The training data contain series of images induced by given motor commands. The robot was put within a circle with an inner diameter of 180±2 cm formed by red bricks (figure 7.1, left).
Random velocities were chosen for the left and right wheels individually (vL and vR). The velocities ranged from -60 mm/sec to 60 mm/sec in steps of 20 mm/sec. The combination with both velocities being zero was discarded.
After a set of velocities was chosen, the robot maintained the given speed. Every two seconds, an image was recorded from the camera (figure 7.3, left) and stored. Recording started after granting the robot a one second acceleration phase. The movement lasted up to a maximum of six shots (five 2 sec intervals) or until the robot got too close to one of the obstacles (this was determined with the help of the same kind of image preprocessing as described in section 7.2.3). In both cases, a new combination of velocities was chosen, and a new recording series started.
In the second case, however, the choice of velocities was restricted. The robot was only allowed to go either forward or backward, depending on if the obstacle was in the back or front. Forward movements were chosen randomly from a subset that fulfills vLvR 0 and vL + vR > 0. Backward movements were chosen in an analogue way ( vLvR 0 and vL + vR < 0).
This data-collection scheme would result in more forward and backward movements against rotational movements (which fulfill vLvR < 0) because the robot was not allowed to do turns when it was close to obstacles. Therefore, when a rotation was possible, the rotational movements were chosen with a higher probability to adjust toward a balanced distribution of velocity combinations.
The actual wheel velocity was recorded during the 2 sec intervals. If it deviated by more than 10 mm/sec from the given value the series was stopped, and the interval's was data discarded. After that, a new series started, as above. The robot was able to pursue this kind of random exploration automatically without getting into physical contact with any obstacle.
In total, 5466 intervals with 6808 images were recorded. The velocities were roughly evenly distributed (figure 7.2). Straight movements (vL = vR) slightly dominated, and there were fewer slow turns like vL = 0 and vR = 20 mm/sec. The explanation is that the lateral friction between wheels and floor made the robot occasionally stick to the ground during slow turns, so these movements were discarded.