This paper presents a black-box modeling approach, locally weighted learning (LWL), for tug handling simulator. Concerned with the problem of parameter drifting and unmodeled dynamics in the conventional mechanism modeling, LWL is proposed for Delta Linda tug to learn the mapping between the input and output directly and employs the empirical regression model instead of the real physical model. Compared with BPNN prediction, the validity and usefulness of the algorithm are illustrated with a 3 degree-of-freedom of Delta Linda tug and LWL is proved to be an accurate manoeuvring modeling tool.