CAREER: Generalizable and Reliable Behavior Synthesis in Uncertain Open-World Environments

对于在精心控制的环境之外的情况下可靠和有用的人工智能系统的需求，人们达成了广泛的共识。这个项目的重点是开发能够在“开放世界”环境中安全规划和行动的自主智能体，在这种环境中，智能体对其将要使用的环境的信息有限。这些智能体可能不确定它们可能遇到的对象的数量、类型和身份，以及它们之间的关系。此外，这些属性的不确定性可能是“非平稳的”，这意味着在代理部署期间环境可能会发生变化。该项目的成果将通过开发在现实的非平稳、开放世界环境中计算安全可靠的人工智能行为的新方法，帮助增加人工智能系统的范围和适用性。为了使人工智能系统更广泛地可访问，该项目还将开发一个自主互动教程系统，用于教授学生不同类型的人工智能规划问题及其解决方案表示。拟议的活动将为理解开放世界规划问题的计算性质发展新的原则和分析方法。它将产生从基于逻辑和概率的方法到人工智能以及理论计算机科学的原理和算法的广泛融合。特别是，它将开发新的表示，以有效地表达开放世界规划问题的定性和决策理论公式，以及解决这些问题的有效算法和实现，同时使用抽象来提高效率和可泛化性。将开发新的方法来利用统计学习技术来提高计算效率，同时确保计算出的代理行为满足开放世界环境中对安全性和可靠性的期望要求。项目期间取得的成果和进展将在具有当代机器人平台的物理和模拟测试台上进行评估。问题生成器和模拟试验台将作为基准公开提供，以帮助再现性并促进这一研究领域的进展。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。

There is broad consensus on the need for AI systems that are reliable and useful in situations going beyond carefully controlled environments. The focus of this project is on developing autonomous agents that can plan and act safely in “open-world” settings, where the agent has limited information about the environment where it will be used. Such agents may be uncertain about the numbers, types and identities of objects that they may encounter, as well as about the relationships between them. Furthermore, the nature of uncertainty about these properties may be “non-stationary”, meaning the environment may change during the agent’s deployment. The outcomes of this project will help increase the scope and applicability of AI systems by developing new methods for computing safe and reliable AI behavior in realistic non-stationary, open-world settings. In order to make AI systems more broadly accessible, this project will also develop an autonomous interactive tutorial system for teaching students about different types of AI planning problems and their solution representations. The proposed activity will develop new principles and analytical methods for understanding the computational nature of open-world planning problems. It will engender broad convergence of principles and algorithms from logic-based and probabilistic approaches to AI, as well as from theoretical computer science. In particular, it will develop new representations for efficiently expressing qualitative and decision-theoretic formulations of open-world planning problems along with efficient algorithms and implementations for solving them while using abstractions for efficiency and generalizability. New methods will be developed to utilize statistical learning techniques for enhancing computational efficiency while ensuring that the computed agent behavior meets desired requirements on safety and reliability in open-world settings. The results and progress made during the project will be evaluated on physical and simulated testbeds featuring contemporary robotics platforms. Problem generators and simulated testbeds will be made publicly available as benchmarks to aid reproducibility and spur progress in this area of research.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.