本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
你开车或者坐车的时候，我们会想，我们打造的其实是一个司机。
When you're driving around or being driven around, say, you know, we think about what we're building as a driver.
我能想象，构建一个庞大的模型，它能理解物理世界如何运作，懂开车的核心要义——包括社交层面，以及什么叫好司机、什么叫差司机。
I can imagine building a big model that understands how the physical world works and understands the important properties of what it means to drive, the social aspects of driving, and what it means to be a good driver as opposed to a bad one.
我觉得我们已经彻底走出了科研和底层技术开发的阶段，进入了加速全球推广和落地的新阶段。
I would say that we've clearly moved past the stage of scientific research and kind of deep core technology development to this new phase of accelerated global scaling and deployment.
Waymo
Waymo
现在每周在多个城市已经能完成近五十万次完全无人驾驶的载客服务，这标志着从长期研究转向了真实世界的规模化应用。
is now doing nearly half a million fully autonomous rides a week across multiple cities, a shift from long term research to real world scale.
本期节目原载于《Cheeky Pint》播客，Waymo联合CEO德米特里·多尔戈夫与约翰·科利森一起拆解了他们是如何打造这套系统的。
In this episode, originally aired on the Cheeky Pint podcast, Waymo co CEO Dmitri Dolgov joins John Collison to break down how they built the system behind it.
从传感器配置，为什么激光雷达依然关键，到模拟训练和评判模型在AI学习中的作用。
From the sensor stack and why LiDAR still matters, to the role of simulation and critic models in training the AI.
他们还聊到，为什么辅助驾驶不会自然演变成完全自动驾驶，全球规模化需要什么条件，以及产品本身正从定制车辆转向全新的网约车经济模式。
They also get into why driver assist won't naturally evolve into full autonomy, what it takes to scale globally, and how the product itself is changing from custom built vehicles to entirely new economies of ride hailing.
德米特里·多尔戈夫是Waymo的联席CEO。
Dmitri Dolgov is co CEO of Waymo.
他2009年作为首批工程师加入谷歌的自动驾驶项目，之后一路晋升，直到2021年接手整个团队。
He joined Google's self driving car project in 2009 as one of its first engineers and was repeatedly promoted until he took it over in 2021.
Waymo是谷歌最成功的‘月亮计划’，现在每周提供超过五十万次完全自动驾驶服务。
Waymo is Google's most successful moonshot and now provides over 500,000 fully autonomous rides each week.
干杯啊。
Cheers, by the way.
是的。
Yeah.
干杯。
Cheers.
你是在俄罗斯长大的吧。
You grew up in Russia.
对吧？
Right?
我在俄罗斯长大。
I grew up in Russia.
对。
Yep.
那时候我那儿还叫苏联。
Then I I was actually Soviet Union.
是啊。
Right.
没错。
Right.
就是这样。
Exactly.
我爸爸是物理学家。
It's My dad is a physicist.
是的。
Yeah.
所以苏联开始解体的时候，他正好拿到了京都大学的一个访问职位，嗯。
So the Soviet Union started falling apart, and then, you know, he, got had a position a visiting position in university, in Kyoto University Mhmm.
待了一年。
For a year.
我们全家就搬过去了，后来他去了伯克利，我也就跟过去了。
We moved there as a family, and then he went to Berkeley, and then kinda tagged along.
然后我高中毕业了，就有点不知道接下来该干嘛。
And then I ran out of, you know I graduated from high school.
是的。
Yep.
我当时在想下一步该做什么，特别怀念俄罗斯那所技术学校。
I was thinking about the next thing I wanted to do, and I really liked that, that that, technical school in Russia.
俄罗斯的
The Russians
对物理特别认真。
are serious about the physics.
确实如此。
They are.
他们确实很厉害。
They are.
所以我就回俄罗斯了，在那儿读完了本科和硕士
So I went back to Russia, and I got my bachelor's and master's
你回俄罗斯是哪一年的事？
in What year was this that you went back to Russia?
1994年。
1994.
好的。
Okay.
所以某种程度上，那会儿算是俄罗斯人最乐观的巅峰时期了
So that was kind of almost peak Russian optimism in a sense where
确实是的
It was.
