Autonomous Driving Learning Preference of Collision Avoidance Maneuvers
Autonomous Driving Learning Preference of Collision Avoidance Maneuvers
复制标题
DOI：
10.1109/tits.2020.2988303
复制
发表时间：
2021-09-01
影响因子：
8.5
通讯作者：
Sonoda, Kohei
中科院分区：
文献类型：
--
作者：
Nagahama, Akihito;Saito, Takahiro;Sonoda, Kohei
最近，随着与SAE自动驾驶2级相对应的自动驾驶系统（ADS）的积极开发，ADS的舒适性受到了极大的关注。在我们以前的研究中，有人提出，ADS的舒适性受到他们的操纵和ADS和驾驶员之间的信息共享程度。然而，即使驾驶员充分了解公认的交通环境和ADS控制的轨迹，驾驶员对ADS的舒适度和信任度也可能不足。这是因为每个驾驶员都有不同的首选轨迹。虽然一些研究人员提出了ADS，学习驾驶员的喜好，这些系统所提出的机动不容易修改，因为离线学习。此外，还对舒适性的主观评价进行了定量研究。本研究提出一个按需学习防撞ADS，学习驾驶员的首选操作，通过驾驶员的干预。在我们的系统中，司机教他们的首选轨迹的系统，只有当他们不满意的机动系统。系统更新参数，并在下一次避让时显示所学习的机动。为了实现所提出的系统，我们采用了修改的风险潜力函数和增益调整方法，以及成本函数和学习方法，逐步和稳定的机动学习。驾驶模拟器实验证明了稳定的轨迹学习和平滑的干预。此外，驾驶员对ADS的学习动作感到满意，并且当使用建议的ADS时，他们对ADS的舒适性和信任度得到改善。
Recently, with the active development of automated driving systems (ADSs) corresponding to SAE automated driving level 2, the comfort of ADSs has gained significant attention. In our previous research, it was proposed that the comfort of ADSs is affected by their maneuvers and the degree of information sharing between ADS and drivers. However, even if the drivers are well-informed of the recognized traffic environment and ADS-controlled trajectory, the comfort and trust of drivers in ADSs could be insufficient. This is because each driver has distinct preferred trajectories. Although some researchers have proposed ADSs that learn driver preferences, the maneuvers presented by these systems are not easily modified, because of off-line learning. In addition, a few quantitative investigations on the subjective evaluation of comfort have been conducted. This study proposes an on-demand learning collision avoidance ADS that learns the preferred maneuvers of drivers through driver intervention. In our system, the drivers teach their preferred trajectory to the system only if they are unsatisfied with the maneuver presented by the system. The system updates the parameters and shows the learned maneuver at the next avoidance. To realize the proposed system, we applied modified risk potential functions and gain-tuning method, as well as a cost function and learning method for gradual and stable maneuver learning. Driving simulator experiments demonstrated stable trajectory learning and smooth intervention. Furthermore, the drivers were satisfied with the learned maneuvers of the ADS, and their comfort and trust in the ADS improved when using the proposed ADS.