Multirobot Behavior Synchronization through Direct Neural Network Communication In: Proceedings of the 5th International Conference on Intelligent Robotics and Applications (ICIRA-2012). New York, NY: Springer-Verlang, 2012. David B. D’Ambrosio, Skyler Goodell, Joel Lehman, Sebastian Risi, and Kenneth O. Stanley Department of Electrical Engineering and Computer Science University of Central Florida Orlando, Florida 32816-2362, USA {ddambro,goodsky,jlehman,srisi,kstanley}@eecs.ucf.edu http://eplex.cs.ucf.edu/ Abstract. Many important real-world problems, such as patrol or search and rescue, could benefit from the ability to train teams of robots to coordinate. One major challenge to achieving such coordination is de- termining the best way for robots on such teams to communicate with each other. Typical approaches employ hand-designed communication schemes that often require significant effort to engineer. In contrast, this paper presents a new communication scheme called the hive brain, in which the neural network controller of each robot is directly connected to internal nodes of other robots and the weights of these connections are evolved. In this way, the robots can evolve their own internal “language” to speak directly brain-to-brain. This approach is tested in a multirobot patrol synchronization domain where it produces robot controllers that synchronize through communication alone in both simulation and real robots, and that are robust to perturbation and changes in team size. Keywords: Evolutionary Algorithms, HyperNEAT, Multirobot Teams, Coordination, Communication, Artificial Neural Networks 1 Introduction As robot technology has matured and large teams of robots have become more commonplace, a research question of growing importance is how to best coor- dinate such robotic teams. While one approach is to coordinate robot teams centrally, scaling such an approach to many robots and mitigating the inherent challenges of limited bandwidth and unreliable communication in the real world may prove problematic [1]. Thus this paper instead focuses on treating robots as autonomous communicating agents, which notably has proven a robust and scalable strategy in nature [2,3]. For example, insect colonies and human society itself operate by this principle. Importantly, communication between agents enlarges the scope of their pos- sible behaviors by enabling coordination and sharing of knowledge. In this way,