And we’re reasonably sure that by the time people turn 16 years old, they’re mature enough to begin operating a motor vehicle. We verify that by demanding that each young driver pass a test before getting their driver's license and being allowed out on the road. Shouldn’t we consider doing something similar for autonomous vehicles to, you know, verify they’re mature enough to drive by themselves?
我们有理由坚信，当人们16岁时，他们已经足够成熟，可以开始驾驶汽车了。 我们要求每个年轻驾驶员在获得驾驶执照并被允许上路之前，需要通过测试，以此来验证这一观点。 我们是否应该考虑对AV做类似的测试，以确认它们是不是足够“成熟”到能自行驾驶？

PHIL KOOPMAN: I often hear people say we should just use a road test like a person would get to get their driver’s license to prove self-driving cars are safe enough. That can help, but I don’t think it’s a complete solution.
PHIL KOOPMAN: 我经常听到人们说，我们应该进行道路测试，就像人们要获得驾驶执照那样的，来证明AV足够安全。 这样做是有一定帮助，但我不认为这是一个完整的解决方案。

We keep coming back to the subject of autonomous vehicles because the problem of making a self-driving vehicle safe has so many aspects to it. One of those facets is figuring out how to certify that a driver is competent. We have certain expectations from a human driver. How do we figure out if a self-driving vehicle is competent?
我们一直在谈论AV这个话题，因为保证一辆AV的安全，是个涉及到多方面的问题。其中之一是得弄明白，如何证明驾驶员能胜任驾驶工作。我们对驾驶员有定量的预期值。但我们如何确定AV是否能胜任呢？

Phil Koopman is the chief technical officer of Edge Case Research and a professor at Carnegie Mellon University. He's one of the key authors of a standard called UL 4600 that helps manufacturers of almost any autonomous system evaluate if they’ve done adequate safety testing. Here’s Koopman on a proposal to determine if a robocar is ready to hit the roads on its own.
Phil Koopman是Edge Case Research的首席技术官，也是卡内基梅隆大学的教授。 他是UL 4600标准的主要制定者之一，该标准可帮助几乎所有自动系统的制造商，评估其是否进行了充分的安全测试。 这是Koopman提出的一项建议，用于评判AV是否已准备好自行上路。

PHIL KOOPMAN: I often hear people say we should just use a road test like a person would get to get their driver’s license to prove self-driving cars are safe enough. That can help, but I don’t think it’s a complete solution. In my mind, a driving test has three pieces. The first is the written test — you have to know all the rules of the road. The second piece is the driving skills test and that’s what people tend to concentrate on. A driving skills test can certainly make sure that you have proficiency at basic maneuvers, I mean it’s necessary, but it’s not everything you need. The most important part is actually the third part which is you have to prove you’re a 16-year old human being or whatever the age is in your jurisdiction. That’s a test too. This is tricky because for machine learning-based systems, they don’t have what’s called general intelligence. They don’t have common sense. And so, being able to prove you have the maturity to deal with unusual situation, things you’re not expecting, is really the hard part here.
PHIL KOOPMAN: 我经常听到人们说，我们应该进行道路测试，就像人们要获得驾驶执照那样的，来证明AV足够安全。 这样做是有一定帮助，但我不认为这是一个完整的解决方案。在我看来，驾驶考试分为三部分。第一个是笔试——你必须了解所有道路法规。第二部分是驾驶技能测试，这是人们倾向于关注的重点。驾驶技能测试当然可以确保你熟练掌握基本操作，这是必要的，但不是你所需的全部。最重要的实际上是第三部分——你必须证明自己已经16岁，或是你所在司法管辖区的适驾年龄。这也是测试的一部分。这很棘手，因为对基于机器学习的系统来说，它们没有所谓的通用智能，它们没有常规意识。因此，最难的部分就是如何能够证明自动系统有能力应对突发情况，应对预期意外的事物。

