Simulation-based Reinforcement and Imitation Learning for Autonomous Sailboat Navigation in Variable Environmental Conditions

Abstract

In light of escalating concerns over climate change, harnessing oceanic data becomes increasingly urgent. Oceans serve as linchpins in understanding the intricate dynamics governing climate phenomena, exerting pivotal influence over global weather patterns and ecological systems. Despite scientific consensus on climate change impacts, including temperature shifts and acidification, a lack of data infrastructure hampers understanding marine ecosystems. This paper presents an adaptive control system for autonomous sailboats that aims to navigate efficiently in varied conditions by favoring the acquisition of oceanographic data. Combining reinforcement and imitation learning, the controller emulates human decision-making, enabling robust navigation. While showcasing promising results, challenges persist in adverse varied conditions. This challenge is exacerbated when confronting necessitating adaptive control mechanisms resilient to wear and tear. Simulators are vital for training due to the need for vast data volumes. Real-world data collection is costly and risky, while simulations accelerate learning. This study employs a simulator operating at ten times real-time speed, significantly simplifying scenario generation by adjusting factors such as sailboat position, orientation, target location, water current, and wind. Nevertheless, this novel approach signifies progress in addressing climate challenges and advancing oceanic research, using advanced computational methods

Keywords

• Reinforcement learning
• Imitation learning
• ROS
• Robotics
• Sailboats
• Autonomous water vehicles