Ball Handling

One of the most important decisions in the robotic soccer domain arises when the agent has control of the ball. In this state, it has the options of dribbling the ball in any direction, passing to any teammate, shooting the ball, clearing the ball, or simply controlling the ball.

In CMUnited-98, the agent uses a complex heuristic decision mechanism, incorporating a machine learning module, to choose its action. The best teammate to receive a potential pass (called potential receiver below) is determined by a decision tree trained off-line [9]. Following is a rough sketch of the decision-making process without all of the parametric details.

To begin with, since kicks (i.e. shots, passes, and clears) can take several cycles to complete (Section 4.1), the agent remembers the goal of a previously started kick and continues executing it. When no kick is in progress (do the first that applies):

• If close to the opponent's goal and no defenders are blocking the path to the goal (defined as a cone with vertex at the ball): shoot or dribble based on the goaltender's position, the position of the closest opponent, and the distance to the goal.
• At the other extreme, if close to the agent's own goal and there is an opponent nearby: clear the ball.
• If approaching the line of the last opponent defender: dribble the ball forward if possible; otherwise send the ball (clear) past the defender.
• If the potential receiver is closer to the goal and has a clear shot: pass to the potential receiver.
• If no opponents are in the direct path to the goal: dribble to the goal.
• If fairly close to the opponent's goal and there is at most one opponent in front of the goal: shoot.
• If no opponents are in the way of one of the corner flags: dribble towards the corner flag.
• If there is a potential receiver: pass.
• If it's possible to hold onto the ball without moving (at most one opponent is nearby): hold the ball.
• Otherwise: Kick the ball away (clear).