Learning Cooperative Trajectories in Mixed Traffic

未来的流量可能会与今天有很大不同。目前，全球正在广泛努力通过使用自动驾驶车辆来自动净化交通。但由于在很长一段时间内将会有车辆，这些车辆不能独立驾驶，同时也会有喜欢自己驾驶车辆的司机，至少在一个重要的过渡期内，我们将有自动驾驶和人工驾驶车辆组成的混合交通。然而，目前对自动驾驶汽车的研究大多集中在每辆车的自主性上。本文提出的项目与优先项目“协同交互汽车”无缝对接，研究如何在混合交通场景下有效地规划和控制多个车辆的轨迹，从而探索混合交通中可证明安全的机动和轨迹集合的学习方法。从理论上看，多个代理的最优调度对我们在本项目中所接近的车辆上的实时应用构成了挑战。为此，应使用现有的学习方法，如基于马尔可夫决策过程模型的强化学习和使用公共整体质量函数的深度Q网络。所开发的方法应在优先计划的仿真环境和真实实验车辆中进行集成、演示和验证。

The traffic of the future is likely to be significantly different from today. Currently, there is widespread global effort to autamate traffic by using self-driving vehicles. But as there will be vehicles for a long time, who can not drive independently and at the same time there will be drivers who like to drive their vehicle themselves, at least for a significant transitional period we will have mixed traffic consisting of self-driving and man-driven vehicles. Most of the current research on self-driving vehicles is focused, however, on the autonomy of each individual vehicle. In the project proposed here, which fits seamlessly into the priority program "Cooperative Interacting Automobiles", we investigate how to efficiently plan and control the trajectories of a number of vehicles in a mixed traffic scenario.The aim of the proposed project is thus to explore learning methods for provably safe maneuver and trajectory collectives in mixed traffic. From a theoretical point of view optimal scheduling for multiple agents constitutes a challenge for real-time applications on a vehicle that we approach in this project. For this purpose, current learning methods shall be applied, such as Reinforcement Learning, which is based on Markov Decision Process models, and deep Q-networks using a common holistic quality function.The developed methods shall be integrated, demonstrated and validated in the simulation environment of the priority program and in real experimental vehicles.