CAREER: Safety Assurances for Perception-Enabled Robotic Systems

从自动驾驶汽车到自主无人机，机器学习驱动的感知组件构成了现代自主系统和机器人的核心部分。自主系统能力主要是通过现代机器学习方法的能力来实现的，它可以优雅地处理丰富的感知输入和输出，从而产生有用的控制信息，最终使机器人能够在新的情况下根据他们所看到的做出智能决策。然而，感知失败可能会引发灾难性的机器人故障，并危及人类安全，最近发生的自动驾驶汽车事故就是一个例子。因此，确保机器人在学习驱动、基于感知的控制器下安全运行，对于使其在高完整性和安全关键应用中得到采用至关重要。本项目将建立一个基础框架，为基于感知控制器的闭环系统提供持续的安全保证，其中在培训时临时提供保证，并在运行时（或运行时）持续监测、更新和改进。具体而言，该项目将：(a)开发用于学习构建稳健感知策略的新技术；(b)为感知策略构建安全监视器，以确保其在运行期间的安全运行；(c)开发一种有原则的方法来大规模挖掘闭环感知故障，并利用它们随着时间的推移提高机器人的安全性。这些结果将通过对异构物理机器人试验台以及逼真模拟器的全面评估来建立，重点是自主检查和自主飞机着陆任务。开发安全感知驱动系统的能力将对广泛的机器人应用产生直接、积极的影响，这些应用对安全性和可靠性非常重要，例如关键基础设施的监控、服务或交付机器人以及自动驾驶汽车。这种影响将通过以下方式得到加强：(a)一项综合教育和推广计划，旨在促进机器人安全讨论，并教育各级教师和学生：K-12、本科生和研究生，以及更广泛的机器人研究界；(b)与业界和规管机构密切合作；(c)专注于传播代码库和实现，并为新的机器人课程提供开源课程材料，包括轮式机器人的动手实验。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。

From self-driving vehicles to autonomous drones, machine learning-driven perception components constitute a core part of modern autonomous systems and robots. Autonomous system capabilities are primarily enabled by the ability of modern machine learning methods to elegantly process rich perceptual inputs and outputs so as to produce useful information for control, ultimately enabling robots to make intelligent decisions in novel situations based on what they see. However, perception failures can cascade to catastrophic robot failures and compromise human safety, as exemplified by recent self-driving car accidents. Therefore, ensuring safe robot operation under learning-driven, perception-based controllers is paramount to enable their adoption in high-integrity and safety-critical applications. This project will establish a foundational framework for providing continual safety assurances for closed-loop systems under a perception-based controller, wherein assurances are provided provisionally at training time, and continually monitored, updated, and improved during operation-time (or runtime). In particular, this project will: (a) develop novel techniques for learning robust-by-construction perception policies; (b) construct safety monitors for perception policies to ensure their safe operation during runtime; and (c) develop a principled approach to mine closed-loop perception failures at scale and use them to improve robot safety over time. These results will be grounded through a thorough evaluation on a heterogeneous physical robotic testbed, as well as photorealistic simulators, with a focus on autonomous inspection and autonomous aircraft landing tasks. The ability to develop safe perception-driven systems will have a direct, positive impact on a broad range of robotics applications where safety and reliability are of high importance, such as surveillance of critical infrastructure, service or delivery robots, and autonomous cars. This impact will be enhanced through: (a) an integrated education and outreach plan designed to facilitate robot safety discussions and educate faculty and students at all levels: K-12, undergraduate and graduate students, and the broader robotics research community; (b) close collaborations with industry and regulatory bodies; and (c) focusing on disseminating codebases and implementations, and open-sourcing curriculum materials for a new robotics course including hands-on labs with wheeled robots.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.