CAREER: Learning Generalizable and Interpretable Embodied AI with Human Priors

本研究旨在利用人工输入来促进自主软件代理学习可推广和可解释的AI控制器。具体来说，人们的输入将被纳入学习框架，以离线（培训前）和在线（培训期间）反馈的形式对学习进行偏向。这种方法在学习问题代理的三个基本组成部分方面是创新的：环境、代理的表示和学习管道。研究者将首先开发生成模型，从人们收集的图像中学习，并为从中学习的代理创建不同的交互环境。研究者将专注于解释自主代理学习到的内部表征。最后，研究人员将整合从与人的在线互动中获得的反馈，以确保最终的人工智能是安全的，并符合人类的偏好。该项目的教育目标是通过将人工智能作为跨学科学科进行教学，将计算机科学、工程学、神经科学和社会科学的教师和学生联系起来。研究者将利用该项目的可解释性框架来帮助学生了解人工智能的内部工作原理，并通过模拟环境使用交互式学习方法来创建沉浸式教育体验，并激励来自不同背景的学生学习自主代理之间的潜在数学原理。本研究的重点是将人类纳入学习具身智能体的三个基本组成部分：环境、智能体表征和学习过程：(1)环境：研究者将设计一个机器学习模型，根据人类经验生成和模拟不同的环境。它将显著提高训练环境的多样性和复杂性，使训练后的智能体能够更好地将其获得的技能推广到未知的情况。(2)智能体表征（Agent representation）：研究者将解读嵌入智能体的学习表征，发现可解释的运动原语，以便人类能够理解人工智能的内部结构，并在具有挑战性的未知场景中控制人工智能的行为。(3)学习过程：研究者将开发无奖励的学习和适应方法，以纳入主动的人类反馈，大大提高人工智能的一致性和安全性。研究人员将对室内和室外的人工智能任务进行研究，包括自动驾驶、家用机器人导航、有腿机器人运动。这个研究项目将人类无缝地集成到机器学习循环中，将可泛化和可解释的嵌入代理带到现实世界的应用中。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 research aims to leverage human input to facilitate learning generalizable and interpretable AI controllers for autonomous software agents. Specifically, input from people will be incorporated into the learning framework to bias the learning in the form of both offline (before training) and online (during training) feedback. This approach is innovative in terms of three fundamental components of the learning problem agent: the environment, the agent’s representation, and the learning pipeline. The investigator will first develop generative models to learn from images collected from people and to create diverse interactive environments for the learning agent from them. The investigator will then focus on interpreting the internal representation learned by the autonomous agent. Lastly, the investigator will incorporate feedback obtained from online interactions with people to ensure that the resulting AI is safe and aligned with human preferences. The education objective of the project is to connect faculty and students from computer science, engineering, neuroscience, and social sciences by teaching AI as an interdisciplinary subject. The investigator will leverage the project’s interpretability framework to help students understand the AI's inner workings and use the interactive learning methods through simulation environments to create immersive educational experiences and motivate students from diverse backgrounds to learn the underlying mathematical principles between autonomous agents.This research focuses on incorporating humans into three foundational components of learning an embodied agent: the environment, the agent representation, and the learning process: (1) Environment: The investigator will design a machine-learning model to generate and simulate diverse environments from human experiences. It will significantly improve the diversity and complexity of the training environments such that the trained agent can better generalize its acquired skills to unseen situations. (2) Agent representation: The investigator will interpret the learned representation of the embodied agent and discover interpretable motion primitives so that humans can comprehend the AI's internals and control the AI's behaviors in challenging unseen scenarios. (3) Learning process: The investigator will develop reward-free learning and adaptation methods to incorporate active human feedback, substantially improving AI alignment and safety. The investigator will ground the research on indoor and outdoor embodied AI tasks, including autonomous driving, household robot navigation, and legged robot locomotion. This research program seamlessly integrates humans into the machine-learning loop to bring generalizable and interpretable embodied agents to real-world applications.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.