Abstract
Collision avoidance for Maritime Autonomous Surface Ships (MASS) remains challenging due to the need for safe, COLREGs-compliant decisions in close encounters. This study proposes a human-centric collision avoidance framework based on a computational intelligence approach called Adaptive Neuro-Fuzzy Inference Systems (ANFIS), trained using high-fidelity bridge simulator data that capture realistic navigator behavior. The fuzzy inference structure is generated using Fuzzy C-Means (FCM) clustering, with the optimal number of clusters determined using the Fuzzy Partition Coefficient (FPC) and Xie-Beni (XB) index.
Separate ANFIS models are developed for crossing, head-on, and overtaking scenarios, and the effectiveness of the proposed method is validated by closed-loop validations using a second-order Nomoto model. Analysis of three simulation cases demonstrates that ANFIS models generate smooth and stable maneuvers, achieve collision free trajectories with safe passing distances, and comply with COLREGs. Moreover, the predicted rudder commands closely resemble human navigator actions, indicating the capability of the approach to capture underlying patterns.
A global sensitivity analysis based on Sobol indices is conducted to quantify the influence of input variables on the ANFIS model output. The results show that different encounter scenarios are governed by distinct dominant features, with crossing and overtaking primarily influenced by geometric variables, while the head-on model exhibits stronger interaction effects among multiple inputs.