那时候正在开放。
It was opening up.
是啊。
It was.
对。
It was.
没错。
Yeah.
确实。
Yeah.
不是的。
No.
我还能想起当时跟妈妈聊过这事。
I actually remember talking to my mom about it.
我爸妈都是在苏联时期长大的。
And, you know, of course, my parents grew up in the Soviet Union.
他们亲眼见过。
They've seen it.
对啊。
Yep.
你懂的吧？
You know?
我是说，他们出生在那之前不久。
I mean, they were born right before Yeah.
战争爆发后，他们经历了特别艰难的岁月。
The war, and then they saw you know, they lived through some really tough times.
我记得跟妈妈聊过这事，其实我回苏联前就在这儿办好了绿卡，是她非要我办的。
And I remember talking to my mom and saying, she she you know, in fact, I got my green card here in The US before I went back, and she insisted that I do it.
是的。
Mhmm.
那时候我根本没打算再回来。
And, I was actually, at the time, I wasn't thinking of coming back.
但那时候我对俄罗斯的发展方向特别兴奋。
But then I was pretty excited about where Russia is and trajectory it's on.
当时我年轻，不懂事，就觉得这路肯定回不去了。
And, you know, being nine young and naive, I was like, there's no there's no turning back.
那你后来怎么又决定回来了？
And so why did decide to come back?
是因为有更大的计划吗？
There's more of a play by
是啊，是有个计划。
play Yeah.
比那个还大。
Than Yeah.
不是的。
No.
是学校的事，这点我很清楚。
School, it was pretty clear to me.
我是想继续学习，没错。
Like, I wanted to continue, studying Mhmm.
就是数学和计算机科学。
You know, math and computer science.
我本科和硕士读的是物理和应用数学，那会儿我也觉得挺对的。
And while the undergrad and master's that I got in physics and applied math, that I think was still Yeah.
俄罗斯的数学和科学教育，基础真的特别扎实。
An incredibly strong kind of foundational, you know, school of Russian math and science.
读研的时候，我特别清楚。
Graduate school Yes.
最好的方式就是去美国。
It was very clear to me that the best way to do it was in The US.
所以我回来了。
So I came back.
我挺惊讶的，英国两家最有价值的初创公司创始人，都是俄罗斯的数学极客，还都上过同一所学校——Revolut的尼古拉和XTX的亚历克斯。
I'm struck by the founders of the two most valuable UK companies are Russian math nerds who both went to the the same school, Nikolay at Revolut and Alex Gurko Alex Gurko at x t x.
不过，这确实是个很强的群体。
But, yeah, it's a it's a strong
侨民群体。
diaspora.
离这儿不远就有一家公司，它的创始人也有类似的背景。
There's a company not far from here where one of the founders also has, you know, a similar pedigree.
对。
Mhmm.
我们公司也是。
We're, a company that Exactly.
也跟这个有关。
Also related to.
没错。
Exactly.
你懂的，那种经典的工程师面试题——你输入 google.com 然后按回车，后面会发生什么？
You know, the classic engineering interview question of, you know, what happens when I type google.com and hit enter
是的。
Yeah.
你知道的，给我讲讲吧，随便哪个部分都行，比如HTTP、DNS这些。
As you know, talk me through, you know, whatever you like, you know, HTTP and DNS and Yeah.
还有BG。
You know, BG.
你可以深入到任何一层技术栈。
You can go down to whatever level of stack you want.
你能不能简单说说，我今天坐Waymo时，技术层面到底发生了什么？
Do wanna maybe just describe when I take a ride in a Waymo today, what's happening at a technical level?
比如，它的架构是怎样的？
Like, what is the architecture?
我先回答你关于实时发生了什么的问题，但这只是故事的一部分，因为我们主要聊的是实时推理这一块。
Let me answer your question of what's happening in real time, but this is going to be only a part of the story because we're gonna be talking about kind of the inference the real time inference part of it.
如果我们想深入展开技术讨论，我觉得更有意思的是把视野放宽，聊聊构建、评估和部署Waymo自动驾驶系统整个生态体系。
And if we wanna, you know, have a deeper, richer technical conversation, it I think it would be interesting also to zoom out and talk about kind of the entire ecosystem of what goes into building, evaluating, and deploying a Waymo driver.
