You're correct, but aren't the robots:
a) observing the scene;
b) recreating the scene/trajectory as they see it; and
c) watching again and repeating
?
1) In the case of unlimited trials what will happen is that fewer clusters will be formed because the robots will follow the "average" trajectory of noise
2) In the case of limited trials more clusters will form (and they will be closer to the original trajectory because the amount of noise contributing to the trajectory is less) and the trajectories they follow will more closely resemble the original "course"
Maybe I'm missing something important. I'll read the paper again.