Session: 08-01-03 AI Driven Autonomous Navigation, Collision Avoidance & Optimization

Submission Number: 181931

Development of Explainable Ai Models Through Behavior Tree Transformation of Asv Collision Avoidance Policies 

Recent advances in deep learning techniques and hardware technology have significantly expanded the use of neural network-based approaches for collision avoidance in Autonomous Surface Vehicles (ASVs). Among these applications, vessels operating on Arctic routes face the critical challenge of avoiding moving obstacles such as icebergs and large ice floes, making robust collision avoidance technology essential for enabling safe autonomous navigation in these environments. While these deep learning models demonstrate strong performance, their black-box nature impedes understanding of internal decision-making processes, thereby constraining opportunities for modification and redesign. This research aims to transform ASV collision avoidance policies into human-interpretable representations through explainable artificial intelligence (XAI) techniques. The proposed methodology converts policies trained using the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm into decision trees and subsequently reconstructs them into behavior trees. This transformation enables domain experts to intuitively comprehend and modify the collision avoidance policies. Furthermore, the research establishes a bidirectional learning framework that utilizes modified behavior trees to retrain the original neural network through imitation learning, effectively integrating human expert knowledge with AI capabilities. Experiments conducted in simulation environments validated the interpretability and modifiability of both the decision tree and the behavior tree representations, confirming that policy performance is preserved throughout the transformation process. This research presents a transparent artificial intelligence framework applicable to ASV collision avoidance across diverse maritime environments, including Arctic routes.

Presenting Author: Jaeyoon Jeon Inha University

Presenting Author Biography: Jaeyoon Jeon is a graduate student in the Department of Naval Architecture and Ocean Engineering at Inha University, specializing in autonomous ship control and reinforcement learning. His research interests include ship maneuvering, collision avoidance systems, and the application of machine learning in marine environments.

Authors:

Jaeyoon Jeon Inha University
Sanghyun Kim Inha University
Jumyeong Lee Inha University
Hyeongseok Yoon Inha University
Juhyeong Oh Inha University
Jisoo Han Inha University