BRIAN SANTO: All right. So what I think we just heard Phil telling us is that, with humans, there is some minimum level of maturity that we can expect at a certain point, the age of 16, and then we need to test to make sure that they actually do have. We give these kids drivers' tests. With a computer, we can't do that. With a computer, we can't assume some base level of maturity, some base level of capability. So we need two things, if I'm understanding it correctly. One is the ability to demonstrate that there's some base level of capability, and the other thing is some assurance that when we test the computer we're testing what we think we're testing.
BRIAN SANTO: 好的。我认为我们刚刚听到Phil告诉我们的是，对于人类，我们可以预期在某个特定的年龄点，如16岁，达到某种最低限度的成熟度，然后我们需要的就是进行测试，确保他们确实是具备了预期的成熟度。我们对这些孩子进行驾驶测试。但对于计算机，我们无法照做。 对于计算机，我们不能假定某些基本水平的成熟度，或是某些基本水平的能力。 因此，如果我理解正确的话，我们需要做两件事。 一是能够证明AV拥有一定基本水平的能力，另一件是能保证在我们进行计算机测试时，我们所测试的，确实是我们认为需要测试的。

JUNKO YOSHIDA: Right. I think your second one is a good one. I think that because an autonomous vehicle today-- I mean, also tomorrow-- is using machine learning, as far as we know, AI today doesn't have a general AI. In other words, that machine learning doesn't have any common sense so to speak, right?
JUNKO YOSHIDA: 对。 我认为你的第二个观点很棒。 我的看法，因为现在——也包括未来——AV正进行机器学习，据我们所知，今天的AI也没有通用智能。 换句话说，机器学习没有任何普遍认知，可以这么说吗？

BRIAN SANTO: Right.
BRIAN SANTO: 是这样的。

JUNKO YOSHIDA: So you can't really teach a machine to have common sense. So really the biggest challenge here is that when we're talking about an autonomous vehicle, A) it's neither a human; B) now is 16 years old. So that's a big gap we have to fill in before we let these robocars roam around the public street, like the street that we drive, right?
JUNKO YOSHIDA: 所以你无法真正教会机器，让它具有常识。 因此真正的最大挑战是，当我们谈论AV时，A）它不是人类； B）也不是16岁。 因此，在让这些AV在公共街道（例如我们驾驶的街道）上路行驶之前，我们必须迈过这样一个大坎，对吗？

BRIAN SANTO: Right, right. Now, let's go back to one of the examples from your story, because it illustrates the difference between a human and a car. A human tends to be able to... a 16-year-old or thereabouts, we expect a 16-year-old human to have a certain level of discernment based on their experience. So when a human sees a ball bounce into the street, a human will probably think, Oh, somebody was playing with that and they might go to retrieve it. There's probably some kid who's about to jump into a street. When an autonomous vehicle sees a ball, it's like, Oh, it's a round object and I shouldn't hit it. But it doesn't anticipate that there might be some kid attached to it, right?
BRIAN SANTO: 是的，是的。 现在，让我们回顾你故事中的一个例子——描绘了人与车之间的区别。 一个人趋向于能够... 一个16岁左右的人，我们预期中16岁的人，能通过他们的经验，获得一定的辨别力。 因此，当一个人看到一个球弹到街上时，他可能会想，哦，是有人在玩那个球，他们可能会跑过来捡球。那么可能有个孩子会突然跑到街道中。 但是当AV看到一个球时，就像，哦，这是一个圆形的物体，我不应该撞它。 AV并不能预料到会有孩子可能伴随这个球一起出现，对吧？

JUNKO YOSHIDA: Or it's a round thing, so I can just hit it.
JUNKO YOSHIDA: 也可能是，这是一个圆形的东西，所以我可以撞它。

BRIAN SANTO: Right. It bounces. It should be able to handle it if I hit it at 30 miles an hour, right?
BRIAN SANTO: 对的， 它会蹦跶。 如果我以每小时30英里的速度撞它，应该能够搞定它，是这样吗？

JUNKO YOSHIDA: Yeah. So prediction is really the hardest thing for AI I was told. You can actually perceive something, whether it's an object that we shouldn't hit or we could hit. But really the biggest thing is that, what is going to happen next? That's the prediction. That's the hard part.
JUNKO YOSHIDA: 是的。对于AI来说，预判确实是最难的事情。 实际上，你可以预知到某些东西，判断这个物体我们应不应该撞到。 不过说真的，最重要的事情是，接下来会发生什么？ 那是一种预判，是很困难的部分。

