Intelligent and Integrated Control of V2X-Enabled Autonomous Vehicles using Deep Reinforcement Learning

决策系统不可靠导致的引人注目的自动驾驶汽车故障阻碍了它们的普及。这些故障通常是由于车辆的车载传感器获得的关于其周围环境的有限和不可靠的信息。这种“感官”挑战可以通过利用车联网（V2X）通信来克服，这将使自动驾驶汽车能够与彼此、附近的基础设施和环境进行无线通信。当通过V2X通信接收到的关键信息与典型的感官数据相结合时，自动驾驶汽车可以更准确、更全面地了解周围环境。实现自动驾驶与V2X通信的完美结合，需要智能集成的控制决策系统。然而，如何在一个集成的框架内共同优化车辆控制和通信控制，以实现安全高效的自动驾驶是一个尚未解决的挑战。该系统需要包括两种类型的控制：(1)车辆的自动驾驶控制；(2)车联网通信控制。与传统的控制方法相比，深度强化学习（DRL）可以将环境智能引入联网和自动驾驶汽车中。我的研究计划将提供一个框架来优化基于DRL的智能集成控制决策系统。它将涵盖三项活动：基于drl的车辆控制，通过V2X增强感知和合作。我将开发能够学习“沟通什么，何时沟通”的DRL算法，以及如何在无成本的理想通信假设下，利用沟通的信息做出智能车辆控制决策。基于非理想V2X通信的鲁棒drl车辆控制。我将研究通信改善控制性能的好处与通信成本之间的权衡，目的是在“什么和什么时候进行通信”问题上做出最佳决策。此外，我将为自动驾驶开发强大而安全的DRL算法，可以处理由于不理想的V2X通信而产生的不确定性。基于drl的车辆与无线电资源联合控制，实现自动驾驶与V2X通信的完美结合。“如何通信”，即如何控制稀缺的无线电资源，与车辆控制密切相关，因为通信性能将影响驾驶感知。我的目标是考虑将这两种类型的控制更紧密地集成在一起，以实现支持v2x的自动驾驶的全局最佳解决方案。通过上述活动，我的研究项目有望提高自动驾驶汽车的可靠性，从而加速其采用，缓解道路拥堵，降低燃油消耗，并为乘客提供更舒适、更安全的体验。

High-profile autonomous vehicle failures caused by unreliable decision making systems have impeded their uptake. These failures are often due to the limited and unreliable information obtained by the vehicle's on-board sensors about its surrounding environment. This "sensory" challenge can be overcome by leveraging vehicle-to-everything (V2X) communications, which would enable autonomous vehicles to communicate wirelessly with each other, nearby infrastructure and the environment. Autonomous vehicles can achieve a more accurate and comprehensive picture of its surrounding when the critical information received through V2X communications is combined with the typical sensory data. An intelligent and integrated control decision making system is required to realize the perfect combination of autonomous driving and V2X communications. However, how to jointly optimize vehicle control and communication control in an integrated framework toward safe and efficient autonomous driving is an unsolved challenge. The system would need to include two types of controls: (1) autonomous driving control of the vehicles and; (2) communication control of the vehicular networks. Compared with the traditional control approaches, Deep Reinforcement Learning (DRL) can introduce ambient intelligence into connected and autonomous vehicles. My proposed research program will provide a framework to optimize an intelligent and integrated control decision making system based on DRL. It will span three activities: DRL-based Vehicle Control with Enhanced Perception and Cooperation through V2X. I will develop DRL algorithms that can learn "what and when to communicate", jointly with how to leverage the communicated information to make intelligent vehicle control decisions under the assumption of ideal communications with no cost. Robust DRL-based Vehicle Control with Non-Ideal V2X Communications. I will investigate the trade-off between the benefit of communications to improve control performance and the cost of communications, with the objective of making optimal decisions on the "what and when to communicate" problem. Moreover, I will develop robust and safe DRL algorithms for autonomous driving that can handle the uncertainty due to non-ideal V2X communications. Joint DRL-based Vehicle and Radio Resource Control toward Perfect Combination of Autonomous Driving and V2X Communications. "How to communicate", i.e., how to control the scarce radio resources, is closely interrelated with vehicle control, as the communication performance will impact the driving perception. My goal is to consider tighter integration of the two types of control to achieve a global optimal solution for V2X-enabled autonomous driving. Through the above activities, my research program promises to improve the reliability of autonomous vehicles thereby accelerating their adoption, easing road congestion, reducing fuel consumption, and providing a more comfortable and safer experience for passengers.