalves UFRN
• Esteban Clua UFF

Abstract

Autonomous sailboats present unique challenges for Unmanned Surface Vessel (USV) research due to the dynamic nature of maritime environments, where variables like wind direction, ocean currents, and obstacles continuously change. This paper provides a detailed comparison of two advanced trajectory planning methods: Deep Reinforcement Learning (DRL) with Artificial Potential Fields (APF) and A* with Proportional-Integral (PI) control. Both techniques are evaluated using a high-fidelity simulation environment in Gazebo, specifically designed to model realistic maritime conditions, including varying wind speeds and moving obstacles. The performance of each approach is measured in terms of obstacle avoidance, trajectory accuracy, and energy efficiency. Quantitatively, DRL with APF reduced navigation time to 4 minutes and 49 seconds, while A* with PI control required 20 minutes and 10 seconds, achieving success rates of 92% and 75%, respectively. The results demonstrate that DRL with APF not only provides smoother and more adaptive trajectories but also significantly improves energy efficiency, making it a more robust and flexible solution for dynamic and uncertain maritime environments. These findings suggest that DRL with APF holds great potential for applications in autonomous maritime navigation, particularly in scenarios that demand real-time adaptability and efficiency.

Keywords: Sea surface, Trajectory planning, Navigation, Wind speed, Sea measurements, Energy efficiency, Trajectory, Collision avoidance, Robots, Underactuated surface vessels, Unmanned Surface Vessel, Autonomous Sailboat, Reinforcement Learning, Artificial Potential Field, A*, PI