CRII: RI: Interpretable Framework and Transformative Applications for Viability in Autonomous Agents

该项目通过研究机器人的“可行性”来推进机器人智能方面的知识。可行性是系统在不利的关键事件后自主维持自身或恢复功能的能力，是健壮智能和自治的理想特征。关键事件的一个例子是机器人倒挂着地，在没有外部帮助的情况下无法站起来。其他的例子还有无人水下航行器处于无法恢复的状态，机械臂卡在杂物中，或者由于机械故障而刹车性能不佳的汽车。这个项目的主要目标是为设计可行的（自我恢复的）人工代理开发一个框架。主要的挑战是制定一个可以为各种情况计算的可行性度量。这个度量需要从电机和传感器读数中计算出来。该项目的第一个任务是从控制系统的基本属性中开发一个有效的代理生存能力度量，以及计算效率的计算方法。第二个任务是证明该框架对以下有用问题的适用性：(i)步行机器人的自我恢复和（ii）汽车制动器的自我维护。该项目不仅促进了对机器人可行性的理解，而且为动力系统的控制和分析提出了新的机器学习方法。为了实现其目标，该项目将李雅普诺夫指数的经典概念扩展到“Agent-Induced Lyapunov exponents”（aiile），该指数优先考虑具有多种潜力的状态，并且具有很大的有效可控状态空间。aiile允许人工智能体自主训练，维持自身或恢复能力，而不需要特定于问题的外部提供的奖励功能。本文将发展已知和未知动力学、部分和完全可观察状态下的AILE度量的计算方法。这种能力将使框架具有广泛的适用性。它将在两个变革性应用中进行演示：运动代理的自我恢复，以及汽车制动器中混乱“粘滑”摩擦的控制。这个项目不是按照传统的学科界限来构建的，而是机器学习、动力系统、力学和信息论之间的桥梁。这个项目将为完全自主的智能体打开新的研究方向和技术的大门，这些智能体具有巨大的潜力，可以在危险的情况下取代和/或协助人类，比如驾驶。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 project advances the knowledge in robot intelligence via research on “viability” of robots. Viability is the ability of a system to autonomously maintain itself or recover functionality after an adverse critical event and is a desired feature of robust intelligence and autonomy. An example of a critical event is a robot landing upside down and being unable to stand up without an external assistance. Other examples are an unmanned underwater vehicle in an unrecoverable state, a robotic arm getting stuck in clutter, or a car with poor braking performance due to a mechanical failure. The principal objective of this project is to develop a framework for the design of viable (self-recoverable) artificial agents. The main challenge is to formulate a metric for viability that can be calculated for various scenarios. This metric needs to be computable from the motor and sensor readings. The first task of this project is to develop an effective metric for agent viability from the essential properties of control systems, along with computationally efficient methods for its calculation. The second task is to demonstrate the applicability of this framework for useful problems in (i) self-recovery of walking robots and (ii) self-maintenance of car brakes. This project not only advances the understanding of robot viability, but also proposes new machine-learning approaches for control and analysis of dynamical systems.To achieve its goals, this project will extend the classical notion of Lyapunov exponents towards ’Agent-Induced Lyapunov Exponents’, AILE, which prioritize states with diverse potentialities, and with a large space of effectively controllable states. AILE allows artificial agents to train autonomously, maintaining themselves or recovering capabilities without the need for problem-specific externally provided reward functions. Computational methods will be developed for the calculation of the AILE metric with known and unknown dynamics, and with partially and fully observable state. This capacity will allow for the broad applicability of the framework. It will be demonstrated in two transformative applications: self-recovery of locomotion agents, and control of chaotic ’stick-slip’ friction in car brakes. This project is not framed along the lines of traditional disciplinary boundaries, but rather bridges between machine learning, dynamical systems, mechanics, and information theory. This project will open doors to new research directions and technology for fully-autonomous agents with an enormous potential for replacing and/or assisting humans in risky situations, such as driving.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.