Collaborative Research: OAC Core: Stochastic Simulation Platform for Assessing Safety Performance of Autonomous Vehicles in Winter Seasons

自动驾驶汽车（AV）的安全性在很大程度上取决于其自动化系统在不同物理环境中部署时的泛化能力（例如感知和决策）。虽然目前自动驾驶汽车的商业化已经被证明可以提高交通安全性，但在冬季恶劣驾驶条件下的自动驾驶汽车安全性能仍然缺乏全面的评估。为了弥补研究差距，该项目旨在开发一个随机仿真平台，该平台可以检查自动驾驶汽车的效率、可靠性和安全性，以防止在广泛的现场实施中出现代价高昂的错误。研究方法以机器学习和物理原理为基础，制定了一种集成和混合的方法来模拟交通流中的随机车辆行为。在不利驾驶条件下，自动驾驶汽车的潜在安全风险将通过车辆行为的动态建模进行评估。该项目将产生一个开源的基于云的仿真平台，允许公众访问测试车辆自动化系统。随着时间的推移，可以通过使用在线机器学习架构来改进仿真模型。这些研究活动将与一系列教育和推广活动紧密结合，包括(i)将先进的计算技术纳入课程，（ii）通过当地K-12的推广工作和夏令营激发年轻一代对科学和工程的兴趣，以及（iii）通过圣地亚哥州立大学的人工智能俱乐部扩大代表性不足的学生群体对计算的参与。西班牙裔服务机构这个多学科研究项目旨在提供改进的模拟算法和计算基础知识，以建立一个先进的网络基础设施工具包。该项目的重点是建立一个随机仿真平台，以评估自动驾驶汽车自动驾驶系统的能力。其目的是开发一种可靠的工具，可以模拟随机车辆行为，研究车辆动力学，并预测冬季恶劣驾驶条件下的潜在自动驾驶安全风险。为此，该项目将首先利用微观交通模型的物理原理来规范模拟车辆交互的机器学习过程。其次，将交通仿真模型与新的车辆动力学模型相结合，对混合交通中多车和单车碰撞概率进行预测；然后研究随机车辆运动，以评估自动驾驶汽车在冰雪路面上的安全性能。第三，这些模型将被集成到一个开源软件包中，包含全面的文档和多个应用案例。预期的交付成果将是一个基于公共云的平台，该平台易于访问，并且能够合并新的数据流以改进模型。在用现场数据验证模型后，该项目将把模拟与现有的自动驾驶系统连接起来进行测试。该项目可以对其他科学和工程领域产生广泛影响，例如物理支持的人工智能，智能和自主系统，以及依赖于模拟数据的其他研究领域。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。

The safety of an autonomous vehicle (AV) highly depends on the generalization capability of its automation systems (e.g., perception and decision-making) when being deployed in diverse physical environments. Although the current commercialization of AVs has been shown to improve traffic safety, AV safety performance under adverse driving conditions in winter seasons still lacks comprehensive evaluation. To bridge the research gap, this project aims to develop a stochastic simulation platform, which examines the efficiency, reliability, and safety of AVs, to prevent costly mistakes in widespread field implementations. The research methods use a foundation of machine learning and physics principles to formulate an integrated and hybrid approach to model stochastic vehicle behaviors in traffic streams. Potential AV safety risks under adverse driving conditions will be assessed with dynamic modeling of vehicle behavior. The project will produce an open-source and cloud-based simulation platform that allows public access to test vehicle automation systems. The simulation models can be improved over time through the use of an online machine learning architecture. The research activities will be closely integrated with a set of education and outreach activities that include (i) incorporating advanced computational techniques into the curriculum, (ii) sparking the interests of younger generations in science and engineering by local K-12 outreach efforts and summer camps, and (iii) broadening the participation of underrepresented student groups in computing through the artificial intelligence club at San Diego State University, a Hispanic serving institution. This multidisciplinary research project aims at contributing improved algorithms in simulation and fundamental knowledge in computing to building an advanced cyberinfrastructure toolkit. The project focuses on producing a stochastic simulation platform that can evaluate the capabilities of AVs' automated driving systems. The motivation is to produce a reliable tool that can model stochastic vehicle behaviors, study vehicle dynamics, and predict potential AV safety risks under adverse driving conditions in winter. To this end, the project will first leverage the physics principles of a microscopic traffic model to regularize the machine learning process for simulating vehicle interactions. Second, both multi-vehicle and single-vehicle crash probabilities in mixed traffic will be predicted by integrating the traffic simulation model with a new vehicle dynamics model. The stochastic vehicle motions will then be studied to assess AV safety performance on icy/snowy pavement. Third, the models will be integrated into an open-source software package with comprehensive documentation and multiple application cases. The expected deliverable will be a public cloud-based platform that is easy to access and is capable of incorporating new data streams for model improvement. After validating the models with field data, the project will connect the simulations with existing automated driving systems for testing. The project can have broad impacts on other science and engineering fields, such as physics-supported artificial intelligence, smart and autonomous systems, and other research domains that depend on simulated data.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.