Fitness was assessed by the robot’s ability to approach the light. I can’t see a specification of whether this was something like continually decreasing distance to the light for a certain amount of time, or arrival at the light, or velocity to the light.

UPDATE: Bongard stated in an email that the fitness function was the average of several light sensors. If the average exceeded a pre-set threshold, the controller is considered successful. The higher the threshold, the closer the robot has to get to the light to count as successful. The threshold was fixed at a level at which most of the time the robots got about half way to the light.

Just curious as to how the body shape progressions were chosen, as that seems to be the one element here that was ‘imposed’ rather than evolved. Of course, that doesn’t diminish from the current finding, but perhaps how the shapes progress affects the results? In line with the thought that the more complicated shapes have the simper controllers embedded in them, maybe a shape progression where the basic forms are more similar to a disabled final form would be even more stable?

On a separate tangent, I’m just curious – do you know how the relative success of each controller was judged? Was it merely the absolute distance, or were there factors for, say, average movement speed in any direction, or similar?