但当你开车或者坐车的时候，总会想到我们做的这个自动驾驶系统。
But when you're driving around or being driven around, say, you always think about what we're building as a driver.
当然，它不是一辆普通的车。
Obviously, it's not a car.
车身上装了一堆传感器，分布在四周。
So it has a number of sensors that are positioned around the vehicle.
我们主要用三种感知方式。
We use three different sensing modalities.
有摄像头，有激光雷达，还有雷达。
There's cameras, there's LiDAR's or lasers Yep.
还有雷达。
And there are radars.
这些是最主要的。
You know, those are the primary ones.
另外还有一些定向麦克风，不过都是辅助的。
There are, you know, some microphones, directional, you know Yep.
麦克风阵列也有，但感知外界主要靠这三种传感器。
Microphone arrays, but those are the primary three for sensing the world.
它们各自的物理特性正好互相补充。
They all have very nicely complementary physical properties.
都能360度无死角覆盖整车，所以Waymo司机随时都能看到周围全貌。
They all have 360 degree coverage around the vehicle, so the Waymo driver sees kind of 360, you know, all the time.
所有这些数据都会传到电脑里，这你肯定能想到。
So all of that data goes into a computer, you would expect.
然后软件处理这些数据，现在全是AI在干活。
And there the software that process now it's, you know, all AI.
对。
Mhmm.
我说的是专门应对现实物理世界的AI。
Like, I'm saying specialized AI in the physical world.
它会处理传感器传来的所有数据。
So it processes the sensor data.
现在呢，用AI的说法来说，就是那些编码器把传感器数据收进来。
Nowadays, you know, talk about it in the, you know, using AI terminology as, you know, encoders that, you know, take this data, in.
然后呢，还有个解码器，也就是执行部分，你可以理解成车里的生成模块。
And then there's the kind of the decoder, the action, you know, the generative part, if you will, in the car.
这个生成任务呢，就是要学会怎么开车。
And the generative task there is to, you know, figure out how to drive.
对吧？
Right?
而这个系统呢，会通过一个专门的接口连到车上，控制车辆的各个动作。
And that is, of course, connected, through kind of a specialized interface to the car where we can actuate the vehicle.
所以你才会看到方向盘自己转，车自己带着你跑。
And, you know, that's why you see the steering wheel, you know, turn, and it drives you around.
明白了。
Okay.
我一上车，
So I get into my car.
传感器主要分三类：激光雷达、雷达和摄像头。
There's three main families of sensors, LiDAR, radar, and, cameras.
然后系统用这些数据先构建对周围世界的模型，比如其他车都在哪儿，之类的。
And then it is using that to first build a model of what's going on in the world, you know, where are all the other cars and things like that.
就像你说的，再根据这些信息做决策，最后控制车子行动。
And then as you say, make decisions and then actuate that with the car.
你现在的驾驶体验，就是这套系统在运作。
That is the the system that you're living in.
所有这些推理都是在车里本地完成的吧？云端根本不参与？
And is all that inference done locally, presumably, Nothing's in the cloud?
实时处理全在车上，不靠云端。
Nothing real time
不在云端。
in cloud.
有些事其实也能在云端做，但不是必须的。
And there are some things that can happen, you know, in the cloud, but they're not required.
明白了。
Got it.
那在云端做的一些锦上添花的功能，能举个例子吗？
What what's an example of a nice to have that happens in the cloud?
你可以想象这种情况，有些事情其实跟自动驾驶测试本身没直接关系。
You can imagine a situation where, we do you know, some of it is not directly related to the test code driving.
对。
Mhmm.
比如你下车之后，我们想确认一下车有没有弄脏。
But, let's say, after you leave the car, we wanna check that, you know, the car is not dirty.
没错。
Yep.
你有没有落下东西，也得检查一下。
You didn't leave anything Yes.
就在那儿。
There.
如果你真落下东西了，嗯。
If you did leave, you know, an item Yeah.
要是你走的时候弄得一团糟，对吧。
Well, if you, you know, left in a a mess Yep.
那我就得把车送回我们的维修点，好好打扫一下。
Then, you know, I wanna send the car to one of our depots, get it cleaned up.
