CAREER: Learning Executable Models of Physical Social Agent Behavior

这项工作的目标是开发能够在观察物理社会代理人时学习可执行行为模型的程序，其中“可执行”意味着模型应该在模拟或移动机器人上运行；而“物理社会代理”指的是感知环境、对环境采取行动的动物、机器人或人，通常是作为一个相互作用的多代理群体的一部分运作。在这项工作中开发的算法将支持以下应用：观察社会性昆虫群体的能力，建立它们的行为模型，然后在模拟中执行和验证模型；观察机动车辆活动的能力（例如，高速公路上的汽车），建立他们的行为模型，然后使用模型在模拟中评估公路设计；以及通过提供团队行为的视频示例来训练移动机器人团队的能力。为了达到这些目的，PI将利用行为生态学、计算机视觉（尤其是活动识别）和基于行为的机器人控制方面的研究，并观察到这些领域所使用的表示惊人地相似。PI将调查自然的、大规模的系统是如何自我组织的，同时，他将研究如何通过编程使人工机器人团队有效地合作；同时进行两种研究的动机是PI希望每一种都能告知另一种。该项目的结果将促进我们对如何组织分布式多智能体系统以获得有效性能的理解。更广泛的影响：在这项工作中开发的技术将深刻影响计算机科学以外的几个学科的研究，包括动物行为学、生态学和动物行为学。为学习和执行社会行为模型而开发的软件将通过互联网提供给研究社区的成员。PI还将开发新的计算机科学课程，这些课程的主题与这项研究交织在一起，包括智能机器人和感知，自主多机器人系统，以及物理多代理系统原理。

The objective of this work is to develop programs that can learn executable models of behavior as they observe physical social agents, where "executable" means that the models should run in simulation or on mobile robots; and "physical social agent" means an animal, robot or person that senses its environment, acts upon its environment, and normally operates as part of an interacting multi-agent group. The algorithms developed in this work will support a applications such as: the ability to observe social insect colonies, build models of their behavior, then execute and verify the models in simulation; the ability to observe motorized vehicle activity (e.g., cars on a highway), build models of their behavior, then use the models to evaluate highway designs in simulation; and the ability to train mobile robot teams by providing video examples of team behavior to be emulated. To these ends, the PI will leverage research in behavioral ecology, computer vision (especially activity recognition), and behavior-based robot control, having observed that the representations utilized in each of these areas are strikingly similar. The PI will investigate how natural, large-scale systems organize themselves and, in tandem, he will study how artificial robot teams can be programmed to cooperate effectively; the motivation for pursuing both lines of research at once is that the PI expects each to inform the other. The results of this project will advance our understanding of how intelligent, distributed multi-agent systems can be organized for effective performance. Broader Impacts: The techniques developed in this work will profoundly influence research in several disciplines outside computer science, including ethology, ecology, and animal behavior. The software developed for learning and executing models of social behavior will be made available over the Internet to members of the research community. The PI will also develop new computer science courses whose topics are intertwined with this research, including one on Intelligent Robotics and Perception, another on Autonomous Multi-Robot Systems, and a third on Principles of Physical MultiAgent Systems.