It predicts the actual positions of the robots

[chichi]

The time elapsed between making an action decision and perceiving the conse- quences of that action in the environment is called the control delay. All physical feedback control loops have a certain delay, depending on the system itself, on the input and output speed and, of course, on the speed at which the system processes information.

In the RoboCup small size league a global vision module is used to determine the positions of all robots on the field. In order to control the behavior of the robots in a way that is appropriate for the situation on the field we need the exact positions of them at every moment. Because of the delay inherent in the control loop, however, the behavior control system actually reacts to the environment four frames ago (about 132 ms). When moving faster – our robots drive at a speed of up to 2 m/s – the delay becomes more significant as the error between the real position and the position used for control grows up to 20 cm.

In order to correct this immanent error we have developed a neural network which processes the positions, orientations, and the action commands sent to the robots during the last six frames. It predicts the actual positions of the robots. These predictions are used as a basis for control. We use real recorded preprocessed data of moving robots to train the network.

The concept of motor prediction was first introduced by Helmholtz when trying to understand how humans localize visual objects (see [6]). His suggestion was that the brain predicts the gaze position of the eye, rather than sensing it. In his model the predictions are based on a copy of the motor commands acting on the eye muscles. In effect, the gaze position of the eye is made available before sensory signals become available.

The paper is organized as follows. The next section gives a brief http://www.fifa15sale.com description to our system architecture. Then we explain how the delay is measured and we present some other approaches to eliminate the dead time. Section 4 describes architecture and training of the neural network used as a predictor. Finally, we present some experimental results and some plans for future research.