MENDIBURU, F. J; MENDIBURU, FERNANDO J.; http://lattes.cnpq.br/8884676140983123; MENDIBURU, Fernando Javier.
Abstract:
In this thesis is addressed the problem of spatially organize behaviors in the context
of automatic design of software controllers for swarm robots. Two new specializations
are proposed for the automatic design of software controller of each robot: modular and
monolithic specialization. The purpose of these new specializations is designing the individual
controller of each robot to produce the swarm collective behavior in which specific distances
among robots must be maintained. Automatic design methods for software controllers in
swarms of robots, previously proposed in the bibliography, do not explicitly consider these
distance restrictions in the specialization of the method. The modular method modifies
the robot’s low-level controller to allow linear and angular velocity control of the robot.
Also, a new individual behavior is included in each robot, called formation. This behavior,
designed using the artificial potential theory, allows to include distance constraints in a
behavior-based controller. The formation behavior allows the adjustment of the interactions
among robots to produce the collective behavior of the swarm. The monolithic method adds
the same formation behavior in a neural network architecture without hidden nodes. We
carry out experiments in simulation of the main automatic design methods presented in the
bibliography that deal with collective tasks and distance restrictions. Also, these methods
were compared with the specializations proposed in this thesis: the modular method and
the monolithic method. These methods were compared in simulation and real robots in
three swarm robotic tasks, using 20 e-puck robots. From the observed results, it can be
concluded that the automatic design methods presented in the bibliography, as well as the
monolithic method proposed in this thesis, failed to produce relevant collective behaviors and
cohesion among robots. However, in the modular method, it is concluded that the introduced
constraints produce cohesion and relevant collective behaviors and the performance of the
designed controllers is significantly better than those obtained with methods present in the
bibliography.