CAREER: Perceivability: Enabling Safe and Secure Autonomy via Synergistic Control, Observation and Learning

这个学院早期职业发展（Career）基金将通过建立可感知性的概念来解决控制和估计理论中的基本问题：可感知性是一个系统的结构属性，它描述了在面对约束的情况下动态构建环境知识的能力。自主系统（如无人机、自动驾驶汽车）必须能够安全、及时地了解它们所处的环境，使它们能够安全地与人类和彼此互动。这个过程取决于系统的结构（例如，动力学、约束）和目标，以及由于传感器故障而不可靠的信息，或由于恶意行为而不可信的信息。基本的可感知性问题是：“在给定的时间范围内，在给定的动态、传感和通信能力下，在给定的学习算法下，安全地学习给定的环境是否可行？”如果答案是否定的，那么人们可能会想：“系统的哪些参数可以改变，从而可以学习环境？”控制和观察之间的协同作用是什么，可以安全地促进知识的产生？”该项目将开发可感知性的基础，以及计算效率的学习和控制技术，以提高系统的安全性、自主性和弹性。该计划还将辅以一项教育和外展计划，该计划将吸引代表性不足的群体和K-12学生，并通过外展活动和机构STEM计划传播结果。可感知性在系统科学中引入了一个改变游戏规则的概念，旨在弥合学习、估计和控制之间的差距，并使系统工程中的新能力成为可能。如果存在一个安全的控制输入，那么在一定的时间范围内，一个环境被称为可感知的，因此，物理系统的安全轨迹，能够收集环境可以学习的数据。因此，可感知性可以被认为是智能系统的一般属性：可达性和可观察性的合并，将知识构建过程与系统动态和约束紧密联系在一起。该项目将研究系统结构和底层控制、估计和学习机制如何(i)在使用不确定（即错误或恶意）信息的情况下，在给定的时间范围内，在安全的系统轨迹上表征环境是否可感知的能力，以及（ii）如何改变系统结构和/或知识构建机制以实现安全的知识生成。这些创新将使自主系统能够在安全关键和时间关键的情况下完成智能、复杂的任务。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 Faculty Early Career Development (CAREER) grant will address fundamental questions in control and estimation theory by establishing the concept of perceivability: the structural property of a system that describes the ability to build knowledge about an environment dynamically, in the face of constraints. Autonomous systems (e.g., drones, self-driving cars) must be able to safely and timely learn the environment they operate in, to enable them to interact safely with humans and each other. This process depends on the structure (e.g., dynamics, constraints) and goals of the system, and information that is unreliable due to sensor faults, or untrusted due to malicious actions. The fundamental perceivability question is: “Is it feasible to safely learn a given environment for given dynamics, sensing and communication capabilities, under a given learning algorithm, within a given time horizon?” If the answer is negative, then one may wonder: “What parameters of the system can be changed such that the environment can be learned? What are the synergies between control and observation that safely enhance the generation of knowledge?” The project will develop the foundations of perceivability, and computationally-efficient learning and control techniques towards increasing system safety, autonomy and resilience. It will be complemented by an educational and outreach program that will engage underrepresented groups and K-12 students and disseminate the results via outreach activities and institutional STEM programs. Perceivability introduces a game-changing concept in systems science that aims to bridge the gap between learning, estimation and control, and enables new capabilities in systems engineering. An environment is called perceivable within some time horizon if there exists a safe control input, and therefore a safe trajectory of the physical system, that enables the collection of data over which the environment can be learned. Perceivability can thus be thought of as a generalized property of an intelligent system: a merging of reachability and observability that tightly links the knowledge-building process with the system dynamics and constraints. The project will investigate how the system structure and the underlying control, estimation and learning mechanisms (i) enable the ability to characterize whether an environment is perceivable within a given time horizon, over safe system trajectories while using uncertain (i.e., faulty or malicious) information, and (ii) how the system structure and/or the knowledge-building mechanism can be altered to achieve safe knowledge generation. The innovations will enable autonomous systems to accomplish intelligent, complex tasks in safety-critical and time-critical situations.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.