Deep learning based autonomous navigation planning in dynamic ice fields

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

School of Science | Master's thesis

Department

Mcode

Language

en

Pages

69

Series

Abstract

Autonomous navigation in Arctic seas is a difficult task due to the highly dynamic and uncertain nature of the environment. Traditional route planning methods often work on a global scale and do not consider the local ice conditions, and the ones that do, either assume static ice conditions or rely on heuristic approximations of collision severity, limiting their applicability in real-world scenarios. This thesis proposes Icenet, a deep learning predictive model integrated with a lattice-based planner for autonomous ship navigation in dynamic ice fields. The neural network is trained on large-scale simulations where ice floe fields with varying concentration levels are generated. By incorporating multi-channel inputs created from combining different ice floe features such as occupancy, thickness, speed, and the location and motion of the ship, the network learns to predict both future ice configurations in the surrounding area and the associated collision cost of the ship-ice interaction. Evaluation on 1000 test trials for each concentration show that Icenet achieves high similarity in ice map prediction (SSIM ≈ 0.93) and collision cost correlation (0.73). Integrated into a lattice-based planner, Icenet outperforms other state-of-the-art planners in terms of path length and collision cost metrics, especially in high ice concentration scenarios.

Description

Supervisor

Polojärvi, Arttu

Thesis advisor

Laaksonen, Jorma
Repin, Roman

Other note

Citation