BRIAN SANTO: And that's the kind of the thing that a human should be able to do when you're in a driver's test, you're checking to see whether that kid can anticipate, understands the rules of the road. And that kid can demonstrate it.
BRIAN SANTO: 这就是人们在进行驾照考试时，应该能做到的事。 你会测试这孩子是否可以预见并理解道路规则。 然后这孩子可以达到这要求。

JUNKO YOSHIDA: Yeah.
JUNKO YOSHIDA: 是的。

BRIAN SANTO: When you do that... And you can give them a license, right? There's some approval that says, Yes, you've demonstrated the minimum level. But with an autonomous vehicle, it might be able to demonstrate some minimum level of awareness, of the ability to drive. Is there a standard? We've been talking about UL 4600. Does UL 4600 give you any framework for figuring out whether you could give an AV a license?
BRIAN SANTO: 当你测试完...你就可以给他们驾照了，对吗？ 有些人表示赞同，你已经达到最低要求。 不过对于AV，它可能具备展现出最低驾驶能力的意识。 现在对此有标准可参考吗？ 我们一直在谈论UL4600。UL4600是否为人们提供了一个框架，以确定是否可以授予一架AV许可证？

JUNKO YOSHIDA: Well, here's the thing. UL 4600 is really very different from normal safety standards. Normal safety standards usually tell you that you follow, I don't know, you follow A, B, C. Therefore your car is safe. I mean, that's a normal, logical way of thinking, right? But what UL 4600 asks, as Phil Koopman tells you, is that, have you done enough of safety? In other words, it doesn't tell you how to do it; it doesn't tell you what needs to be done; but it does tell you, have you done enough. In other words, they are asking the developers of autonomous vehicles to think of all the things that could happen to your cars, and we don't care how you did it, but tell us your strategy. How did you design your system so your system can reasonably... your system can find a way to mitigate or minimize the risk, right? So what they are asking for is "reasonable" safety.
JUNKO YOSHIDA: 事情是这样的。 UL 4600实际上与正常的安全标准有很大不同。 正常的安全标准通常会告诉你要遵守，要遵循A，B，C。照做之后，你的汽车就是安全的。 我是说，这是正常的逻辑思维方式，对吗？ 但是，正如Phil Koopman告诉你的，UL 4600要求的则是，你是否为保证安全性做了足够工作？ 换句话说，它不会告诉你如何参照执行；它不会明确告诉你需要做什么； 但它确实引导你，了解你做得是否足够。 换种说法，他们正在要求AV的开发人员要考虑到汽车可能发生的所有事情，我们不在乎你是如何做到的，而是需要你告诉我们你的应对策略。 你是如何设计系统的，以使你的系统可以合理地……可以找到一种降低或最小化风险的方法，对吗？ 他们要求的是“合理的”安全性。

BRIAN SANTO: To figure out how to get your autonomous vehicle to be at least equivalent to a 16-year-old kid.
BRIAN SANTO: 为了弄明白如何让你的AV至少具备相当于一个16岁孩子应有的常识。

JUNKO YOSHIDA: Right. Right.
JUNKO YOSHIDA: 对，对。

BRIAN SANTO: So that's the gap that Koopman talks about. That's kind of his point in bringing up a 16-year-old. You need some criteria by which you can license an autonomous vehicle as a reasonable driver, right?
BRIAN SANTO: 这就是Koopman所说的差距。 这就是他提到16岁孩子这个例证的观点。你需要一些标准。通过这些标准，你可以给AV授予许可，认证其为合格的驾驶员，对吗？

JUNKO YOSHIDA: Exactly. You know, you have kids, right? You have daughters. And then you always ask your daughters, Did you think of that? When something really bad happens. And essentially the UL 4600 is asking AV designers, Have you thought of that? Did you think of that? And they could say, Yeah, we thought of that. And here's how we did it. Or here's how we do it, right?
JUNKO YOSHIDA: 确实如此。你有孩子，对吗？ 你有女儿。 然后当真正糟糕的事情发生时。你总是会问你的女儿们，你想到这样的情况了吗？ 从本质上讲，UL 4600在问AV的设计人员，你预想过这样的场景吗？ 你有去想吗？ 他们可以说，是的，我们想到了，这就是我们的应对方法， 或是说我们的处理办法，对吧？

