NRI: FND: Scalable and Customizable Intent Inference and Motion Planning for Socially-Adept Autonomous Vehicles

该项目将通过解决自动驾驶汽车（AVs）的一个重要而具有挑战性的问题：自动驾驶汽车与人类驾驶汽车的交互，促进科学进步，促进国家繁荣和安全。目前，缺乏能够让自动驾驶汽车以安全且善于社交的方式与周围多辆汽车互动的理论。这个国家机器人计划（NRI）项目将通过开发一种新的算法框架来解决这一关键需求，使自动驾驶汽车能够预测其他车辆的行为，并根据当地的驾驶文化定制其运动。该项目通过推进控制工程、机器学习和认知科学领域的知识，为国家利益服务。该项目还将在实现广泛采用的自动驾驶汽车方面迈出重要一步，这有望提高交通系统的效率和安全性。项目结果将通过项目网站、开源模拟软件和公共数据集发布。该项目的影响将通过各种教育活动扩大，包括新的协作自动驾驶课程，本科生研究项目，以及通过实验室参观向当地社区推广。这个项目旨在回答两个基本的研究问题：1)意图推理和运动规划的什么形式能够创造出适合社交的运动，以及2)这些形式的什么体现可以实现多车交互的可扩展性和改变驾驶文化的可定制性？为了回答这些问题，研究团队将追求以下三个目标。首先，项目团队将建立一个贝叶斯博弈模型来表示车辆交互，并制定机械意图推理和运动规划策略。其次，将开发一个消息传递神经网络，以实现可扩展的意图推理和多辆周围车辆的运动预测。第三，将开发一种社会关注机制，使自动驾驶汽车能够主动优先考虑各种控制因素，例如对周围车辆的安全和礼貌。开发的算法将通过与自动驾驶汽车制造商和研究机构合作设计的驾驶模拟器，在十字路口和高速公路等现实驾驶场景中进行验证。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 will promote the progress of science, and advance the national prosperity and safety, by tackling an important and challenging problem for autonomous vehicles (AVs): interaction of AVs with human-driven vehicles. Currently, there is a lack of theory that allows an autonomous vehicle to interact with multiple surrounding vehicles in a safe and socially-adept manner. This National Robotics Initiative (NRI) project will address this critical need by developing a novel algorithm framework for an autonomous vehicle to be able to anticipate other vehicles, behavior and customize its motion according to the local driving culture. This project serves the national interests by advancing knowledge in the fields of control engineering, machine learning, and cognitive science. The project will also make an important step in making widely adopted autonomous vehicles a reality, which promises to increase transportation system efficiency and safety. Project results will be disseminated through a project website, open-source simulation software, and public datasets. The impacts of this project will be broadened through various educational activities, including a new class on collaborative autonomous driving, undergraduate research projects, and outreach to the local community through lab tours.This project aims at answering two fundamental research questions: 1) what formalisms of intent inference and motion planning are capable of creating socially-adept motions, and 2) what embodiment of these formalisms can achieve scalability for multi-vehicle interactions and customizability for changing driving cultures? To answer these questions, the research team will pursue the following three objectives. First, the project team will build a Bayesian game model to represent vehicle interactions, and develop mechanistic intent inference and motion planning policies. Second, a message passing neural network will be developed to enable scalable intent inference and motion prediction of multiple surrounding vehicles. Third, a social attention mechanism will be developed that allows an autonomous vehicle to actively prioritize its various control considerations, e.g., safety and courtesy towards the surrounding vehicles. The developed algorithms will be validated with real-world driving scenarios such as intersections and highways in a driving simulator designed through collaboration with autonomous vehicle manufacturers and research institutes.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.