EAGER: Embodiment of Human Values Profiles in the Control of Autonomous Vehicles

这个Mind，Machine，and Motor Nexus(M3X)早期概念探索性研究奖助金(AGIRE)项目为自动驾驶汽车与其人类驾驶员和乘客之间的可信交互提出了一个新的愿景。尽管自动驾驶汽车正在开发以提高人类安全，但此类车辆的故障是不可避免的，特别是当由于人类疲劳、分心或反应时间相对于自动驾驶汽车的计算速度较慢而无法及时将控制权移交给人类司机时。该项目寻求设计自动车辆控制器，以模拟负责任的人类司机的碰撞行为，以确保“道德优雅”的故障--故障模式很可能符合人类的伦理审查标准。该项目将通过开发可信赖的自动驾驶汽车控制器来模拟负责任的人类司机的实时驾驶和碰撞行为，从而促进科学进步和促进国民健康。通过这样做，该项目将推动自动车辆控制领域的最先进水平，同时还将开发一个车辆控制器设计框架，承诺增加人类对未来自动驾驶汽车的信任。该项目还将开发资源，以鼓励未被充分代表的群体参与STEM领域。这项工作的长期目标是开发一种新的自动车辆控制器设计框架，将模仿负责任的人类司机的碰撞行为和决策，从而展示出“道德优雅”的故障模式。本文研究了两个方面的问题。第一个研究试图通过调查来预测人类的驾驶行为，这些调查能够在一大批受试者中确定“仁慈的”和“权力的”价值观的心理指标。受试者还将参与对撞车场景的虚拟和物理模拟。研究1的目的是描述表达善意和权力价值特征的人在撞车行为上的差异。第二项研究研究了一种新的方法，将不同的值轮廓转换为自动车辆的实时控制器。L的研究采用了一种新的方法，与现有的道义论和结果论方法截然不同，以设计用于AV控制的伦理人工智能系统。通过将低级别实时控制模型的参数与从虚拟和物理模拟中对碰撞场景做出反应的司机收集的数据进行匹配，PI团队将捕捉人类感觉运动控制的每时每刻的变化，这些变化比高级有意识的决策更有可能在碰撞场景中主导反应。如果成功，这项工作可能会赋予未来的自动驾驶汽车值得信赖的控制器，其故障模式可能符合人类的道德审查标准。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This Mind, Machine, and Motor Nexus (M3X) EArly-concept Grant for Exploratory Research (EAGER) project advances a novel vision for trustworthy interaction between autonomous vehicles and their human drivers and passengers. Although autonomous vehicles are being developed to increase human safety, failures with such vehicles are inevitable, especially when timely handover of control to a human driver is infeasible due to human fatigue, distraction, or slow reaction times relative to the computational speed of the autonomous vehicle. This project seeks to design autonomous vehicle controllers that mimic the crash behaviors of responsible human drivers in a way that ensures "ethically graceful" failures - failure modes that are likely to meet human standards of ethical scrutiny. This project will promote the progress of science and advance the national health by developing trustworthy controllers for autonomous vehicles that mimic the real-time driving and crash behaviors of responsible human drivers. By doing so, the project will advance the state of the art in autonomous vehicle control while also developing a framework for vehicle controller design that promises to increase human trust in the autonomous vehicles of the future. The project also will develop resources to inspire involvement of underrepresented groups in STEM fields.The long-term goal of this work is to develop a new framework for autonomous vehicle controller design that will mimic the crash behaviors and decisions of responsible human drivers and thus exhibit "ethically graceful" failure modes. Two studies are researched. The first seeks to predict human driving behavior using surveys capable to identify psychological measures of "benevolent" and "power" Value Profiles in a large cohort of human subjects. Subjects will also engage in both virtual and physical simulations of crash scenarios. The goal of Study 1 is to characterize differences in crash behavior of people expressing benevolent and power Value Profiles. The second study researches a novel approach to translating the different Value Profiles into real-time controllers for the autonomous vehicles. The research l takes a novel approach that is quite different from existing deontological and consequentialist approaches to the design of ethical AI systems for AV control. By fitting the parameters of a low-level, real-time control model to data collected from drivers responding to crash scenarios in both virtual and physical simulations, the PI team will capture the moment-by-moment variations in human sensorimotor control that are more likely to dominate responses during crash scenarios than are high-level conscious decisions. If successful, this work may endow future autonomous vehicles with trustworthy controllers with failure modes that are likely to meet human standards of ethical scrutiny.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.