Learning Coordination for Multi-Autonomous Multi-Human (MAMH) Agent Systems with Guaranteed Safety

许多真实世界系统的操作涉及人类和自主代理的共存。这些代理之间的协调不充分可能导致严重的性能降级或安全风险。该项目旨在开发一种新的多自治多人协作框架，以增强算法的可扩展性和安全性保证。与传统的优化和机器学习方法相比，该框架解决了两个主要挑战：(I)系统的非合作性，这是由于人类和机器人之间的信息不对称、人类偏好的异质性以及人类在与机器人合作时的决策自私所引起的；(Ii)协调安全，这在有人类代理的情况下是至关重要的，但很难使用传统的黑盒学习模型来衡量。为了应对这些挑战，本研究的智力价值在于创新地融合了博弈论、机器学习、人类建模和网络控制理论，形成了一种多自治多人协调的框架，提高了模型的透明度和可学习性。该框架的核心是一种新颖的人-响应对齐机制，允许系统中的自主代理不仅被动地适应人类的行为，而且还能巧妙地指导他们，从而提高整个系统的效率和安全性。为了促进这一点，将以分布式训练分布式执行的方式开发计算可扩展和高效的算法，纯粹基于代理用于通信和计算的本地资源。这项工作的更广泛影响延伸到各种工程实践，包括交通协调、人-机器人协作以及涉及人类用户的电力/物联网系统。该项目特别强调劳动力发展和教育。精心设计的“RoboArt”活动将吸引K-12学生参与，培养创造力、解决问题的技能和STEM曝光率。该项目还将为高中生和大学生提供多学科学习和研究机会，确保全面进入不断发展的机器人和机器学习领域。此外，该项目将为研究界贡献有价值的数据集，强调可访问性和可重用性，以促进该领域正在进行的创新。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 operation of many real-world systems involves the co-existence of human and autonomous agents. Inadequate coordination among these agents can lead to significant performance degradation or safety risks. This project aims to develop a novel framework for Multi-Autonomous Multi-Human coordination, which enhances algorithmic scalability and safety guarantees. Compared with traditional optimization and machine learning approaches, the proposed framework addresses two major challenges: (i) the non-cooperative nature of the system, which arises from information asymmetry between humans and robots, heterogeneity in human preferences, and human selfishness in decision-making when working with robots; and (ii) coordination safety, which is of critical importance in the presence of human agents but is difficult to measure using traditional black-box learning models. Additionally, human behaviors are subject to uncertainties, which may easily deviate the actual coordination from intended ones.To address these challenges, the intellectual merits of this research lie in its innovative integration of game theory, machine learning, human modeling, and network control theory, resulting in a framework for Multi-Autonomous Multi-Human coordination that enhances both model transparency and learnability. Core to the framework is a novel human-response alignment mechanism, allowing autonomous agents in the system to not only passively adapt to human behaviors but also subtly guide them, enhancing the efficiency and safety of the entire system. To facilitate this, computationally scalable and efficient algorithms will be developed in the manner of distributed-training-distributed execution, purely based on agents’ local resources for communication and computation. The broader impacts of this work extend to various engineering practices, including traffic coordination, human-robot teaming, and power/IoT systems involving human users. The project has a special emphasis on workforce development and education. A carefully designed "RoboArt" event will engage K-12 students, fostering creativity, problem-solving skills, and STEM exposure. The project will also offer multidisciplinary learning and research opportunities for high school and university students, ensuring inclusive access to the evolving field of robotics and machine learning. Furthermore, the project will contribute valuable datasets to the research community, emphasizing accessibility and re-usability to facilitate ongoing innovation in the field.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.