The Soccer Server system [5] used at RoboCup-97 [2] provides a rich and challenging multiagent, real-time domain. Sensing and acting is noisy, while inter-agent communication is unreliable and low-bandwidth.
In order to be successful, each agent in a team must be able to sense and act in real time: sensations arrive at unpredictable intervals while actions are possible every 100ms. Furthermore, since the agents get local, noisy sensory information, they must have a method of converting their sensory inputs into a good world model.
Action capabilities range from low-level individual skills, such as moving to a point or kicking the ball, to high-level strategic collaborative and adversarial reasoning. Agents must be able to act autonomously, while working together with teammates towards their common overall goal. Since communication is unreliable and perception is incomplete, centralized control is impossible.
This article presents the CMUnited-97 approaches to the above challenges which helped the team to the semifinals of the 29-team RoboCup-97 simulator tournament. Section 2 introduces our overall agent architecture which allows for team coordination. Section 3 presents our agents' world model in an uncertain environment with lots of hidden state. Section 4 lays out the agents' hierarchical behavior structure that allows for machine learning at all levels of behavior from individual to collaborative to adversarial. Our team's flexible teamwork structure, which was also used by the CMUnited-97 small-size robot team [7], is presented in Section 5. Section 6 concludes.