FIGURE 2.

Snapshots of the three simulated real-world environments. Here, the paths and the landmasses are fixed for each instance of the environment, though each instance samples a random subset of the historical AIS data to create the dynamic obstacles. Thus, the dynamic obstacles are initialized with an initial position and velocity as shown with the black arrows and will follow the trajectories shown with the red dashed lines. As in the training environments, the agent’s goal is to safely navigate the vessel along the path from the starting position to the goal.