CAREER: Automated Design of Decentralized Robust and Explainable Swarm Systems (ADDRESS)

该学院早期职业发展(Career)项目的目标是识别和测试新的科学原理，以设计具有可靠集体行为的群体机器人系统。在对自然界中非凡的合作行为的观察的推动下，简单机器人的大型团队承诺前所未有的任务效率和弹性优势，而不是复杂的独立系统。这些有益的能力对于未来的灾难应对、环境监测、军事行动和太空探索，以及美国在这些领域继续保持科学领导地位都是至关重要的。然而，大多数现有的设计大型集体操作的移动机器人行为的方法都受到两个关键限制：缺乏群体级别的可预测性能保证，以及无法适应不同的不确定环境和操作规模。该奖项支持理论研究，通过利用机器学习工具与严格的工程设计方法相结合，使系统分析和定制单个机器人的知识表示和物理设计如何影响其集体行为，从而同时解决这些限制。这些理论贡献将通过设计和测试新的小型空中和地面机器人来付诸实践，这些新的小型空中和地面机器人适用于集群应用，在时间紧迫的紧急反应和污染清理领域具有广泛的社会影响。让当地应急利益攸关方参与的外联工作将使人们能够了解将这种群体机器人技术转化为实践的潜在障碍。这个多学科项目还将在设计和机器人学的交叉领域为工程学学生提供新颖的体验式学习环境和多样性倡议，并通过丰富研究生课程和在旗舰机器人学会议上组织新的研讨会来促进这些领域的高级技能发展。本研究的首要目标是调查核心假设，即可靠的群体系统可以通过从可证明的最优专家解决方案中模仿个体代理行为和并发剪裁代理形态来进行计算设计。这里的“可靠性”包括随之而来的集体行为的概括性、可扩展性和数学可解释性，将在分散的群体机器人系统的背景下进行分析，这些系统包括手掌大小的轮式机器人和多旋翼无人机，其任务是提供目标搜索和集体运输行动。为了实现这一目标，本研究设想了以下三个关键的基本贡献：1)开发可学习的个体智能体知识的尺度不可知表示，通过新颖的贝叶斯搜索和图论模型来调节其任务规划过程；2)识别混合模仿学习方法，以适应不同环境中的个体智能体行为，同时最小化随后的集体行为与可证明是最优的离线解的偏差；3)发展基于新颖的约束策略梯度和协同进化方法的计算方法，以同时设计智能体的形态和智能体行为的学习，使得随后的形态复杂性最优地促进必要的行为适应。这些贡献将提供对“可靠性”和“形式和行为的相互作用”的更多了解，预计这将影响除群体机器人之外的广泛的多智能体和分散系统，例如各种网络物理系统。这项综合教育计划包括：1)基于群体电脑游戏和以保护为重点的无人机飞行实验，为工科学生(包括代表性不足的少数族裔)创建新的体验式学习计划，以及2)新的研究生课程和会议工作室。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The objective of this Faculty Early Career Development (CAREER) project is to identify and test new scientific principles to design swarm robotic systems with dependable collective behavior. Motivated by observations of phenomenal cooperative behavior in nature, large teams of simple robots promise unprecedented task efficiency and resilience benefits over sophisticated standalone systems. These beneficial capabilities are fundamental to the future of disaster response, environment monitoring, military operations and space exploration, and to the continuing scientific leadership of the U.S. in these domains. However, most existing approaches to designing the behavior of mobile robots that operate as large collectives suffer from two key limitations: the lack of predictable performance guarantees at the swarm level and the inability to adapt to different uncertain environments and scales of operation. This award supports theoretical research to concurrently tackle these limitations by leveraging machine learning tools combined with rigorous engineering design approaches that enable systematic analysis and tailoring of how knowledge representation and the physical design of individual robots influence their collective behavior. These theoretical contributions will be reduced to practice by designing and testing new small aerial and ground robots for swarm applications with broad societal impact in the areas of time-critical emergency response and pollution clean-up. Outreach efforts engaging local emergency-response stakeholders will allow understanding of potential barriers to transitioning such swarm-robotic technologies to practice. This multidisciplinary project will also enable novel experiential learning environments and diversity initiatives for engineering students at the intersecting fields of design and robotics, and facilitate advanced skill development in these fields by enriching the graduate curricula and organizing a new workshop at a flagship Robotics conference.The overarching goal of this research is to investigate the central hypothesis that dependable swarm systems can be computationally designed via imitation learning of individual agent behavior from provably-optimal expert solutions and concurrent tailoring of agent morphology. Here "dependability" encompasses the generalizability, scalability and mathematical explainability of the ensuing collective behavior, which will be analyzed in the context of decentralized swarm robotic systems, comprising palm-sized wheeled robots and multirotor drones, that are tasked to provide target search and collective transport operations. To accomplish this goal, the following three key fundamental contributions are envisioned in this research: 1) develop learnable scale-agnostic representations of the individual agent's knowledge that regulates its task-planning processes embodied by novel Bayesian search and graph-theoretic models; 2) identify hybrid imitation learning approaches for adapting the individual agent behavior over varying environments, while minimizing the deviation of the ensuing collective behavior from provably-optimal offline solutions; 3) develop computational methods based on novel constrained policy gradient and co-evolution approaches to concurrently design agent morphology along with the learning of agent behavior, such that the ensuing morphological complexity optimally facilitates the necessary behavioral adaptations. These contributions will provide an increased understanding of "dependability" and the "interplay of form and behavior" that is expected to impact a broad range of multi-agent and decentralized systems beyond swarm robotics, such as various cyber physical systems. The integrated education plan involves the creation of 1) new experiential learning programs for engineering students, including underrepresented minorities, based on swarm computer games and conservation-focused drone flight experiments, and 2) a new graduate course and conference workshops.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.