Collaborative Research: SLES: Guaranteed Tubes for Safe Learning across Autonomy Architectures

自动驾驶系统（如自动驾驶汽车和送货无人机）会因任务和环境的复杂性而面临安全挑战。尽管这些系统已经证明了自己的学习和适应能力，但确保这些系统的安全并非易事。安全可能受到各种因素的威胁，包括但不限于意外变化、恶劣的天气条件或未知的障碍物。这些系统的车载学习解决方案主要用于电脑游戏等非关键情况，在这些情况下，安全问题最小。该提案旨在解决各种应用场景（例如，城市空中交通中的自动驾驶汽车和飞行车辆）中对端到端安全的迫切需求。我们提出了一种新颖的解决方案，称为“数据支持单纯形”或“DeSimplex”。DeSimplex建立在坚实的数学原理和系统的方法上，用于收集数据并使用数据来改进系统的性能。它提供了一个可以被证明是安全的框架，使具有学习功能的系统即使在面对极端事件或环境危害时也能进行调整并表现良好。这项工作对于更广泛的应用至关重要，这些应用涉及在不可预测的、苛刻的物理环境中安全高效地运行自主系统，包括自动驾驶汽车和3D城市空中交通的飞行器。提出的工作为推进对端到端学习系统安全的理解奠定了基础，这是网络物理系统、机器人和机器学习中的一个基本问题。我们的目标是追求以下两个相互关联的研究重点：(i)通过可靠的不确定性量化方法提高高性能自治，以确保数据驱动的适应性和对不确定性的精确测量；（ii）开发高保证自治架构，并建立可验证的可观察性和可控性的切换规则。将开发一个框架，结合高性能和高保证自主性的优势，促进自适应学习，准确的不确定性量化和可验证的安全措施。将开发新的方法(i)基于政策的闭环学习以提高性能，（ii）可靠的不确定性量化以提供数据驱动的适应性，以及（iii）系统级别的可验证的可观察性和可控性。将采用系统的双策略方法来安全收集数据，以实现所提出的高性能自治，以协调数据的三个所需属性：安全性、策略性和闭环。拟议的框架将在严格的程序中进行验证，从模块化模拟测试到集成和部署在真实的空中和地面车辆上。这项研究得到了美国国家科学基金会和开放慈善机构的合作支持。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。

The autonomous systems that operate by themselves (e.g., autonomous cars and delivery drones) expose safety challenges dependent upon the complexity of the missions and environments. Although such systems have demonstrated the ability to learn and adapt on their own, ensuring the safety of these systems is not trivial. Safety can be endangered due to various factors, including but not limited to unexpected changes, inclement weather conditions, or unknown obstacles. The onboard learning solutions of these systems have mostly been used in non-critical situations like computer games, where safety concerns are minimal. This proposal aims to address the urgent need for end-to-end safety in learning-enabled systems across various application scenarios, e.g., self-driving cars and flying vehicles in urban air mobility. We propose a novel solution called "Data-enabled Simplex" or "DeSimplex.” DeSimplex is built on solid mathematical principles and systematic methods for collecting data and using the data for the system’s performance improvement. It provides a framework that can be proven to be safe and allows learning-enabled systems to adjust and perform well even when faced with extreme events, or environmental hazards. The proposed work is crucial for wider applications that involve the safe and efficient operation of autonomous systems in unpredictable, demanding physical environments, including autonomous cars and flying vehicles of 3D urban air mobility.The proposed work lays the groundwork for advancing the comprehension of safety in end-to-end learning-enabled systems, a foundational problem in cyber-physical systems, robotics, and machine learning. We aim to pursue the following two interconnected research thrusts: (i) improving high-performance autonomy with reliable uncertainty quantification methods to ensure data-driven adaptability and precise measurement of uncertainties and (ii) developing high-assurance autonomy architectures and establishing switching rules for verifiable observability and controllability. A framework will be developed that combines the strengths of high-performance and high-assurance autonomy, facilitating adaptive learning, accurate uncertainty quantification, and verifiable safety measures. Novel methods will be developed for (i) on-policy, closed-loop learning to boost performance, (ii) reliable uncertainty quantification to provide data-driven adaptability, and (iii) verifiable observability and controllability at the system level. A systematic, dual-strategy approach will be pursued for safe data collection for the proposed high-performance autonomy to reconcile the three desired properties for data: safety, on-policy, and closed-loop. The proposed framework will be validated in a rigorous procedure from modular simulation testing to integration and deployment on real aerial and ground vehicles.This research is supported by a partnership between the National Science Foundation and Open Philanthropy.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.