There are several objects in the world, such as the goals and the field markers which remain stationary and can be used for self-localization. Mobile objects are the agent itself, the ball, and 21 other players (10 teammates and 11 opponents). These objects are represented in a type hierarchy as illustrated in Figure 1.
Figure 1: The agent's object type hierarchy.
Each agent's world model stores an instantiation of a stationary object for each goal, sideline, and field marker; a ball object for the ball; and 21 player objects. Since players can be seen without their associated team and/or uniform number, the player objects are not identified with particular individual players. Instead, the variables for team and uniform number can be filled in as they become known.
Mobile objects are stored with confidence values within [0,1] indicating the confidence with which their locations are known. The confidence values are needed because of the large amount of hidden state in the world: no object is seen consistently. While it would be a mistake to only remember objects that are currently in view, it is also wrong to assume that a mobile object will stay still (or continue moving with the same velocity) indefinitely. By decaying the confidence in unseen objects over time, agents can determine whether or not to rely on the position and velocity values .
All information is stored as global coordinates even though both sensor and actuator commands are specified in relative coordinates (angles and distances relative to the agent's position on the field). Global coordinates are easier to store and maintain as the agent moves around the field because the global coordinates of stationary objects do not change as the agent moves, while the relative coordinates do. It is a simple geometric calculation to convert the global coordinates to relative coordinates on demand as long as the agent knows its own position on the field.
The variables associated with each object type are as follows: