自动驾驶汽车与日常生活

作者	Barry Brown
时间	06 February 2017
来源	The Social Life of Autonomous Cars \| IEEE Journals & Magazine \| IEEE Xplore (scut.edu.cn)

Abstract

Until the day comes when all vehicles are fully autonomous, self-driving cars must be more than safe and effi cient—they must also understand and interact naturally with human drivers.	在所有车辆完全自动驾驶的那一天到来之前，自动驾驶汽车必须更加安全和高效，它们还必须理解人类驾驶员并与之自然互动
C ars such as the Tesla Model S and the Volvo XC90 now feature advanced self-driving functions, with tens of thousands of these vehicles on roads worldwide and more appearing every year. In addition, Tesla and other companies like Delphi and Google are testing fully autonomous cars, which have traveled millions of miles on American roads.	特斯拉Model S和沃尔沃XC90等汽车现在具有先进的自动驾驶功能，全球有数以万计的此类汽车上路，每年都有更多的汽车出现。此外，特斯拉和德尔福（Delphi）和谷歌（Google）等其他公司正在测试全自动驾驶汽车，这些汽车在美国公路上行驶了数百万英里。
We’re in the midst of a global fi eld test of autonomous driving technology, yet results from these tests are proprietary, with little publically available data. Occasionally, fl aws in the technology are exposed by videos taken by in-car dashcams and passengers’ mobile phones and uploaded to social media sites like YouTube, prompting media discussion and sometimes controversy. For example, in December 2016, on the fi rst day of a trial launch of a fl eet of Uber self-driving cars with human monitors in San Francisco, a motorist’s dashcam captured one such car driving through a red light and narrowly missing a pedestrian. Hours after the video appeared on YouTube, California authorities forced Uber to cease testing until it had obtained proper permits.1At Stockholm University, we’ve developed a new method that provides a quick, partial view of self-driving systems using public videos.2These videos reveal some of the challenges of adapting autonomous cars to human social activity on the road.	我们正在进行自动驾驶技术的全球现场测试，但这些测试的结果是专有的，几乎没有公开的数据。偶尔，车内行车记录仪和乘客手机拍摄的视频会暴露出该技术的漏洞，并上传到YouTube等社交媒体网站，引发媒体讨论，有时引发争议。例如，在2016年12月，在旧金山，一名司机的行车记录仪捕捉到一辆优步无人驾驶汽车闯红灯，险些撞上一名行人。视频出现在YouTube上几个小时后，加州当局强迫优步停止测试，直到获得适当许可。1在斯德哥尔摩大学，我们开发了一种新方法，可以使用公共视频快速、部分地查看自动驾驶系统。2这些视频揭示了让自动驾驶汽车适应人类在路上的社交活动所面临的一些挑战。

REPURPOSING ONLINE VIDEOS

YouTube is the world’s largest repository of third-party videos. For our fi rst study, we used a range of terms to search this repository for clips involving both semiautonomous cars with driver assistance functionality and fully autonomous test cars. We mostly excluded promotional videos and instead focused on reviews and travelogues, many of which contain long stretches of silent driving and commentaries on system actions.

YouTube是世界上最大的第三方视频存储库。在我们的第一次研究中，我们使用了一系列术语来搜索这个存储库中的剪辑，这些剪辑既包括具有驾驶员辅助功能的半自动汽车，也包括完全自动测试汽车。我们大多排除了宣传视频，而是专注于评论和游记，其中许多包含长时间的无声驾驶和对系统行为的评论。

We collected a corpus of 93 video clips—totaling 10.5 hours—recorded in the US, the UK, Germany, France, Sweden, Hong Kong, Iceland, and Canada. The average length is 9 minutes, with 7 of the clips over 30 minutes.Most illustrate Tesla’s driver-assistance system, Autopilot, and three show similar systems in the Volvo XC90and a Honda Civic. Nine videos, totaling 11 minutes, recorded Google’s self- driving cars. In addition, a South by Southwest (SXSW) presentation on the Google project shows several interesting incidents.

