Roles such as leading and following can emerge naturally in human groups. However, in human–robot teams, such roles are often predefined due to the difficulty of scalably learning and adapting to them. In this work, we enable a robot to efficiently learn how group dynamics emerge and evolve in human teams and we leverage this understanding to plan for influencing actions for autonomous robots that guide the team toward achieving a common goal. We first develop an effective and concise representation of group dynamics, such as leading and following, by enforcing a graph structure while learning the weights of the edges corresponding to one-to-one relationships between the agents. We then develop an optimization-based robot policy that leverages this graph representation to attain an objective by influencing a human team. We apply our framework to two types of group dynamics, leading-following and predator–prey, and show that our structured representation is scalable with different human team sizes and also generalizable across different tasks. We also show that robots that utilize this representation are able to successfully influence a group to achieve various goals compared to robots that do not have access to these graph representations (Parts of this work has been published at Robotics: Science and Systems (RSS) (Kwon et al. in Proceedings of robotics: science and systems (RSS), 2019. https://doi.org/10.15607/rss.2019.xv.075).

领导和跟随等角色可以在人类群体中自然出现。然而，在人机团队中，由于难以大规模学习和适应这些角色，这些角色通常是预先定义好的。在这项工作中，我们使机器人能够有效地学习群体动态如何在人类团队中出现和演变，我们利用这种理解来规划影响自主机器人的行动，引导团队实现共同目标。我们首先通过执行图结构同时学习与代理之间的一对一关系对应的边的权重来开发团队动态的有效且简洁的表示，例如领先和跟随。然后，我们开发了一种基于优化的机器人策略，该策略利用此图表示通过影响人类团队来实现目标。我们将我们的框架应用于两种类型的群体动态，领先跟随和捕食者-猎物，并表明我们的结构化表示可以随着不同的人类团队规模进行扩展，并且也可以在不同的任务中推广。我们还表明，与无法访问这些图形表示的机器人相比，利用这种表示的机器人能够成功地影响一个群体以实现各种目标（这项工作的部分内容已发表在 Robotics: Science and Systems (RSS) ( Kwon 等人在机器人学论文集：科学与系统 (RSS)，2019 年。https://doi.org/10.15607/rss.2019.xv.075）。