Autonomous Underwater Vehicle Docking Under Realistic Assumptions Using Deep Reinforcement Learning
This paper addresses the challenge of docking an Autonomous Underwater Vehicle (AUV) under realistic conditions. Traditional model-based controllers are often constrained by the complexity and variability of the ocean environment. To overcome these limitations, we propose a Deep Reinforcement Learni...
| Authors: | , |
|---|---|
| Format: | article |
| Status: | Published version |
| Publication Date: | 2024 |
| Country: | España |
| Institution: | Varias* (Consorci de Biblioteques Universitáries de Catalunya, Centre de Serveis Científics i Acadèmics de Catalunya) |
| Repository: | Recercat. Dipósit de la Recerca de Catalunya |
| OAI Identifier: | oai:recercat.cat:10256/25812 |
| Online Access: | http://hdl.handle.net/10256/25812 |
| Access Level: | Open access |
| Keyword: | Aprenentatge profund Deep learning Aprenentatge automàtic Machine learning Aprenentatge per reforç Reinforcement learning Vehicles submergibles autònoms Autonomous underwater vehicles Vehicles submergibles -- Sistemes de control Submersibles -- Control systems |
| Summary: | This paper addresses the challenge of docking an Autonomous Underwater Vehicle (AUV) under realistic conditions. Traditional model-based controllers are often constrained by the complexity and variability of the ocean environment. To overcome these limitations, we propose a Deep Reinforcement Learning (DRL) approach to manage the homing and docking maneuver. First, we define the proposed docking task in terms of its observations, actions, and reward function, aiming to bridge the gap between theoretical DRL research and docking algorithms tested on real vehicles. Additionally, we introduce a novel observation space that combines raw noisy observations with filtered data obtained using an Extended Kalman Filter (EKF). We demonstrate the effectiveness of this approach through simulations with various DRL algorithms, showing that the proposed observations can produce stable policies in fewer learning steps, outperforming not only traditional control methods but also policies obtained by the same DRL algorithms in noise-free environments |
|---|