我们收集了美国、英国、德国、法国、瑞典、香港、冰岛和加拿大的93个长达10.5小时的视频剪辑的语料库。平均长度为9分钟，其中7个片段超过30分钟。大多数展示了特斯拉的驾驶员辅助系统Autopilot，三个展示了沃尔沃XC90中的类似系统和本田思域。9段视频共11分钟，记录了谷歌的自动驾驶汽车。此外，谷歌项目（SXSW）演示显示了几个有趣的事件。

THE SOCIAL ROAD

We drew upon well-established linguistic and sociological methods that analyze unobtrusive video recordings of humans to better understand driving behavior and some of the potential problems posed by autonomous vehicles. Perhaps the most interesting revelation comes frominteractions between such cars and other drivers.	我们借鉴了成熟的语言学和社会学方法，这些方法分析了人类的不引人注目的视频记录，以更好地理解驾驶行为和自动驾驶汽车带来的一些潜在问题。也许最有趣的启示来自这类汽车与其他驾驶员之间的互动
Consider Tesla’s Autopilot system, which has three main functions: › adaptive cruise control, which relies on distance to the car ahead to accelerate and decelerate; › auto-steering, which uses road markings and, if that fails, the vehicle ahead, to maintain lane position; and › lane-change assistance, which moves the car to an adjacent lane if it’s safe to do so after the driver activates the turn indicator.	以特斯拉的Autopilot系统为例，它有三个主要功能： ›自适应巡航控制，依靠与前车的距离来加速和减速； ›自动转向，使用道路标记，如果失败，则使用前方车辆，以保持车道位置； ›变道辅助，在驾驶员激活转向指示灯后，如果安全的话，可将车辆移至相邻车道。
These functions are far from fully autonomous driving, but the Tesla can travel for considerable periods of time in autonomous mode.	这些功能远非完全自动驾驶，但特斯拉可以在自动驾驶模式下行驶相当长一段时间。
In the YouTube videos we collected, most of the time Autopilot drives without incident. Yet, due to its simple mechanics, Autopilot sometimes misunderstands other drivers’ actions. While people can often discern other drivers’ intentions as well as mood or character—aggressive, hesitant, selfi sh, unpredictable, and so on— based on changes (or the absence of changes) in their car’s speed, direction, and so on, Autopilot lacks this ability.	在我们收集的YouTube视频中，大多数时候Autopilot都没有发生事故。然而，由于其简单的机制，自动驾驶有时会误解其他驾驶员的行为。尽管人们通常可以根据汽车速度、方向等的变化（或没有变化）来辨别其他驾驶员的意图，以及情绪或性格的攻击性、犹豫、自私、不可预测等等，但Autopilot缺乏这种能力。

图1 特斯拉自动驾驶系统的“粗鲁”行为，因为司机发出信号，表示要向左侧车道移动，为一辆皮卡车驶入高速公路腾出空间。（a）一辆从左车道后方驶来的银色汽车减速，以允许特斯拉司机在卡车进入特斯拉司机（右）前方的右车道时越过（左）。（b）这辆银色汽车的司机认为特斯拉司机拒绝了他的礼貌提议，于是加速进入空档（左），但Autopilot还是执行了变道，打断了另一名司机（右）。
Consider, for example, the fairly common situation depicted in Figure 1. On a divided highway with two lanes in each direction, a Tesla driver is in the right lane in heavy traffic. Ahead on his right he sees a pickup truck merging onto the highway from a connector road. The Tesla driver activates his left-turn signal, indicating his desire to move to the left lane to make room for the truck. A driver in a silver car approaching from behind in the left lane sees the signal and stops accelerating, offering a space for the Tesla driver to enter (Figure 1a). Autopilot, however, doesn’t recognize this polite “gesture” and continues to maintain the Tesla’s position. The silver car’s driver, concluding that the Tesla driver has rejected the offer, starts accelerating again. Autopilot, determining the lane change to still be safe, directs the Tesla to move over (Figure 1b).	例如，考虑图1中描述的相当常见的情况。在一条双向两车道的分车道公路上，一名特斯拉司机在交通繁忙的情况下位于右侧车道。在他右边的前方，他看到一辆皮卡车从一条连接道路汇入高速公路。特斯拉司机启动左转向信号灯，表示他想向左侧车道移动，为卡车腾出空间。一名驾驶银色轿车的司机从左车道后方驶来，看到信号后停止加速，为特斯拉司机提供了进入的空间（图1a）。然而，Autopilot并不认可这种礼貌的“手势”，并继续保持特斯拉的位置。这辆银色汽车的司机断定特斯拉司机拒绝了这一提议，于是再次开始加速。自动驾驶仪确定变道仍然安全，指示特斯拉靠边行驶（图1b）。
The action produced by the Tesla’s robotic coordination might not “bother”another autonomous driving system, but the silver car’s driver understandably perceives this as doubly rude: The Tesla driver first apparently spurns his offer and then cuts in front of him at the last second. The Tesla driver acknowledges this rudeness in the video: “Well, we pulled in ahead of that guy and from what I saw it wasn’t something he was exactly encouraging.” This example demonstrates that, even if autonomous cars can be safe and reliable, they fail to recognize the social nature of the road. Human drivers aren’t always “in tune” with one another, but they’re not just algorithmic agents—in most cases they can detect the same kind of subtle cues that we exhibit in personal interactions.	特斯拉的机器人协调产生的行动可能不会“麻烦”另一种自动驾驶系统，但这辆银色汽车的司机可以理解地认为这是双重粗鲁：特斯拉司机首先显然拒绝了他的提议，然后在最后一秒在他面前插嘴。这位特斯拉司机在视频中承认了这种粗鲁：“好吧，我们把车开到了那个家伙的前面，据我所见，这并不是他真正鼓励的事情。”这个例子表明，即使自动驾驶汽车可以安全可靠，他们也没有意识到道路的社会性质。人类驾驶员并不总是彼此“协调一致”，但在大多数情况下，他们不仅仅是算法代理，他们可以检测到我们在个人互动中表现出的相同类型的微妙线索。

SEEING A GAP AS JUST A GAP

In addition to not understanding what others on the road are doing, autonomous driving systems have trouble recognizing what their own actions might communicate to human drivers. This problem is highlighted in Figure 2, which is extracted from a SXSW 2016 talk narrated by Chris Urmson, principal engineer of Google’s self-driving car project.	除了不了解路上其他人在做什么之外，自动驾驶系统还存在问题认识到他们自己的行为可能会传达给人类驾驶员。图2突出了这个问题，该图摘自谷歌自动驾驶汽车项目总工程师克里斯·厄姆森（Chris Urmson）讲述的SXSW 2016演讲。

图2 人类驾驶员被自动驾驶技术弄糊涂了。一名驾驶红色车辆的司机穿过高速公路，试图挤进两辆谷歌自动驾驶汽车之间的狭窄空间，进入左转车道（如图）。驾驶员可能错误地将第二辆车为保持安全跟车距离而留下的间隙视为进入空间的提议。
In this example, a human driver crosses a four-lane highway and then enters a left-turn lane just before an intersection, wedging his red vehicle in the narrow space between two self-driving cars. Prompting laughter from the audience, Urmson mocks the driver’s behavior: “And of course, we have people who do I don’t know what sometimes on the road, like this guy pulling out between two self-driving cars. You have to ask, what are you thinking?”	在本例中，一名人类驾驶员穿过一条四车道公路，然后在十字路口前进入左转车道，将他的红色车辆楔入两辆自动驾驶汽车之间的狭窄空间。Urmson嘲笑了司机的行为：“当然，我们有一些人，我不知道有时在路上会发生什么，比如这个人在两辆自动驾驶汽车之间停车。你必须问，你在想什么？”
Yet, while the human driver might not show the best judgment, his behavior is understandable. As drivers, we learn the importance of maintaining a safe distance behind the car ahead. This principle is likewise incorporatedinto autonomous driving systems. In this case, the second self-driving car slowly approaches the intersection and leaves a gap between itself and the fi rst self-driving car. However, as the Tesla example shows, gaps in the road aren’t just safe following distances— they can also be interpreted by other drivers as off ers to enter the space. The human driver might have perceived the gap between the self-driving cars as such an off er.	然而，虽然人类驾驶员可能没有表现出最佳判断，但他的行为是可以理解的。作为司机，我们了解到与前面的汽车保持安全距离的重要性。这一原则也被纳入自动驾驶系统。在这种情况下，第二辆自动驾驶汽车缓慢地接近十字路口，并在自己和第一辆自动驾驶车辆之间留下一个间隙。然而，正如特斯拉的例子所示，道路上的空隙不仅仅是安全的跟车距离，其他驾驶员也可以将其解释为进入空间的障碍物。人类驾驶员可能已经意识到自动驾驶汽车之间的差距如此之大。
The driver might also simply have tried to force his way in. People don’t always drive sensibly or legally, especially in situations that can lead to anger or confusion such as in construction zones or heavy traffi c. On busy roadways, some drivers will exploit any space they can fi nd. In this case, the self-driving car approaching the intersection dealt with the human driver’s action appropriately, slowing down and allowing the red vehicle to enter the space.	司机也可能只是试图强行进入。人们并不总是理智地或合法地开车，尤其是在可能导致愤怒或混乱的情况下，例如在施工区或交通拥挤的地方。在繁忙的道路上，一些司机会利用他们能找到的任何空间。在这种情况下，接近十字路口的自动驾驶汽车适当地处理了人类驾驶员的动作，减速并允许红色车辆进入空间。

