Generating Collective Behavior of Multi-legged Robotic Swarm with Deep Neuroevolution

Abstract

In this paper, the controller of the multi-legged robotic swarm is designed by deep neuroevolution, which is a technique to train a deep neural network by using artificial evolution. The computer simulations are conducted with a 3D physics engine called Bullet. An aggregation task is examined with varying the sensor range to discuss the behavior. The results show that deep neuroevolution was able to generate collective behavior of the multi-legged robotic swarm. Moreover, the robotic swarm showed a potential behavior that might be useful to achieve more complex tasks.