Principled Robotic Decision Making via Reachability

通过可达性进行有原则的机器人决策

• 批准号: RGPIN-2019-04605
• 负责人: Chen, Mo
• 金额: $ 2.04万
• 依托单位: Simon Fraser University
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2022
• 资助国家: 加拿大
• 起止时间: 2022-01-01 至 2023-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=750550
• 关键词: Principled; Robotic; Decision; Making; via

Principled Robotic Decision Making via Reachability Many autonomous systems, such as unmanned aerial vehicles and autonomous cars, have potential to make great positive impact in many aspects of our lives. However, currently robotic systems tend to operate in controlled environments, often in the absence of other agents. To realize their potential, robotic systems still need to become both safer so that they do not harm humans and other robots, and smarter so that they and humans can achieve better mutual understanding. This can be done through more principled robotic decision making algorithms that account for both domain knowledge and real-world data. To make robots safer, we use reachability analysis, a safety verification method that involves computing the reachable set based on the dynamics, or possible behaviours, of a robotic system. The reachable set quantifies the set of states or configurations of a robotic system from which a target set of states can be reached. This target set of states may represent a set of goal states which are desirable, or a set of dangerous states which are undesirable. Reachability analysis has previously been successfully applied to small scale robotic systems with general nonlinear dynamics, under the influence of process noise, disturbances, and other agents in the environment. Since it accounts for key prior properties of robotic systems such as dynamics and control action constraints, reachability analysis provides a principled decision making framework that my research program aims to advance. To make robots smarter, machine learning has been used extensively; its data-driven nature has the potential to allow robots to sense their environment and to perform complex tasks. However, currently a large amount of data is needed for a robotic system to learn to perform a task, and the learned behaviour transfers poorly across different environments and tasks. Reachability analysis has the potential to help overcome these challenges by summarizing key properties of robotic systems into a form that is compatible with machine learning algorithms. Conversely, machine learning also has the potential to address challenges in safety verification by inferring human intent. In this way, reachability analysis is a bridge that connects traditional principles of robotic decision making to modern data-driven techniques. My research program aims to develop practical, principled robotic decision making algorithms based on reachability analysis. At a high level, the objectives are two-fold: 1. To make robots safer by addressing the computational challenges in reachability analysis and making reachability analysis more practical through its incorporation with realistic perception systems 2. To make robots smarter and more understandable by utilizing reachability to augment machine learning algorithms, and by incorporating insights from machine learning into reachability in tasks such as human-robot interactions

基于可达性的原则性机器人决策许多自主系统，如无人机和自动驾驶汽车，都有可能在我们生活的许多方面产生巨大的积极影响。然而，目前机器人系统倾向于在受控环境中运行，通常在没有其他代理的情况下运行。为了实现它们的潜力，机器人系统仍然需要变得更安全，以便它们不会伤害人类和其他机器人，也需要变得更智能，以便它们和人类能够实现更好的相互理解。这可以通过更有原则的机器人决策算法来实现，这些算法既考虑了领域知识，也考虑了现实世界的数据。为了使机器人更安全，我们使用可达性分析，这是一种安全验证方法，涉及基于机器人系统的动力学或可能的行为计算可达集。可达集合量化机器人系统的状态集合或配置，从中可以到达目标状态集合。该目标状态集可以表示期望的一组目标状态或不期望的一组危险状态。可达性分析已被成功地应用于受过程噪声、扰动和环境中其他因素影响的具有一般非线性动态的小规模机器人系统。由于它考虑了机器人系统的关键先验属性，如动力学和控制动作约束，可达性分析提供了一个原则性的决策框架，我的研究计划旨在推进这一框架。为了让机器人变得更智能，机器学习得到了广泛的应用；它的数据驱动性质有可能使机器人能够感知环境并执行复杂的任务。然而，目前机器人系统需要大量的数据来学习执行任务，并且学习的行为在不同的环境和任务中传输很差。可达性分析通过将机器人系统的关键属性总结为与机器学习算法兼容的形式，有可能帮助克服这些挑战。相反，机器学习也有可能通过推断人类意图来应对安全验证方面的挑战。通过这种方式，可达性分析是连接传统机器人决策原则和现代数据驱动技术的桥梁。我的研究计划旨在开发实用的、基于可达性分析的原则性机器人决策算法。在更高的层面上，目标有两个：1.通过解决可达性分析中的计算挑战使机器人更安全，并通过将可达性分析与现实感知系统相结合使可达性分析更加实用2.通过利用可达性来增强机器学习算法，并通过将来自机器学习的见解结合到诸如人-机器人交互等任务的可达性中，使机器人更智能和更容易理解