NRI: Small: Online Training of Hierarchical Multirobot Teams

这个项目涉及一种新的机器人从演示中学习的方法，允许实验者实时教给多个机器人非平凡的、不同种类的集体行为。训练多个机器人是具有挑战性的，因为尽管实验者可能知道他希望实现的紧急集体行为，但他不太可能知道机器人的个人行为及其相互作用将实现这一目标。这个项目将机器人团队或群体分解成由领导者和下属组成的社会等级。然后，小群体在层次结构的底部训练相对简单的集体行为，然后用更复杂和抽象的行为逐步训练领导者群体，直到所有训练的顶层根行为都被学习。该项目检查机器人组内的同质行为、机器人之间的异质行为，以及在层次结构中的任何级别上两者的混合。该方法建立在单机器人行为训练的前人工作的基础上，该方法使用由有限状态自动机和学习的转移函数组成的行为的分解层次。除了机器人，该项目还允许为动画以及社会科学和生物学中新兴的基于代理的模型中的虚拟代理开发群体行为。这项工作与教学和研究机器人实验室、社会科学中基于代理的建模人员以及K-12贫困学校的机器人教育工作相结合。

This project involves a new approach to robot learning from demonstration which allows experimenters to teach multiple robots nontrivial, heterogenous collective behaviors in real-time. Training multiple robots is challenging because while the experimenter may know what emergent collective behavior he wishes to achieve, it is unlikely that he knows what individual robot behaviors, and their interactions, will achieve it. This project decomposes robot teams or swarms into a social hierarchy of leaders and subordinates. Small groups are then trained at the bottom of the hierarchy relatively simple collective behaviors, then groups of leaders are trained progressively with more complex and abstract behaviors, until all trained top-level root behaviors have been learned. The project examines homogeneous behaviors within robot groups, heterogeneous behaviors among the robots, and mixtures of the two at any level in the hierarchy. The method builds on prior work in single-robot behavior training, which used decomposed hierarchies of behaviors consisting of finite-state automata and learned transition functions. Beyond robotics, the project also allows the development of group behaviors for virtual agents found in animation and in emerging agent-based models in the social sciences and biology. The work integrates with a teaching and research robotics laboratory, with agent-based modelers in the social sciences, and with robotics education efforts in K-12 underprivileged schools.