Improving the performance of self-organized robotic clustering: modeling and planning sequential changes to the division of labor Jung-Hwan Kim and Dylan A. Shell Abstract— Robotic clustering involves gathering spatially dis- tributed objects into a single pile. It is a canonical task for self- organized multi-robot systems: several authors have proposed and demonstrated algorithms for performing the task. In this paper, we consider a setting in which heterogeneous strategies outperform homogeneous ones and changing the division of labor can improve performance. By modeling the clustering dynamics with a Markov chain model, we are able to predict performance of the task by different divisions of labor. We propose and demonstrate a method that is able to select an open-loop sequence of changes to the division of labor, based on this stochastic model, that increases performance. We validate our proposed method on physical robot experiments. I. I NTRODUCTION Studies of self-organized multi-robot systems (MRS) con- sider multiple agents, each with limited individual capa- bilities, but with the capacity for synergistic interaction in order to perform tasks collectively. Unlike the more common intentional distributed robot teams, the group’s functionality emerges through feedback mediated by the environment and is the product of action rather than representation or calcu- lated reasoning [1]. Self-organized MRS have several poten- tial advantages: simple hardware allows for the production of cheap, specialized, and robust units which exploit economies of scale. However, since the robots in self-organized MRS have limited sensing and manipulation capabilities, it can be difficult to improve the speed of collective performance. It is already known that merely increasing the number of robots will not improve the speed of the system above a certain threshold because of the interference between team members [2], [3]. Principled methods for maximizing system performance (in terms of speed and/or quality) remains challenging for self-organized robot swarms. In our previous work [4], [5], we introduced a novel approach for object clustering, one of the most widely studied task domains for self-organized MRS. The approach we demonstrated consisted of two complementary behaviors: twisting and digging (Fig. 1 illustrates both). Conceptually, twisting behavior is likely to deliver the object into the central region, while digging behavior makes gaps between objects and the boundary. Each robot was assigned with one of these behaviors for the duration of a clustering experiments. With a mix of robots executing the two com- plementary behaviors, the robots detached the objects from the boundary and successfully generated a single central cluster as shown in Fig. 2. Certain mixes of the behaviors Jung-Hwan Kim and Dylan A. Shell are with Department of Computer Science and Engineering, Texas A& M University, College Station, Texas, USA. jnk3355,dshell at cse.tamu.edu Fig. 1: Trajectories of Twisters and Diggers on the boundary region. Basically the trajectories differ by the way they move away from the boundary wall. Fig. 2: The clustering process. (a) initial configuration and (b) final configuration. outperformed other mixes and in different respects. For example, the mix of 2T3D (2 Twisters and 3 Diggers) had reliable performance compared to other cases while mix 1T4D (1 Twister and 4 Diggers) formed a cluster efficiently in the shortest observed time although it failed in one of its trials. This suggests that, given a preference between reliability and efficiency, an appropriate mix (or distribution of labor) could be determined. In this paper, we attempt to address the question of how to maximize the system’s performance by computing a policy for altering the robot division of labor as a function of time. This research considers a sequencing strategy based on the hypothesis that since clustering performance is influenced by the division of labor, it can be improved by sequencing different divisions of labor. We construct a model in order to predict clustering behavior (in terms of likelihood of success and speed) and propose a method that uses the model’s predictions to select a sequential change in labor distribution. Both of these aspects are performed off-line at design time. The model is calibrated with values from experiments in which robots maintain a constant distribution of labor. Then the analysis step is conducted in order to produce a labor 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) November 3-7, 2013. Tokyo, Japan 978-1-4673-6357-0/13/$31.00 ©2013 IEEE 4314