CAREER: Towards a Theory of Robust Learning & Control for Safety-Critical Autonomous Systems

未来的自主系统，如自动驾驶汽车和敏捷机器人，将负责在不断发展和不确定的条件下执行复杂而复杂的任务，使用从复杂、高维、丰富的感知模式（如摄像头）收集的信息。由于这种普遍存在的不确定性和复杂性，反馈控制回路将继续在自治系统中普遍存在。经典的控制理论技术需要复杂而详细的动力系统模型，并假设信息是由简单的单输出传感设备（如加速度计）提供的，这些假设在未来自主系统的设想中显然是失败的。相反，虽然机器学习和强化学习技术可以适应不确定的动态条件和丰富的感知模式，但它们往往只关注性能而忽略了安全性/鲁棒性问题，或者即使它们解决了安全性，也只适用于有限类别的系统。尽管最近在学习和控制的原则集成方面取得了进展，但在可以被认证为安全、健壮和高性能的系统类别与现实世界的自治系统之间仍然存在很大的差距。该项目旨在制定一项研究计划，为鲁棒学习和鲁棒控制的新理论奠定基础，同时解决现实世界中各种安全关键型自主系统的安全和性能挑战。该项目的研究成果将整合到协同教育计划中，该计划包括在宾夕法尼亚大学开发研究生和本科课程，旨在丰富工程师自主和控制系统教学课程。宾夕法尼亚大学还将利用公开的教育平台和外展项目，为STEM专业的学生建立进入大学的渠道，增加代表性不足的少数族裔的代表性，推进围绕自主和控制系统的公共沟通，并传播研究成果。该提案认为，需要一种新的跨学科视角来研究鲁棒控制和鲁棒机器学习，以释放基于学习的控制在安全关键复杂、动态和不确定场景中的真正潜力。推力将开发新的鲁棒的基于学习的控制策略，明确表征和解释学习和控制管道中的不确定性的影响。第一个重点是通过鲁棒学习和鲁棒控制技术的协同集成，利用当代高容量模型（如深度神经网络）学习控制未知的动态系统，旨在减轻分布转移对闭环性能的有害影响。第二个重点是通过开发工具，将控制理论的鲁棒性和稳定性保证扩展到具有复杂、高维传感模式（如相机）的系统，这些工具允许将这些复杂的感知传感器抽象地视为适用于传统鲁棒控制方法的“噪声虚拟传感器”。最后，第三个推力启动了对学习控制器的鲁棒性和样本复杂性的基本限制的研究，使用基于感知的控制作为案例研究。因此，通过控制理论、机器强化学习、统计学习理论和鲁棒优化等工具的跨学科组合，该项目将开发新的广泛适用的联合鲁棒学习和鲁棒控制工具，这些工具将具有强大的性能、鲁棒性、安全性和样本效率保证。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。

Future autonomous systems, such as self-driving cars and agile robots, will be tasked with performing sophisticated and complex tasks under continuously evolving and uncertain conditions, using information gleaned from complex, high-dimensional, rich perceptual sensing modalities (e.g., cameras). Due to this ubiquitous uncertainty and complexity, feedback control loops are and will continue to be pervasive in autonomous systems. Classical control theory techniques require intricate and detailed models of dynamical systems, and assume that information is provided by simple, single output sensing devices (e.g., accelerometers), assumptions that clearly fail in the scenarios envisioned for future autonomous systems. Conversely, while techniques from machine and reinforcement learning can accommodate both uncertain dynamic conditions and rich perceptual sensing modalities, they tend to either focus solely on performance and ignore safety/robustness concerns, or if they do address safety, are only applicable to a limited class of systems. Despite recent progress in principled integration of learning and control, there still exists a wide gap between the class of systems that can be certified as safe, robust, and high-performing, and real-world autonomous systems. This project aims to develop a research plan that builds the foundations of a novel theory of robust learning and robust control that simultaneously addresses the challenges of safety and performance across a wide range of safety-critical real-world autonomous systems. The research outcomes of this project will be integrated into a synergistic education plan which includes developing both graduate and undergraduate courses at the University of Pennsylvania aimed at enriching the curriculum for teaching autonomy and control systems to engineers. Publicly available education platforms and outreach programs within the University of Pennsylvania will also be leveraged to build a pipeline for STEM majors entering college, to increase representation among underrepresented minorities, to advance public communication around autonomy and control systems, and to disseminate research results.This proposal argues that a novel cross-disciplinary perspective on robust control and robust machine learning is required to unlock the true potential of learning-based control in safety-critical complex, dynamic, and uncertain scenarios. Thrusts will develop novel robust learning-based control strategies that explicitly characterize and account for the effects of uncertainty in the learning and control pipeline. The first thrust focuses on learning to control an unknown dynamical system using contemporary high-capacity models, such as deep neural networks, through a synergistic integration of robust learning and robust control techniques aimed at mitigating the deleterious effects of distribution shift on closed-loop performance. The second thrust focuses on extending the robustness and stability guarantees of control theory to systems with complex, high-dimensional sensing modalities such as cameras, by developing tools that allow for such complex perceptual sensors to be abstractly viewed as “noisy-virtual sensors” that are amenable to traditional robust control methods. Finally, the third thrust initiates a study of the fundamental limits of the robustness and sample-complexity of learning-enabled controllers, using perception-based control as a case study. Thus, through an interdisciplinary mix of tools from control theory, machine & reinforcement learning, statistical learning theory, and robust optimization, this project will develop novel broadly applicable joint robust learning and robust control tools that come with strong guarantees of performance, robustness, safety, and sample-efficiency.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.