CPS: TTP Option: Medium: Collaborative Research: Smoothing Traffic via Energy-efficient Autonomous Driving (STEAD)

研究显示，世界上交通最拥堵的10个城市中有5个在美国。交通拥堵给美国各地的交通系统带来了不必要的负担，提高了运输成本和能源足迹。车辆自动化为减少交通流量和提高交通基础设施的效率创造了机会。特别是，该项目旨在减少幽灵交通堵塞的能源足迹，即密集的交通在没有明显原因的情况下停止，以及拥堵时的走走停停。研究小组的目标是通过将一小部分联网和自动驾驶汽车（cav）插入有驾驶员的正常交通中，也就是有人驾驶的交通，将走走停停的交通拥堵的总能源足迹减少至多40%。这项工作将建立混合自治模型（自动驾驶汽车和有人驾驶交通的组合），并测试这部分自动驾驶汽车以受控方式平滑交通流量的能力，从而减少能源足迹。这项研究结合了数学、控制理论、机器学习和交通工程。该项目包括四所大学，并与行业和政府合作。该项目还将吸引学生和社区利益相关者，包括州和联邦运输机构以及自动驾驶汽车制造商。具体来说，实现交通平滑和减少环境足迹的技术贡献包括用于稀疏控制设置的新的平均场最优控制公式，其中只有一部分车辆是自动驾驶汽车，并且可以控制。研究人员将开发基于深度强化学习的数据驱动控制算法，旨在实现在分析方法推导显式控制器过于复杂的情况下（例如，由于多车道，坡道和人类驾驶风格的高度变化）的控制。他们还将开发基于可满足性模凸优化的工具，以实现这些控制器的安全性和鲁棒性。该方法将首先使用微模拟工具进行验证，以评估其效率和有效性。一旦在模拟中得到验证，该项目将根据系统的实时控制命令，在有人驾驶车辆上进行现场测试，由100名人类驾驶员根据通过手机应用程序传递的目标速度和车道的控制信号执行该算法。之后，该系统将在项目过渡到实践部分的高速公路路段上进行多达20辆自动驾驶汽车的测试。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。

Studies show five of the top 10 most-gridlocked cities in the world are in the United States. Traffic congestion puts undue burden on transportation systems across the United States, raising transportation costs and the energy footprint. Vehicle automation creates an opportunity to reduce traffic and improve efficiency of the transportation infrastructure. In particular, this project aims to reduce the energy footprint of phantom traffic jams, where dense traffic comes to a halt for no apparent reason, and also stop-and-go-waves in congestion. The research team aims to reduce the overall energy footprint of stop-and-go congestion by up to 40% via a small portion of connected and autonomous vehicles (CAVs) inserted into normal traffic with drivers, also known as manned traffic. The work will build models of mixed autonomy (a combination of CAVs and manned traffic), and test the ability for this portion of CAVs to smooth the flow of traffic in a controlled manner, and thus reduce the energy footprint. The research combines mathematics, control theory, machine learning, and transportation engineering. The project includes four universities and engages industry and government partners. The project will also engage students and community stakeholders, including State and Federal transportation agencies and CAV manufacturers.Specifically, the technical contributions enabling traffic smoothing and reduction in the environmental footprint include new mean-field optimal control formulations for sparse control settings where only a subset of vehicles are CAVs and can be controlled. Investigators will develop data-driven control algorithms based on deep reinforcement learning designed to enable control in settings where analytical approaches to derive explicit controllers are too complex (e.g., due to multi-lane, ramps, and high variation of human driving styles). They will also develop tools based on Satisfiability Modulo Convex optimization to enable safety and robustness of these controllers. The approach will first be validated using microsimulation tools to assess their efficiency and their validity. Once validated in simulation, the project will then field test the algorithm with manned vehicles following real-time control commands of the system, executed by 100 human drivers following control signals communicated via a phone app with target speeds and lanes. After which, the system will be tested with up to 20 CAVs inserted onto a freeway stretch in the Transition to Practice component of the project.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.