In his master’s thesis, Nils Petter Alfsen, in collaboration with the Norwegian Maritime AI Centre and Kongsberg Maritime, investigated how an AI model can predict the movements of nearby vessels and help autonomous ships chart a safe course.

Ship captains and vessel traffic service (VTS) operators draw on extensive experience to interpret and anticipate vessel traffic, taking into account local traffic patterns, weather conditions and grounding risks. For autonomous ships, this knowledge must be built into software and algorithms.

Alfsen used historical Automatic Identification System (AIS) data, which include information about vessel position, course and speed, to train a transformer model on vessel traffic in the area where ferries between Horten and Moss cross the Oslo Fjord in Norway.

The model uses 30 minutes of AIS history from nearby vessels to predict a set of possible trajectories over the next ten minutes. The aim is to give the autonomous vessel controller a better understanding of vessel interactions and a stronger basis for selecting course and speed by considering multiple possible developments in the surrounding traffic.

The illustration shows two vessels, S1 and S2, with their trajectories over the preceding 30 minutes (black dashed lines), the model’s three predicted trajectories for the next ten minutes ranked by probability (coloured lines), and the vessels’ actual trajectories over the same prediction period (blue dashed lines).

Simulations show that this type of trajectory prediction can improve autonomous collision avoidance in coastal traffic. The results indicate significant potential for combining AI-based AIS trajectory prediction with probabilistic collision-avoidance control, which accounts for uncertainty in the future movements of other vessels. Compared with a constant-velocity predictor, which assumes that other vessels maintain their current course and speed, Alfsen’s AI model helped the controller make better manoeuvring decisions in several of the simulated situations.

The model is lightweight and can run locally on board with modest computing resources. The model and code are available to MAI partners for further research.

The thesis was supervised by Tor Arne Johansen, Centre Director at NTNU.

Read the master’s thesis in the Norwegian Research Information Repository (NVA).