*5.2.* AGGREGATION

Figure 3 (center) shows the performance of TuttiFrutti and EvoColor in AGGREGATION. In this mission, TuttiFrutti performs significantly better than EvoColor.

Also in this case, from visual inspection, TuttiFrutti produced control software that effectively uses the capabilities that the robots have of displaying and perceiving colors. As we expected, TuttiFrutti designs collective behaviors in which robots reach and remain in the blue zone by moving towards blue walls. This behavior is often complemented with navigation or communication strategies that boost the efficiency of the swarm. For example, some instances of control software include a repulsion behavior that drives robots away from the green walls—robots reach the blue zone faster by avoiding unnecessary exploration in the green zone. In other instances, robots that step in the blue zone, or perceive the blue walls, emit a signal of an arbitrary color—other robots then follow this

signal to reach the blue zone. In this sense, robots communicate and collectively navigate to aggregate faster. Finally, some instances combine the two strategies.

EvoColor designed collective behaviors in which robots use the colors displayed in the arena. Robots explore the arena until they step in one of the black regions—either at the blue or green zone. If robots step in the green zone, they move away from the green walls and reach the blue zone. If robots step in the blue zone, they attempt to stand still. In this sense, robots react and avoid the green walls as a strategy to aggregate in the blue zone.

**Figure 4.** Instance of control software produced by TuttiFrutti for STOP. The probabilistic finite state machine shows the effective modules in black and non-reachable modules in light gray. Circular modules represent the low-level behaviors and rhomboid modules represent transition conditions.

The control software produced by TuttiFrutti and EvoColor showed a significant drop in performance when ported to the physical robots. As observed in STOP, the difference in mean performance between simulations and experiments with physical robots is larger for EvoColor than TuttiFrutti. Robot swarms that use the control software produced by TuttiFrutti display the same collective behaviors observed in simulation. The decrease in performance occurs because few robots that leave the blue zone do not return as fast as observed in the simulations. The control software produced by EvoColor does not port well to the physical robots—that is, robots appear to be unable to reproduce the behaviors observed in the simulation. Robots ramble in the arena and seem to react to the presence of their peers, however, no specific meaningful behavior could be identified by visual inspection.

Figure 5 shows an example of the control software produced by TuttiFrutti for AGGREGATION. Robots start in COLOR-FOLLOWING displaying yellow (*δ* = *Y*) and move towards cyan robots (*γ* = *C*). When they perceive the blue walls (*δ* = *B*), COLOR-DETECTION triggers and the robots transition to a second module COLOR-FOLLOWING in which they move towards the blue walls (*δ* = *B*) while emitting a cyan signal (*γ* = *C*). By cycling in these behaviors, robots can navigate to the blue zone either by moving towards the blue walls or by following the cyan signals that other robots emit. The transition

conditions FIXED-PROBABILITY, GRAY-FLOOR and NEIGHBOR-COUNT trigger the COLOR-FOLLOWING behavior that allows the robot to return to the aggregation area.

**Figure 5.** Instance of control software produced by TuttiFrutti for AGGREGATION. The probabilistic finite state machine shows the effective modules in black and non-reachable modules in light gray. Circular modules represent the low-level behaviors and rhomboid modules represent the transition conditions. Modules labeled as C-FOLLOWING stand for the low-level behavior COLOR-FOLLOWING.
