A swarm robotics approach to cooperative package-pushing problems with evolving recurrent neural networks

Swarm robotic systems are a kind of multi-robot systems consisting of many homogeneous autonomous robots without any type of global controllers. Since a task given to a whole system cannot be solved by a single robot, cooperative behavior should be developed in a robotic swarm by a certain emergent mechanism. In this paper, an evolutionary robotics approach, i.e., the method that the robot controllers are designed by evolving artificial neural networks, is adopted. However, it is well known that the evolvability of an artificial neural network is strongly dependent on its topological structure. Therefore, this paper empirically finds a better artificial neural network structure by conducting computer simulations. Four types of recurrent artificial neural networks are compared with a benchmark called the cooperative package pushing problem. We find that the best performance is obtained with a recurrent neural network of which hidden layer has the small-world properties.