Machine teaching in Swarm Metaverse under different levels of autonomy.

Shepherding algorithms enable scalable swarm control via the utilization of one or a few control agents. Despite their demonstrated effectiveness in controlling swarms of point-particle agents, shepherding algorithms have been barely evaluated in controlling realistic swarms of uncrewed vehicles (UxVs). Furthermore, existing shepherding algorithms face significant challenges in dealing with complex environments such as those featuring obstacles. We address these research gaps by studying the use of human demonstrations for teaching herding behaviours to machine learning controllers. In particular, we focus on how the level of autonomy used for collecting human demonstrations affects the effectiveness of the resulting swarm controller performance. Our experimental investigation shows that demonstrations collected under a high level of autonomy result in a significantly higher success rate than those collected under a low level of autonomy. Our findings highlight that providing high-level commands for the human demonstrator is more effective even when the demonstrations is used for training a low-level controller.This article is part of the theme issue 'The road forward with swarm systems'.