Assessing the Effect of Self-Assembly Ports in Evolutionary Swarm Robotics Kazi Shah Nawaz Ripon * , Eirik Jakobsen * , Christopher Tannum * , Jean-Marc Montanier † * Department of Computer and Information Science, Norwegian University of Science and Technology, Norway † Protolab – Softbank Robotics Europe Email: ksripon@idi.ntnu.no, eirikjak@gmail.com, christannum@gmail.com, montanier.jeanmarc@gmail.com Abstract—Self-assembly in swarm robotics is essential for a group of robots in achieving a common goal that is not possible to achieve by a single robot. Self-assembly also provides several advantages to swarm robotics. Some of these include versatility, scalability, re-configurability, cost-effectiveness, extended reliabil- ity, and capability for emergent phenomena. This work investi- gates the effect of self-assembly in evolutionary swarm robotics. Because of the lack of research literature within this paradigm, there are few comparisons of the different implementations of self-assembly mechanisms. This paper reports the influence of connection port configuration on evolutionary self-assembling swarm robots. The port configuration consists of the number and the relative positioning of the connection ports on each of the robot. Experimental results suggest that configuration of the connection ports can significantly impact the emergence of self- assembly in evolutionary swarm robotics. Keywords—Evolutionary robotics, self-assembly, machine learning, connection port configuration, docking mechanism. I. I NTRODUCTION By drawing inspiration from social insects [1], [2] and other self-organizing systems [3], [4], swarm robotics approaches the coordination of a number of autonomous robots, which need to interact and to cooperate to achieve a common goal [5]. The core idea is to capitalize on simple interactions among robots in order to solve complex problems by means of emergent behaviour. Among these behaviours, self-assembly is the autonomous organization of components into patterns or structures [6]. It can provide multiple advantages in robotics such as robustness through redundancy and cost reduction through design of large number of robots [7]. Another crucial benefit is the versatility achieved when robots self-assemble into new structures based on the specific task to solve [8]. For example, the movement of robots is improved when they are able to overcome larger obstacles in an environment. In order to foster most of this advantage, it is crucial to learn when and how to self-assemble to face the environmental conditions at hand. This is the objective of functional self-assembly [9]. In order to achieve functional self-assembly, robots have to learn autonomously the most suited behaviours. To date, au- tonomous learning of self-assembly has been achieved through reinforcement learning [10] and evolutionary algorithms [11]. This work will consider only the evolutionary algorithms. Evo- lutionary Robotics (ER) studies how to automate the design of control systems for autonomous robots, using algorithms based on Darwinian evolution [9]. This approach has been already successfully used in the design of swarm behaviours [12], [13]. Previous studies on self-assembly have shown the difficulty to evolve such behaviour [14]. A key goal of this work is to study the mechanisms that can promote the learning of functional self-assembly through evolution. Previous studies on the evolution of self-organization have focused on using differing self-assembly mechanisms, such as the architecture used by the self-assembly system, the actions that are performed by the robots to achieve an organized structure, or the docking mechanism hardware [9], [14], [15]. However, existing works do not provide any direct comparison among the mechanisms in order to justify which is the most suitable to promote the evolution of self-assembly. This hinders further research on the topic. This work studies one specific aspect of the self-assembly mechanism: the configura- tion (number and relative positioning) of the connection ports. Mainly, we studied the influence of ports configuration on the emergence of self-assembly when the robots are given basic learning capabilities. For the experiment, we designed a simple predator/prey scenario, where the evolved robots (prey) can self-assemble to gain certain advantages over its predators. The experiments are simulated with a heavily modified version of roborobo platform [16]. The remainder of this paper is organized as follows. Sec- tion II presents the state of the art. Section III contains the system description and implementation, including modifica- tions which were made to the existing roborobo framework. Section IV describes the results of the experiment, as well as an analysis and reasoning for their given state. Section V concludes the work including scopes for future work. II. RELATED WORKS Multiple works have discussed and compare the various self-reconfigurable robots [17]–[20]. Unfortunately, most of them introduce new platforms without comparing their per- formances to previous works [17]–[19]. To date, there exists no such work justifying a self-assembly system as better than others considering any specific property. Latest articles for the design of self-reconfigurable robots propose versatile robots able to emulate previously cited robots [21], [22]. Nevertheless, their performance with regards to the simulated counter-parts is not studied. [20] compares