Online Maneuver Learning and Its Real-Time Application to Automated Driving System for Obstacles Avoidance
Online Maneuver Learning and Its Real-Time Application to Automated Driving System for Obstacles Avoidance
复制标题
DOI：
10.1109/tiv.2022.3146622
复制
发表时间：
2023-04
影响因子：
8.2
通讯作者：
Takumi Tatehara;Akihito Nagahama;T. Wada
中科院分区：
文献类型：
--
作者：
Takumi Tatehara;Akihito Nagahama;T. Wada
随着自动驾驶技术的进步，驾驶员逐渐扮演乘客的角色。在驾驶员角色的这种转变期间，已经提出了使规划路径适应各个驾驶员的学习方法，以提高自动驾驶系统（ADS）提供的舒适度和信任度。然而，现有方法应用从在特定情况下至少进行一次的手动驾驶数据中进行的离线学习，即使它们可以接受边驾驶边学习的请求（按需学习）。虽然有几种在线学习方法可用，但它们的学习结果还没有被应用于车辆操纵的真实的时间。专注于避障，我们提出了按需在线学习个人驾驶员的首选路径，并调查所提出的方法是否提高了学习后的ADS提供的舒适性和信任。与现有的方法不同，所提出的ADS可以在自动驾驶和手动驾驶之间平滑和任意地过渡，从而学习首选的机动动作，其结果可以实时应用于避障。因此，建议的ADS包括ADS和驾驶员之间的相互转移的转向权限和曲线修改类似的比例微分控制，根据ADS规划的路径和实际车辆轨迹之间的误差。驾驶模拟和问卷调查的实验结果表明，所提出的ADS方法提高了驾驶员的舒适性和信任度。此外，建议的ADS方法学习的路径偏好的个别司机在回避演习在一定程度上，并逐步调整路径根据驾驶员的偏好，通过反复学习。所提出的方法可能有助于ADS的方便性和快速适应各种交通状况。此外，所提出的方法可以作为在线学习和实时应用到其他ADS操作的基础。
As automated driving technology advances, drivers are gradually taking the role of passengers. During this transition of the driver’s role, learning methods to adapt the planned path to individual drivers have been proposed to improve the comfort and trust provided by automated driving systems (ADSs). However, existing methods apply offline learning from data of manual driving conducted at least once in a specific situation, even if they can accept the request to learn while driving (on-demand learning). Although several online learning methods are available, their learning results have not been applied in real time for vehicle maneuvering. Focusing on obstacle avoidance, we propose on-demand online learning of preferred paths for individual drivers and investigate whether the proposed method improves the comfort and trust provided by the ADS after learning. Unlike the existing methods, the proposed ADS can smoothly and arbitrarily transition between automated and manual driving, thereby learning preferred maneuvers whose results can be applied in real-time to obstacle avoidance. Accordingly, the proposed ADS includes mutual transfer of steering authority between the ADS and driver and curve modification resembling proportional–derivative control according to the error between the ADS planned path and actual vehicle trajectory. Experimental results from driving simulations and questionnaire survey responses show that the proposed ADS method improves the comfort and trust of drivers. In addition, the proposed ADS method learns the paths preferred by individual drivers during avoidance maneuvers to some extent and gradually adjusts the path according to the driver’s preference through repeated learning. The proposed method may contribute to the convenience of ADSs and their fast adaptation to various traffic situations. Furthermore, the proposed method can serve as basis for online learning and its real-time application to other ADS operations.