SENSING AND ESTIMATION ON A MODULAR TESTBED FOR SWARM ROBOTICS Gregory K. Fricke, Devendra P. Garg ∗ Robotics and Manufacturing Automation Laboratory Mechanical Engineering and Materials Science Department Duke University {gkf4, dpgarg}@duke.edu Dejan Milutinovi ´ c Applied Mathematics and Statistics Department University of California at Santa Cruz dejan@soe.ucsc.edu ABSTRACT Collective robotics offers the promise of enhanced perfor- mance and robustness relative to that of individual robots, with decreased cost or time-to-completion for certain tasks. Having many degrees of freedom, the multi-robot control and estima- tion problems are challenging, specifically when the solutions require a great amount of communication among the robots. While numerical simulation is a critical tool in swarm robotics research, verification of obtained results under a physical real- ization of the swarm is far from routine. Therefore, we have de- veloped and used a sensor-integrated testbed for the validation of cooperative-robotics algorithms, observation of swarm behavior, and measurement of system performance. INTRODUCTION Conducting experiments in swarm robotics is challenging, time-consuming, and expensive, leading many investigators in the field to extensively rely on computer simulations. While simulation is indispensable for algorithm development and per- formance estimation, the physical realization of a swarm system is important for full understanding and validation of swarm be- havior. Also, physical experiments shed light on the nature of uncertainty sources in the sensors, actuators, and communication pathways specific to mobile robotic systems while illustrating the logistical difficulties involved in implementing swarm behavior. Robot pose estimation is critical for feedback control of robot trajectory. Methods for pose estimation and robot local- ization have been extensively studied, utilizing methods such as dead-reckoning [1–6], on-board and off-board computer vi- sion [7,8], environmental sensing [9,10], collective or distributed localization [11, 12], and simultaneous localization and map- ∗ Address all correspondence to this author. ping (SLAM) [13] using laser-scanning systems [14, 15], on- board cameras [16], RFID tags and readers [17, 18], or ultrasonic rangefinders [10]. Pose estimation and tracking of individual robots are critical to the metrology required for complete evaluation of experimen- tal results. In the case of multi-robot systems, the scalability of the metrology system is a critical concern. With multiple robots, reliance on advanced vision-based techniques quickly surpasses all but the most advanced computational systems, imposing the restriction that performance data must be evaluated off-line with- out real-time constraints. Figure 1. Khepera-II arena of the RAMA Lab.