要是你把手机落在车里了，我们得能检测出来，没错。
If you left an item there, you know, on your phone, alright, we want to, detect that Yep.
然后通知你，把东西寄回给你。
And then, you know, send it to our listing phone and let you know.
对吧？
Right?
所以我们用的是一个不在车上的模型来处理这事，而不是非得装在车上，毕竟这跟开车没关系，不是实时任务。
So that, you know, we do with the kind of a by asking a model that actually lives off board as opposed to having to put it on the car, right, because it's not a real time task related to, you know, the driving.
这就算一个例子了，对。
So that's one example of Yeah.
有一件事
Something that
推特上老有人在争论自动驾驶的各种问题。
There are all these debates that go on on Twitter around self driving.
比如端到端方案和那种模块化方案的分歧。
So I can think of, you know, end to end versus the more kind of modular, approach.
还有只用摄像头，还是用一堆传感器的争论。
There's, cameras only versus array of sensors.
我不太清楚，这些争论对行业专家来说真有那么重要吗？还是说，其实早有定论，只是成了算法训练的素材？
And I can't tell, are these debates actually interesting to an expert in the field, or do you think these are just settled matters and they're just grist for the algorithm?
我明白这些问题的来由。
I understand where the questions are coming from.
但我觉得，这些讨论往往太笼统，把很多真正关键的细节都丢掉了——在我看来，最有意思的技术问题，恰恰藏在这些细节里。
I do find that, kind of, often, the way they're posed and the way the debate happens is losing a lot of the nuance and a lot of detail that really matters are, to me, the most interesting technical questions are in that level.
因为我们打造Waymo自动驾驶系统的方式，是从一个强大的云端基础模型开始的。
Because the way we think about the in building the Waymo driver, it starts with a large off board foundation model.
对。
Mhmm.
我能想象构建一个大型模型，它能理解物理世界如何运作，明白驾驶的核心要素、社交层面，以及什么是好司机、什么是差司机。
I can imagine building a big model that understands how the physical world works and understands, the the important properties of what it means to drive, the social aspects of driving, and what it means to be, you know, a good driver as opposed to a bad one.
这就是基础。
So that's the foundation.
然后我们把它细化成三个主要的离线教师模型。
Then we specialize it into, what we call, know, three main off board teachers.
这些依然是大型、高容量的离线模型。
There are still large, high capacity off board models.
分别是Waymo驾驶员、模拟器，还有评判系统。
There's the Waymo driver, there is the simulator, and then there's the critic.
对吧？
Right?
之后再把这些模型压缩成更小的版本，这样就能更快地运行推理了。
And those then get distilled into smaller models that you can run, you know, inference on faster.
所以Waymo驾驶员就成了
So the Waymo Driver becomes,
你知道的
you know,
核心，主心骨，没错。
the backbone, the male backbone Yep.
车里头的那个核心。
Of what's, in the car.
模拟器呢，负责生成我们云端的合成环境，用来训练和闭环评估系统；至于评判者，
The simulator, of course, is what powers our synthetic generative environment that can run on the cloud for training and for evaluation in closed loop of the system, and the critic So
激光雷达有在本地运行吗？
does XALiDAR ever run locally?
没有。
No.
完全没有。
No.
它不运行在本地。
It doesn't.
对。
Yeah.
不过我觉得有意思的是，解码器的工作方式、模型的运作机制——如果你想想模拟器里那个生成真实世界、模拟其他人行为、车辆动向的任务，其实都离不开这一点。
However, what I think is interesting, in a way, the way the decoder works, the way the model works, if you think about the generative task in the simulator of kind of, you know, creating those realistic worlds and how, you know, other people behave, how, you know, cars Yeah.
行人，没错。
Pedestrians Yeah.
还有骑车的人，这些对象之间如何相互关联，是车上实时任务必须解决的核心能力，而这一点，其实和模拟器里要做的事是相通的。
Cyclists, in order and the task that you have to solve on the car in real time, there is this ta you know, fundamental shared capability of understanding how these objects relate to each other Mhmm.
如果你在车上运行系统，就得预测它们未来可能的行为；而在模拟器里，你则通过采样生成这些概率性的行为模式。
And predicting what they might do in the future if you are running on the car and then generating those, you know, some sampling, those probabilistic behaviors in the simulator.
所以虽然用的是不同的模型，但正因如此，共享的基础模型才能同时支撑这两边。
So it's it's different models, but there is you know, this is why the shared foundation model is able to power both.
明白了。同样，说到评判模块，它的任务就是找出那些关键事件，然后给出明确的判断，没错。
Got And similarly, if you think about the critic, like, the job of the critic is to find interesting events and then, you know, be opinionated Yes.
关于什么是好的行为，对。
About what's good behavior Yes.
还有什么是，对。
And what's Yes.
坏的行为。
Bad behavior.
本质上都是同样的理解。
Similar fundamental understanding.
对吧？
Right?
如果你在车上做推理，还是得从这些可能的未来场景里，挑一个你想要去实现的，对。
If you're running, you know, inference on the car, you still have to, like, figure out which of the multiple hypotheses of these future worlds you wanna, you know Yes.
然后据此做出决策。
Take action to steer towards.
是的。
Yeah.
对吧？
Right?
好的。
Okay.
这些都基于同一个基础模型吗？
And these are all downstream of the same foundation model?
没错。
That's right.
先从基础模型开始。
So start with the foundation model.
对的。
Yep.
然后你再做微调，让它专精于某个任务，但还是用原来的模型。
You know, then you, you know, specialize in fine tune, still off board model.
这些就是老师模型，然后你把每个老师模型的知识都提炼出来。
Those are the teachers, and then you distill each one of the teachers kind of distill Yes.
训练它自己的学生，没错。
Trains its own student Yes.
还有驾驶员、模拟器、评判者。
And the driver, the simulator, the critic.
对。
Yes.
你二十年前就开始做自动驾驶了。
You started working on self driving twenty years ago.
是啊。
Yeah.
从技术演进的角度看，这只是一个算力堆砌的故事吗？就是说，我们得有足够的计算资源才行？
As you think about the tech evolution, is this just a scaling law story where we had to be able to throw enough compute at us?
还是说，我们得等某些架构创新出现才行？
Were there architectural approaches we needed to wait to, have be invented?
还是说，我们其实走了二十年的弯路，才终于找到正确方向？
Was it just a story of we needed twenty years of going down the wrong cul de sacs before we eventually arrived at the right approach?
以你现在的认知来看，2015年你能成功推出一个像Waymo这样的产品吗？还是说当时还缺些关键技术？
You know, could you knowing what you know now, could you have a successful Waymo in market in 2015, or was there some enabling technology?
不行。
No.
这些年技术上的突破至关重要，尤其是人工智能方面。
Technology, breakthroughs that happened over the years were critically important, primarily in AI Mhmm.
还有其他领域，比如计算能力。
But also in other areas, like, you know, compute.
你得有强大的算力才能跑得动。
You got a heavy compute that you need run.
对。
Yep.
我不觉得这是试了一千条死胡同，然后回头改道，最后才找到唯一正确路径的故事。
Now, I wouldn't characterize it as, like, going, you know, a thousand different dead ends and then having to retract and then finding, like, the one right path.
我觉得这更像一个不断迭代、逐步演进的过程。
I would characterize it as, yeah, iterative learning and evolution.
是的。
Yes.
后来，Transformer 模型就出现了。
And then, you know, transformers came around.
但你知道，像 Transformer 这种架构其实非常通用。
But, you know, transformers, for example, are very general Mhmm.
架构。
Architecture.
对吧？
Right?
没错。
Yep.
支撑了大语言模型的发展，没错。
Powers, LLMs, powers, you know Yes.
我们的模型。
Our models.
但你怎么把这些用到那个领域里，我觉得这才是关键，不是随便就能搞定的
But how you apply them to that space, I think this is where Isn't just fallout
是那些做转录的人带来的影响。
of transcribers.
没错。
Exactly.
对吧？
Right?
当然了，大家总爱聊架构什么的。
And, of course, you know, people like to talk about architectures.
是啊。
Yep.
架构当然重要，但说到底，真正关键的是你的评估指标、评测方法、训练技巧，还有新数据。
But architecture, you know, is important, but, really, a lot of it comes down primarily to your metrics, to your evaluation mechanisms, to, you know, all of the training recipes, and, of course, new data.
对。
Yes.
大语言模型擅长处理文本，或者说，其实是处理词元。
LMs are good at text or, I mean, tokens Mhmm.
特别是，它们在那些有单一文本语料库的领域表现最好，比如编程——因为本来就全是文字，特别方便。
Specifically, and obviously perform best at domains that have some kind of single corpus of text they can work on, like coding, where it's very helpful that everything was just kind of textual already.
成功的一部分，就在于把各种领域都转化成文本形式，这样就能用大语言模型来处理了。
And part of the success has been creating textual representations for domains so that we can then, you know, put LMs against them.
你能说说，你是怎么编码你所看到的世界的吗？
Can you describe how you encode the world that you're seeing?
是的。
Yeah.
你的做法是直接建一个三维模型，相当于一个三维位图吗？
I mean, are you just building a three d model, like a three d bitmap, essentially?
还是
Or
这其实就引出了一个问题：编码器和解码器之间的接口到底该是什么样子。
So this is where I think we can get a bit into the this question of what is the interface between the encoder and the decoder parts.
对。
Yep.
而且这其实也涉及到你之前提到的那个问题——大家总爱争论要不要端到端。咱们先聊聊端到端到底什么意思，然后再回头说说编码器和解码器之间的接口到底该咋设计。
And I think that touches also on the, you know, the thing you flagged earlier where people like to, you know, debate end to end or not end to so the of the way, let's make you know, talk a little bit about end to end and then get back to, like, what is the interface between those two.
对吧？
Right?
那咱们说的‘端到端’，到底指的是啥？
So when we say end to end, what do we mean?
意思就是，它是个很大的机器学习模型。
We mean that, it is some large ML model.
不过通常呢，大家不会把它一股脑儿全堆成一个整体。
Typically, you don't build them monolithically.
而是会拆成好几个部分，不同的子图结构。
You have, you know, different parts and different subgraphs.
但关键在于，你得能让梯度从前到后、再从后往前顺畅地传过去。
But what's important is that you can propagate and backprop the Yep.
梯度和损失函数能够沿着所有层级反向传播，这样每一层都能学习到对最终任务有意义的权重和特征表示。
You know, gradient and the the the loss function all through the different layers so they can you know, every layer, you can learn, you know, the the weights and the representations that matter for the the final task.
你不用把数据硬塞进编码器和解码器之间那个狭窄的过渡环节里。
You don't force it through some, you know, narrow funnel between, let's say, the encoder and the decoder.
对
Yeah.
我对端到端的理解很简单，就是输入像素数据，直接输出自动驾驶的操控指令，就是说
I think of a simple view of end to end being, you know, pixels go in and car actions come out, which
可能有点
may be a
过于简化了
bit of an oversimplification.
不过
But
对
Yeah.
说得太对了。
That's exactly right.
这其实就是最基础、最普通的版本了。
And if like, this is kind of the basic vanilla version of it.
对吧？
Right?
你想啊，要打造一个能完全自动驾驶的司机系统，得需要些什么？
There if you think about the you know, what will it take to build the driver that's capable of fully autonomous operations.
没错。
Yep.
你得考虑整个生态系统——司机、模拟器、评判系统，全都得有。
You think about this entire ecosystem of the driver, the simulator, the critic.
没错。
Yep.
如果你只做像素输入、轨迹输出这一套，想同时搞定这三块，还达到我们要求的安全性和性能，根本不可能，更别提大规模应用了。
If that's all you do, like, you know, pixels in, trajectories out, it becomes very difficult to do all of those three and achieve the high level of safety and performance that we require, and it becomes very difficult to kind of do it at scale.
不过呢，这确实是入门最简单的方式。
And, however, if you know, that's, it it's kind of a very easy way to get started.
对吧？
Right?
你只要收集点数据就行，就像现在大模型圈里大家都做的那样。
You collect some data, kinda like, know, in all of you to the LLM world.
对吧？
Right?
最简单的做法就是挑一个现成的模型。
The easiest thing you can do is have you know, you know, pick a model.
没错。
Yeah.
现在入门最方便的办法，就是直接用一个视觉语言模型。
The easiest way to get started nowadays would be just, you know, take a VLM.
它本来就自带了和语言对齐的摄像头编码器。
It already has a, kind of a language aligned camera encoder.
对。
Yep.
然后它有个解码器，能根据图像生成文字，你知道的。
And then it has a decoder that, you know, will can predict you know, generate text by you know?
你可以微调它，比如说：
And you can fine tune it and say, hey.
别生成文字了，生成运动轨迹吧。
Instead of text, generate trajectories.
这真的很容易做到。
You know, very, very doable.
其实我们之前发过一篇论文，
In fact, we you know, a while ago, we published a paper
对。
Yes.
叫Emma，就是干这事儿的。
Called Emma that did exactly that.
是的。
Yes.
而且在正常情况下，它开车开得相当不错。
And it will actually, I mean, in a nominal case, drive pretty darn well Mhmm.
这简直让人震惊。
Which is mind blowingly impressive.
这太搞笑了。
That is very funny.
对。
Yeah.
你的意思是，
And, I mean, there's something to You're
你可以说，
saying you
你可以直接拿一个现成的模型，
can take an off the shelf model,
这跟开车根本没关系。
which has nothing to do with, driving to start with.
没错。
That's right.
你还能得到不错的效果。
And you'll get these good results.
没错。
That's right.
是的。
Yeah.
明白了。
Get it.
在正常情况下。
In in a nominal case.
是的。
Yeah.
我明白了。
I get it.
我只是想说清楚一点。
Just wanna be clear.
这跟你实际需要的差距可不是一星半点。
It's orders of magnitude away from what you need.
对。
Yeah.
是啊。
Yeah.
别在街上试，但确实管用。
You should not try it on the street, but it works.
比如说吧，
But for example,
要是真谈马的话。
if talking horse.
它能说话，这真让人惊叹。
It's impressive that it's talking.
你懂的。
You know?
没错。
Exactly.
就是这个意思。
Exactly.
其实你真正想做的产品，可能根本不是全自动驾驶，而只是一个驾驶辅助系统，对吧。
And you can actually the product that you wanted to build was maybe a driver assist system, not a fully autonomous system Yep.
那可能你只需要做到这一步就够了。
Then maybe that's all you need to do.
对。
Yep.
而且说实话，做这种程度的系统，根本用不着那么复杂的模拟器和评估机制，因为对可靠性的要求低多了。
And and then, yeah, for that, you don't need all this other machinery of the simulator and the critic, so that that's because the number of nines is drastically lower.
但这其实挺有意思的，因为背后确实有那么点直觉，能解释为什么这种方法有效。
But there is this is interesting because there, you know, there is some intuition behind, you know, why that works.
你想想开车最难的部分，其实跟聊天挺像的。
If you think about the hard parts of driving, it's, you know, not unlike, you know, having a conversation.
对。
Mhmm.
只不过在大模型的世界里，你是在建模语言，或者说是建模句子和词语之间的对话。
Except, yeah, if in the LLM world, right, having you know, you're modeling language or maybe modeling a dialogue in the space of sentences and words.
而开车难的地方，也在于这种多主体、社交互动的特性。
What makes driving hard is also this kind of multi agent social interactive part of it.
对吧？
Right?
我一动，就会影响你。
If I do something, that's gonna affect you.
也会波及别人。
It's gonna affect somebody else.
而且历史背景挺重要的。
It's kinda and the history matters.
这不是个局部的、光靠几何就能解决的问题。
It's not local and just geometric.
上下文很关键。
Context matters.
语义也很重要。
Semantics matters.
所以，但它用的不是语言那种词句的方式。
So, but it's in a different you know, it's not in the language of words.
而是用一种身体语言之类的东西来表达。
It's in the language of kind of well, body language, if you will.
对吧？
Right?
为什么会这样？如果你用这种方法，其实已经有实证支持了。
How so and we see that empirically validated if you, you know, do this approach.
好的。
Okay.
那我们就假设把这个东西做出来。
So then let's say we build this thing.
就用摄像头，像素输入，输出轨迹。
Just cameras camera encoder, pixels go in, trajectory go out.
