Reinforcement learning-based tracking control of autonomous underwater vehicles for seafloor platform data collection

• Using AUVs as commuter platforms to shuttle between seafloor stations and the surface for near-real-time data collection. • The proposed tracking method establishes underwater optical communication links in three-dimensional space. • A reinforcement learning-based controller to control the commuter AUV, reducing pointing errors and energy consumption. • This method maintains pointing error of distance within 1.10 m in actual machine experiments. • This near-real-time data retrieval solution allows for more efficient monitoring of our marine environment. Real-time or near real-time access to seafloor platform observations will allow more efficient monitoring of our marine environment in response to possible changes. Data collection from the seafloor remains a time-consuming and costly endeavor due to communication limitations. In response to this challenge, we propose using an autonomous underwater vehicle (AUV) as a commuter vehicle to track the seafloor platform and establish a short-range optical communication link in three dimensions. A reinforcement learning-based tracking controller is trained in a simulated environment to minimize pointing errors and energy consumption during data retrieval. After experimental validation in a simulated environment, we deployed the tracking controller on the actual AUV Tri-TON and conducted sea experiments in Shimoda, Japan. The experimental results show that Tri-TON completed link pointing in 18 s, maintaining the pointing error of distance within 1.10 m, bearing angle error within 10.54 ∘ , and elevation angle error within 1.70 ∘ .