SOMETIMES IT’S GOOD TO BE A CREEP

Human drivers also interpret other drivers’ inaction. This too can cause problems for autonomous driving systems, as Figure 3 shows.	人类驾驶员也会解释其他驾驶员的不作为。这也会给自动驾驶系统带来问题，如图3所示。

图3 自动驾驶技术带来的攻击性驾驶。（a）一辆谷歌自动驾驶汽车在十字路口右前方的一辆白色汽车的司机面前到达一个四向停车点。（b）自动驾驶汽车在行驶过程中似乎犹豫不决，这促使司机急转弯，导致谷歌汽车突然刹车，险些与后面的车辆相撞。
In this example, a Google self- driving car arrives at a four-way stop just before a driver in a white car on the cross street. Research on four-way stops underlines the importance of creeping into the intersection to let other drivers know that you’re attentive to the situation and ready to take your turn. Although the Google car edges forward, the motion is insuffi cient to signal an urgency to proceed. The driver on the cross street interprets this as hesitation and accordingly moves into the intersection fi rst. This causes the Google car to brake abruptly, much like a novice driver, which in turn causes the driver behind to also stop quickly to avoid a rear-end collision.	在这个例子中，一辆谷歌自动驾驶汽车在十字路口的一辆白色汽车的司机之前，到达一个四向停车点。对四向停车的研究强调了悄悄进入十字路口的重要性，让其他司机知道你已经注意到情况并准备好转弯。尽管谷歌汽车向前推进，但这一动议并不足以表明行动的紧迫性。十字路口上的司机将此解释为犹豫，因此首先进入十字路口。这会导致谷歌汽车突然刹车，就像新手司机一样，这反过来也会导致后面的司机快速停车，以避免追尾。
Our videos also show drivers, apparently annoyed by the Google’s car slowness, tailgating the vehicle to “urge” it forward through intersections.	我们的视频还显示，司机们显然对谷歌的汽车速度慢感到恼火，他们尾随车辆在十字路口“催促”它前进。

