The Dual Dynamics control architecture, proposed by Jäger [3], describes reactive behaviors in a hierarchy of control processes. Each layer of the system is partitioned into two modules: the activation dynamics that determines whether or not a behavior tries to influence actuators, and the target dynamics, that determines strength and direction of that influence. The different levels of the hierarchy correspond to different time scales. The higher level behaviors configure the lower level control loops via activation factors that determine the mode in which the primitive behaviors are. These can produce qualitatively different reactions if the agent encounters the same stimulus again, but has changed its mode due to stimuli that it saw in the meantime.
A more detailed description of our control architecture, that is based on these ideas, is given in [1]. The robots are controlled in closed loops that use different time scales. We extend the Dual Dynamics scheme by introducing a third dynamics, the perceptual dynamics. Here, either slow changing physical sensors are plugged in at higher levels, or readings of fast changing sensors, like the ball position, are aggregated to slower and longer lasting percepts. Since we use temporal subsampling, we can afford to implement an increasing number of sensors, behaviors and actuators in the higher layers.
Behaviors are constructed bottom up: First, processes that react quickly to fast stimuli are designed. Their critical parameters, e.g. a target position, are determined. When the fast behaviors work reliably, the next level can be added. This level can now influence the environment either directly by moving slow actuators or indirectly by changing critical parameters in the lower level.
In the lowest level of the field player only two behaviors are implemented. These are a parameterized taxis behavior and an obstacle avoidance behavior.
On the next level various behaviors use them. They approach the ball, dribble with the ball, kick the ball, free the ball from corners, home, and so on. Here, we also implemented a ball prediction that allows for anticipative actions. On the third level, we dynamically adjust the home positions of the players, such that they block opponent players or position themselves freely to receive passes.
Each of our robots is controlled autonomously from the lower levels of the hierarchy using a local view to the world. For instance, we present the angle and the distance to the ball and the nearest obstacle to each agent. In the upper layers of the control system the focus changes. Now we regard the team as the individual. It has a slow changing global view to the playground and coordinates the robots as its extremities to reach strategic goals.
In the first team level, we decide which robot should take the initiative and go for the ball. The remaining players are assigned to supporting roles. Passing and strategy changes would be implemented in higher team layers.
