CAREER: Toward Lifelong Safety of Autonomous Systems in Uncertain and Interactive Environments

该奖项全部或部分由《2021年美国救援计划法案》（公法117-2）资助。这项教师早期职业发展计划（Career）拨款将资助研究，使自主系统能够在与人类密切互动的情况下安全运行，例如，在下一代制造基础设施中，从而促进科学进步，促进国家繁荣。直到最近，为了防止伤害和死亡，人类与机器人在物理上是分开的。现代机器人技术关注的是人类和协作机器人系统一起完成相同的任务。在这种交互环境中的安全隐患是人为错误的发生。至关重要的是，将安全意识反应编程到协作机器人系统中，以确保即使在任务或环境发生变化时也能保证安全行为。该项目将开发一种新的算法框架，用于自动机器人系统的安全保证，旨在在安全可以管理时实现最佳性能，在无法预测和补偿不可避免的故障时，并从过去的错误中学习。该框架将提高自主系统的可信度，同时最大限度地减少人力在部署和维护方面的努力，这是在不确定和交互式环境中赋予智能机器人完全自主的关键步骤，包括工业机器人和自动驾驶等应用领域。通过研究和教育的紧密结合，该项目将有助于机器人和自主领域新的跨学科培训，向公众传播机器人安全研究，并通过远程操作的机器人平台提供互动学习的机会。与先进机器人制造研究所、钢铁女孩机器人项目和乔特罗斯玛丽霍尔大学预科学校的合作伙伴关系将为研究生提供在小型制造商实习的机会，并扩大目前代表性不足的群体的个人研究参与。本研究旨在为跨任务安全监护人理论做出基础贡献，该理论在不手动调优的情况下增强现有硬件平台，监控并优化其名义上的面向任务的控制动作，以满足代表安全要求的约束，并在时变不确定性下实现这些目标。它通过研究数据高效的模型学习算法来实现这一目标，该算法可以准确地跟踪交互环境的动态，并通过设计自适应控制器来根据新学习的动态模型安全地调整控制策略。提出了一种基于责任的进化对抗学习方法，使自适应安全控制算法在给定可用资源限制的情况下达到最优性能。安全卫士和智能优化器方法的评估是在模拟和实验中实现的，使用自动驾驶车辆在不同的交通条件下与人类操作的车辆进行交互，以及在涉及机械臂操纵器和其他人类或机器人代理的空间共享应用中。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2).This Faculty Early Career Development Program (CAREER) grant will fund research that enables autonomous systems to operate safely in close interaction with humans, as required, for example, in next generation manufacturing infrastructure, thereby promoting the progress of science, and advancing the national prosperity. Until recently, humans were physically separated from robots to prevent injuries and fatalities. Modern robotics focuses on humans and collaborative robotic systems working together on the same tasks. A safety hazard in such interactive environments is the occurrence of human errors. It is critical that safety conscious responses be programmed into collaborative robotic systems to guarantee safe behavior even when tasks or environments change. This project will develop a new algorithmic framework for safety assurance of autonomous robotic systems that aims for optimal performance when safety can be managed, anticipates and compensates for inevitable failures when it cannot, and learns from past mistakes. This framework will increase trustworthiness of autonomous systems while minimizing human efforts in deployment and maintenance, critical steps toward granting full autonomy to intelligent robots in uncertain and interactive environments, including such application domains as industrial robotics and autonomous driving. Through close integration of research and education, this project will contribute to new interdisciplinary training in robotics and autonomy, accessible dissemination of research in robot safety to the public, and opportunities for interactive learning through a remotely operated robotic platform. Partnerships with the Advanced Robotics for Manufacturing Institute, the Girls of Steel Robotics program, and the Choate Rosemary Hall college-preparatory school will be leveraged to provide opportunities for graduate student internships with small manufacturers and broaden participation in research of individuals from currently underrepresented groups.This research aims to make fundamental contributions to a theory of cross-task safe guardians that augment existing hardware platforms without manual tuning, monitor and optimally modify their nominal task-oriented control actions to satisfy constraints representing safety requirements, and accomplish these objectives under time-varying uncertainty. It achieves this aim by investigating data-efficient model learning algorithms that accurately track the dynamics of an interactive environment, as well as by designing adaptive controllers that safely adjust the control strategy according to newly learned dynamic models. A responsibility-based evolutionary adversarial learning approach is developed to enable the adaptive safe control algorithm to achieve optimal performance given limits on available resources. Evaluation of the safe guardian and intelligent optimizer approaches is achieved in simulation and experimentally using autonomous vehicles interacting with human-operated vehicles in different traffic conditions, as well as in space-sharing applications involving robot arm manipulators and other human or robotic agents.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.