THE UNCANNY VALLEY OF AUTONOMOUS CARS

Research on human social interaction has revealed that many of our actions—complaints, invitations, and so on—are preceded by “pre-actions” that communicate our intent and thus prepare others for what we’re about to do. For example, placing a hand on a stand-up microphone is a pre-action to making an announcement. Because of the driving environment’s complexity and the intricate rules of engagement, drivers carry out many pre-actions to avoid potential collisions or to assert themselves in congested conditions.	对人类社会互动的研究表明，我们的许多行为、抱怨、邀请等都是在“预先行动”之前进行的，这些行动传达了我们的意图，从而为其他人准备好我们要做的事情。例如，将一只手放在立式麦克风上是发布公告的预先行动。由于驾驶环境的复杂性和复杂的接战规则，驾驶员会采取许多预防措施，以避免潜在的碰撞或在拥挤的情况下维护自己。
Programming autonomous vehicles to execute and recognize often subtle pre-actions is challenging, which poses a dilemma for designers akin to the “uncanny valley” problem in computer graphics and robotics. As computer- generated human fi gures and humanoid robots become lesscartoonish and more realistic, their residual shortcomings become more visible, sometimes evoking unease or revulsion. Similarly, as self-driving cars become better at the mechanics of driving, yet still not “behave” exactly like human drivers, they could arouse feelings of anger or frustration. For example, always slowing down to avoid a collision—a statistically safe but not always the best response—could encourage drivers to cut off or zoom past an autonomous car, ironically increasing the risk of a collision.	对自动驾驶汽车进行编程以执行和识别通常是微妙的预先动作是一项挑战，这给设计者带来了一个类似于计算机图形和机器人技术中的“uncanny valley”问题的困境。随着计算机生成的人形和人形机器人越来越少卡通化和更加现实化，他们的残余缺点变得更加明显，有时会引起不安或反感。同样，随着自动驾驶汽车在驾驶技术上变得更好，但仍然没有像人类驾驶员那样“表现”，它们可能会引起愤怒或沮丧。例如，总是减速以避免碰撞，这在统计上是安全的，但并不总是最佳的反应，这可能会鼓励驾驶员切断或快速驶过自动驾驶汽车，具有讽刺意味的是，会增加碰撞的风险
In the early days of the automobile, a person walked ahead of a car with a red flag to alert other road users. Until autonomous vehicles can navigate the social road as well as actual roads, should they likewise be equipped with a system to warn human drivers or to signal their intent in ambiguous situations such as four-way stops? This prompts the question of how identifying a vehicle as autonomous would influence human drivers. Would they be more cautious and patient, or try to take advantage of a “dumb” artificial driver?	在汽车的早期，一个人走在一辆挂着红旗的汽车前面，提醒其他道路使用者。在自动驾驶汽车既能在社会道路上行驶，又能在实际道路上行驶之前，它们是否应该同样配备一个系统，以警告人类驾驶员，或者在四向停车等模棱两可的情况下发出其意图的信号？这就引出了一个问题，即将车辆识别为自动驾驶将如何影响人类驾驶员。他们会更加谨慎和耐心，还是试图利用一个“愚蠢”的人工司机？
My goal here isn’t to critique the current generation of self-driving cars, which are still in the early stages of development. Rather, it’s to point out that driving isn’t just a mechanical operation but also a complex social activity. Until the day comes when all vehicles are fully autonomous, self- driving cars must be more than safe and efficient—they must also understand and interact naturally with human drivers. So long as most vehicles on the roadway continue to be operated by people, self-driving car designers must consider how their choices impact other drivers as well as their own vehicles’ passengers. If not, the social road could get a lot bumpier.	我在这里的目标不是批评当前一代的自动驾驶汽车，它们仍处于发展的早期阶段。相反，这是为了指出驾驶不仅仅是一种机械操作，也是一种复杂的社会活动。在所有车辆完全自动驾驶的那一天到来之前，自动驾驶汽车必须更加安全和高效，它们还必须理解人类驾驶员并与之自然互动。只要道路上的大多数车辆继续由人驾驶，自动驾驶汽车设计者就必须考虑他们的选择如何影响其他驾驶员以及自己车辆的乘客。如果没有，社会道路可能会变得非常崎岖。

References

1.J.F. Rodriguez, “Video Appears to Show Uber Self-Driving Car Running Red Light in SF,”San Francisco Examiner, 14 Dec. 2016; www.sfexaminer.com /uber-self-driving-vehicle-appears -launch-red-light-first-day-sf.

2. B. Brown and E. Laurier, “The Trouble with Autopilots: Assisted and Auton-omous Driving on the Social Road,” to be published in Proc. 2017 CHI Conf. Human Factors in Computing Systems (CHI 17), 2017.

BARRY BROWN is a professor of human–computer interaction and research director of the Mobile Life Centre at Stockholm University. Contact him at barry@dsv.su.se.