BRIAN SANTO: Right. So there is an additional step after UL 4600 is what I'm hearing. There's that step where you need some sort of test, some way of discerning that the AV that you produce by conforming to the expectations of UL 4600, the AV you produce is doing what you're expecting. Is that right?
BRIAN SANTO: 对。我听说，在通过UL 4600之后，还有一个额外的步骤。 在这一步中，你需要进行某种测试，通过某种方式辨别出你所生产的，已通过符合UL 4600要求的AV，正按照你的预期运行。是这样的吗？

JUNKO YOSHIDA: Actually, 4600 is not a road test. So it doesn't give you a test. But it does ask you all the questions if you thought about all this unexpected things. And they wouldn't accept your answer, Oh yeah, we thought about it. No, that's not good enough. You have to prove what you did with evidence. Evidence backed by rigorous engineering. That's what he's asking. That's what UL 4600 is asking. So it's part of the process that, in designing, what you need to think about in advance.
JUNKO YOSHIDA: 实际上，4600不是路考。 因此，它不会给你测试。 但是，它的确会问你所有问题，问你是否预想到了所有这些意外的场景。 而且他们不只单单接受答案，“是的，我们考虑了。” 不，那还不够。 你必须证明自己确实做了这些工作。要有严谨的工程支持作为证据。 这就是Koopman的要求。 这就是UL 4600的要求。 所以，这是设计过程中，你得事先考虑的部分。

For example, they don't tell you whether you should use lidar, you shouldn't use this sensing technology or you should use how many cameras. They don't specify any of those things. They don't care.
例如，他们没有告诉你是否应该使用激光雷达，没有说你不应该使用这种传感技术或是应该使用多少个摄像头。他们没有指定任何这些东西。他们在乎的不是这些。

BRIAN SANTO: As long as there are results.
BRIAN SANTO: 只要结果。

JUNKO YOSHIDA: Exactly. And you need to prove, we have done this much testing and this testing shows, gives you the evidence, this is reasonably safe by using this sensor technology and that sensor technology, a combination of the three or whatever, right?
JUNKO YOSHIDA: 确实如此。你需要证明，我们已经进行了这么多的测试，并且这些测试的结果向你提供了证据，证明通过使用这项感器技术和那项传感器技术（将这三者或其他因素结合使用）是相当安全的，对吗？

BRIAN SANTO: So do we have anything that's like analogous to a Department of Motor Vehicles who's going to take a look at this, whether it's a Beemer or Volkswagen or a Ford or a Prion or whatever and says, Okay, go parallel park here. Now take it on the highway. Do we have a Department of Motor Vehicles yet?
BRIAN SANTO: 那么，我们有没有类似于机动车辆管理所(DMV)的人员来研究这个问题，无论是Beemer、Volkswagen还是Ford、Prion或者其他品牌。然后说，好的，在这里并行泊车，现在上高速公路。 我们有DMV（管理AV）吗？

JUNKO YOSHIDA: No. I think we're far from it. We're really far from it. Unfortunately. I mean, we have a long way to go, right?
JUNKO YOSHIDA: 不，我认为我们差得远。 真的还差得远。 不幸的是，我们还有很长的路要走，对吧？

BRIAN SANTO: Well, next thing to do, right?
BRIAN SANTO: 好吧，这是接下来要做的事，对吧？

JUNKO YOSHIDA: Yup. It's kind of a downer, but it's a lot of challenge. And I think that's why people are excited about autonomous vehicles in one way or the other because there's so much to do, right? And I just love the idea of engineers asking themselves if they thought enough about the safety.
JUNKO YOSHIDA: 对。挑战还很多，这有点令人沮丧。 但我认为这就是为什么人们对AV感兴趣的原因之一，因为还有很多事情需要去完成，对吗？ 工程师会问自己，是否对安全性考虑的足够多，这样的想法是我喜欢的。

BRIAN SANTO: So are you a pretty good driver?
BRIAN SANTO: 那你一定是一个好司机？

JUNKO YOSHIDA: No, I’m terrible! (Laughs)
JUNKO YOSHIDA: 不，我开车技术很烂！（笑声）

BRIAN SANTO: Junko’s blog on the subject, “When your Teenage Robot Can Drive,” is on the web site.
BRIAN SANTO: Junko在EE Times Blog上发表了题为“When your Teenage Robot Can Drive”的博客文章。