Autonomous sailing with sim-to-real reinforcement learning
Kiki J.A. Bink (Maritime Research Institute Netherlands (MARIN), Student TU Delft)
Bülent Düz (Maritime Research Institute Netherlands (MARIN))
Gabriel D. Weymouth (TU Delft - Mechanical Engineering)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Autonomous sailing offers a sustainable alternative for reducing greenhouse gas emissions in maritime transport, aligning with global environmental targets. This study explores the application of reinforcement learning (RL) to autonomous sailing, addressing challenges in handling dynamic and unpredictable environmental conditions. Leveraging a sim-to-real transfer methodology, RL agents were trained in a simulation environment with the domain randomization technique to enhance adaptability and robustness, and tested in real-world scenarios using a robotic sailboat in the Offshore Basin at MARIN. The study quantified the reality gap between simulation and real-world environments, identifying key discrepancies in actuator latency and simulation modeling accuracy. In real-world basin experiments, the best-performing RL agent successfully completed the course in 12 out of 12 runs. Unlike conventional controllers, the trained agents demonstrated enhanced sailing capabilities like roll tacking and recovery from wind-stalled conditions. This work advances the understanding of autonomous sailing control and highlights pathways to bridge the reality gap, contributing to the broader adoption of RL in dynamic real-world applications.
help_outline