Data-driven material-based encoding for swarm decentralized controllers

Efficient and versatile collaborative swarm algorithms are essential for various applications, including coordinated motion, efficient delivery systems, and environmental monitoring and resource searching. However, many current algorithms are developed in an ad hoc manner, lacking a foundation in physical principles that often drive emergent dynamics in nature. Drawing inspiration from biological bottom-up approaches, decentralized swarms coordinate multiple units to behave as a discrete, deformable macroscopic body. In this study, we introduce a strategy to encode material-based principles into swarms, enabling them to mimic a unified continuum response to external stimuli. Leveraging extensive synthetic data from a versatile discrete bio-lattice simulation framework, we employ neural networks to train a robotic swarm to emulate structured materials and react to fields of attractants or repellents. To assess performance, we thoroughly test our approach using both simulation campaigns and a real-world swarm of ground drones (i.e., kilobots). We propose benchmark tests and demonstrate the system’s capability to integrate material-based behaviors into decentralized swarm control, paving the way for the transfer of material-based behaviors into collaborative algorithms.