Our control architecture is shown in Figure 1. It is based on the Dual Dynamics scheme developed by H. Jäger [3,4]. The robots are controlled in closed loops that use different time scales and that correspond to behaviors on different levels of the hierarchy.
We extend the Dual Dynamics concept by introducing a third element, namely the perceptual dynamics, as shown on the left side of the drawing. Here, either slow changing physical sensors, such as the charging state indicators of the batteries, are plugged-in at the higher levels, or the readings of fast changing sensors, like the ball position, are aggregated by dynamic processes into slower and longer lasting percepts. The boxes shown in the figure are divided into cells. Each cell represents a sensor value that is constant for a time step. The rows correspond to different sensors and the columns show the time advancing from left to right.
A set of behaviors is shown in the middle of each level. Each row contains an activation factor from the interval [0,1] that determines when the corresponding behavior is allowed to influence actuators.
The actuator values are shown on the right hand side. Some of these values are connected to physical actuators that modify the environment. The other actuators influence lower levels of the hierarchy or generate sensory percepts in the next time step via the internal feedback loop.
Since we use temporal subsampling, we can afford to implement an increasing number of sensors, behaviors, and actuators in the higher layers without an explosion of computational cost. This leads to rich interactions with the environment.
Each physical sensor or actuator can only be connected to one level of the hierarchy. One can use the typical speed of the change of sensor readings to decide where to connect a sensor. Similarly, the placement of actuators is determined by the time constant they need to produce a change in the environment. Behaviors are placed on the level that is low enough to ensure a timely response to stimuli, but that is high enough to provide the necessary aggregated perceptual information, and that contains actuators which are abstract enough to produce the desired reactions.