Towards an Evolutionary Design of Modular Robots for Industry - Foundations on Natural and Artificial Computation - page 58

Fig. 4. Screenshots of the final morphology in action
In addition, a clear general trend can be observed in the evolution, where
the number of modules per robot increases quickly during the first 30 iterations,
when the algorithm is exploring the search space. In the last part of the evolution,
the number of modules is stabilized to an average number of around 9 modules
per robot. In practical terms, this evolution leads to a final morphology that is
shown in Fig. 4 and that is able to cover a distance of around 4 m (performing
a random search in the same conditions we obtain a maximum distance of 2 m).
It can be observed that the robotic unit has four branches of modules that are
fixed to the base. Two of them produce the movement of the robot and the other
two stabilize it. Furthermore, it presents a hinge attached between the mobile
part of the slider, fixed to the base, and a slider that acts like a leg. It constitutes
the first branch which produces movement. The other branch is composed by
a hinge fixed to the base and to the mobile part of one slider. Another slider
is joined in parallel with this slider to elevate the front of the robot. In order
to stabilize the robot in its motion, a slider and a rotational (axis) module are
attached in the left side. The end of this slider touches the ground in a certain
position of the motion and prevents the robot from falling. Finally, there is a
branch made up of two hinges which are fixed at a port in the right side of the
base, and which prevent the displacement of the robot in the Y axis direction.
5 Conclusions
In this paper we have introduced a constructive evolutionary strategy that per-
mits obtaining both the morphology and the control parameters for modular
multicomponent robotic systems. This algorithm has been designed to take into
account the intrinsic deceptiveness and lack of information in many areas of
the search space and provides an easy way to balance the different evolutionary
scales at which the evolution of morphology and control must take place in order
to be able to cooperatively coevolve these two elements. The algorithm has been
linked to the Gazebo simulator for the evaluation of the fitness of the individuals
it produces, thus obtaining a complete evolutionary design and testing system
for modular robots. The operation of this approach has been shown through its
application to the generation of a robot and control strategy to a typical bench-
mark problem using a modular architecture developed in our group with very