*5.4. Aggregate Results*

TuttiFrutti and EvoColor obtain similar results when the control software is evaluated with simulations. On the other hand, TuttiFrutti is significantly better than EvoColor when the control software is ported to the physical robots. It has already been pointed out that when control software developed in simulation is ported to a real-world platform, due to the reality gap one might observe both a drop in performance [53] and a substantial modification of the collective behavior [69]. The entity of these effects might depend on the design method, and some design methods might be more robust than others [53]. Our results indicate that EvoColor is more affected by the reality gap than TuttiFrutti across the three missions considered. This is apparent both in the entity of the performance drop we measured and in the fact that the collective behaviors of the control software generated by EvoColor are often dramatically differently in simulation and in the real world, while the ones of the control software generated by TuttiFrutti are essentially unchanged.

By introducing TuttiFrutti, we also investigated the impact of an extended design space in the optimization process of AutoMoDe. The size of the design space in Vanilla and Chocolate is *O*(|*B*| <sup>4</sup> <sup>|</sup>*C*<sup>|</sup> <sup>16</sup>), as estimated by Kuckling et al. [70]. *B* and *C* represent, respectively, the number of modules in low-level behaviors and transition conditions. Using the same method as Kuckling et al., we estimate the design space in TuttiFrutti to be *O*(|4*B*| <sup>4</sup> <sup>|</sup>*C*<sup>|</sup> <sup>16</sup>)—that is, 256 times larger than the one searched by Chocolate. Notwithstanding the larger design space, we do not find evidence that TuttiFrutti is affected by the increased number of parameters to tune. Indeed, TuttiFrutti produced effective control software for all missions considered.
