Learning Representative Vessel Trajectories Using Behavioral Cloning

The data-driven method consisted in predicting trajectories by mimicking the underlying policy of human captains. The decisions made by these experts are recorded by AIS signals and can be fused with factors like weather conditions, current tide level, ship destination, etc. to get a comprehensive definition of the state at each time.

The motivation was paving the way for near real-time forecasts of vessel trajectories from a given snapshot of the situation, instead of the costly history of the potentially large number of vessels present in a scene.