SLES: Vision-Based Maximally-Symbolic Safety Supervisor with Graceful Degradation and Procedural Validation

该项目旨在开发新技术，以确保自动驾驶汽车和家用机器人等自主机器人系统的安全性。这种系统的安全性至关重要，因为机器人与物理世界和人互动，在缺乏安全的情况下可能会产生不利的互动结果。大多数下一代机器人系统预计将包含使用机器学习构建的组件。这种基于学习的组件可以产生新的功能，但也容易出现不可预测的行为或故障，使整个系统使用起来不安全。该项目旨在为这一问题制定一个普遍的解决方案。其主要思想是构建一个安全监督器，这是一个软件模块，可以持续监控机器人的动作，并根据需要进行干预以确保安全。安全监督员的作用类似于驾驶教练，他观察练习驾驶员并在必要时接管。该项目开发的技术将广泛用于构建安全有效的机器人系统。该项目的研究通过研究培训、课程开发和推广活动与K12、本科和研究生教育相结合。该项目开发了构建基于视觉的安全监督器的技术，该安全监督器赋予整个系统称为优雅降级的安全属性，这意味着整个系统在不熟悉或未知的情况下不会发生灾难性故障;相反，整个系统将检测到环境的不熟悉性质，并切换到安全和保守的行动。为此，项目团队开发了象征性的场景表示以及推理算法，这些算法可以产生可解释和可验证的安全评估和决策，这些评估和决策对不熟悉的场景具有鲁棒性。为了严格测试和评估安全主管，项目团队开发了程序验证算法：通过程序生成的合成视觉数据进行验证。程序生成是从符号计算机程序生成合成数据的过程，它提供了所有粒度级别的完全控制，并轻松实现长尾事件和新场景的系统模拟。除了程序验证，项目团队还对现实世界的机器人进行评估，重点是导航和重新安排任务。该奖项反映了NSF的法定使命，并通过使用基金会的智力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 project aims to develop new technology to ensure the safety of autonomous robotic systems such as self-driving cars and home robots. Safety for such systems is critical because robots interact with the physical world and people, resulting in potential for adverse interaction outcomes in the absence of safety. Most next generation robotic systems are expected to contain components built using machine learning. Such learning-based components can result in new capabilities but can also be prone to unpredictable behavior or failures, making the full system unsafe to use. This project seeks to develop a general solution to this problem. The main idea is to build a safety supervisor, a software module that continuously monitors the actions of a robot and intervenes as needed to ensure safety. The safety supervisor functions similarly to a driving coach, who watches the practicing driver and takes over when necessary. Techniques developed in this project will be broadly useful for building safe and effective robotic systems. Research in this project is integrated with K12, undergraduate, and graduate education through research training, course development and outreach events.This project develops techniques for constructing a vision-based safety supervisor that endows the full system with the safety property called graceful degradation, meaning that the full system will not fail catastrophically under unfamiliar or unknown scenarios; instead, the full system will detect the unfamiliar nature of the circumstance and switch to actions that are safe and conservative. To this end, the project team develops symbolic scene representations together with reasoning algorithms which produce interpretable and verifiable safety assessments and decisions that are robust to unfamiliar scenarios. To rigorously test and evaluate the safety supervisor, the project team develops algorithms for procedural validation: validation through procedurally generated synthetic visual data. Procedural generation is the process of generating synthetic data from symbolic computer programs, which provide full control at all levels of granularity and easily enable systematic simulation of long-tail events and novel scenarios. In addition to procedural validation, the project team also performs evaluation on real-world robots with a focus on navigation and rearrangement tasks.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.