Imitation Learning for Collision Avoidance in Unmanned Surface Vehicles via Behavioral Cloning from a Deterministic Planner

This paper presents the application of imitation learning—specifically, Behavioral Cloning—to collision avoidance and path planning for Unmanned Surface Vehicles. The proposed approach enables machine learning models to learn navigation policies directly from expert demonstrations generated by a deterministic collision-avoidance algorithm. In this framework, the expert algorithm provides consistent and interpretable decision-making aligned with maritime navigation rules, while the learning-based model aims to approximate this behavior in a data-driven manner. A supervised learning pipeline is developed in which state–action pairs are extracted from simulated encounter scenarios and used to train multiple machine learning models. The study focuses on tree-based ensemble methods, including Random Forest, Extremely Randomized Trees (Extra Trees), XGBoost, and CatBoost, due to their robustness and suitability for nonlinear regression tasks. The trained models are evaluated on a set of representative single-target-ship encounter scenarios to assess their ability to reproduce expert decisions. The results demonstrate that Behavioral Cloning can effectively capture the underlying decision policy of the expert system, achieving high agreement with the reference algorithm. However, the present study is limited to simulated single-target encounters and does not explicitly account for factors such as multi-ship encounters, sensor noise, communication or perception delays, or environmental disturbances. Therefore, the results should be regarded as an initial assessment of the feasibility of the proposed learning-based approach rather than comprehensive validation for practical autonomous navigation. Future work will address these limitations and investigate the approach under more diverse and realistic operating conditions.