For an off-line evaluation of the anticipation performance, test data were collected using a slightly different random exploration scheme. The goal was to get random movement sequences instead of series with a constant motor command.
For each interval, a new random velocity combination was used. A recording series consisted of eight 2 sec intervals starting from zero velocity. The first interval was discarded, thus leaving seven intervals each under identical conditions. The wheel velocity was monitored in the last 0.7 sec of each interval. If its mean value deviated by more then 10 mm/sec from the given value the whole recording series was discarded. Therefore, the limited acceleration required to make the random choice of velocities slightly dependent on the previous choice. Absolute velocity changes of more than 60 mm/sec for each wheel were not allowed.
The choice of velocities in each interval further depends on the encounters with obstacles. If the obstacle was in the front the robot moved backward, and if it was in the back the robot moved forward (same way as done for the training data). Additionally, the robot responded to an obstacle on the left or right side by turning the front of the robot toward the obstacle (choosing | vL| > | vR|, vL > 0, and vR < 0 for obstacles on the right side, and accordingly for the left side, changing roles of vL and vR). The turn toward the obstacle was then followed by a backward movement.
Such a response to the obstacles allowed the rear part of the robot to keep a larger distance to the circle. This was necessary since the mirror was located in the frontal part of the robot (see figure 7.1), and thus the bottom part of an obstacle close to the rear part of the robot was occluded. Totally, 138 test series were recorded, with a total of 966 intervals and 1104 images.