这质量在正常情况下开车是够用的。
It the quality is sufficient to, you know, drive in the normal case.
但要应付那些罕见的极端情况，达到我们要求的超人级安全标准，还远远不够。
It's not sufficient to deal with the long tail of, you know, the edge cases and hit the high bar of superhuman safety that we require.
那接下来你就会问了：还需要点啥？
So then, you start asking the question, what what else do you need?
对。
Yes.
如果你训练的时候，光是观察别人是怎么开车的，嗯。
And, if all you did was kind of observing how other people drive when you trained Yep.
展开剩余字幕（还有 480 条）
系统可能只是被动观察人们怎么开车、怎么互动，也可能让你自己开车，再用模仿学习来训练它，是的。
The system, maybe observing, you know, just passively how people drive and how they interact, maybe also, you know, driving the car yourself and then using imitative learning Yes.
但别忘了，光靠这个还不够。
To train it, Mind that that's not enough.
你得搞闭环训练。
You have to do something in closed loop.
你得像RLFT那样做点什么，就是那种跟我们看到的并行的方法，对吧？
Kinda have to you know, you have to do things like RLFT, which is also, you know, parallel to what we see also in RLFT?
RLFT。
RLFT.
基于强化学习的微调。
Reinforcement learning based fine tuning.
哦，明白了。
Oh, okay.
对。
Yeah.
是的
Yeah.
就跟那种结合人类反馈的强化学习差不多。
So, you know, similar to the reinforcement learning with human feedback Yeah.
对
Yeah.
在大语言模型领域里。
In in the LLM world.
对吧？
Right?
没错
Yeah.
咱们得搞个真正的闭环驾驶系统，就是让车去尝试各种不同路况，再给它奖励信号，好让它别跑偏。
Wanna do maybe closed loop proper closed loop driving where, you know, you explore all kinds of different situations, and then you give it a reward signal to kind of keep it in distribution.
为此，我们需要一个高度真实的模拟器。
For that, then we need a realistic simulator.
对吧？
Right?
如果你想搞个靠谱的强化学习系统，就得对奖励函数有自己的判断。
You also you know, if you wanna have a good RL system, you need to have an opinion for the reward function.
这时候就轮到评论家出场了。
This is where the critic comes in.
对吧？
Right?
如果你搞的是纯端到端的系统，那咱们来看看这个模拟器。
If you have a purely end to end system, let's look at the simulator.
那接下来你怎么办？
Now what do you do?
你就只能从像素直接推到轨迹了。
You have to you're you're then constrained to just go from pixels to trajectories.
对吧？
Right?
就这么简单，你只要把系统跑起来就行。
That's that's all, you know, you you can run And the system on.
这空间维度太高了，想生成所有东西根本是个难题。
It it's a very high dimensional space, it's a it's a, you know, hard problem generate everything.
就算你解决了这个问题，用像素直接到轨迹再模拟整个流程来训练或评估，效率也低得吓人。
But even if you solve that, it just becomes incredibly inefficient to run it in the in the full way of pixels to trajectories and simulation for training or for evaluation.
这时候就需要中间表示了。
So this is when intermediate representation come in.
在这个任务里，其实已经有一些中间表示在用了。
There are some intermediate representations in the world in this task group.
在现实世界里，我们知道哪些是靠谱的。
You know, in the physical world, we know are correct.
是的。
Yes.
但光靠这些还不够。
They are not sufficient.
是的。
Yes.
但它们并不会限制通用性。
But they're not generality limiting.
对吧？
Right?
你看，这儿有个物体。
You know, there's an object here.
还有，比如说，道路这个概念。
There's, you know, a concept of a road.
有标识牌。
There's signs.
有速度限制。
There's speed limits.
所以，这就是为什么要在此基础上做增强。
So this is where augmenting that Yes.
从编码器解码器学到的那些表征，那些嵌入向量，再加上更结构化的表示方式
Learned representation, those learned embeddings from the encoder decoder with that, you know, more, you know, structured representation
是的。
Yes.
这就是我们所做的，我们发现这实际上为我们提供了更多可以调节的选项来模拟。
Is what we do, and we find that this kind of gives us additional knobs to simulate Yes.
在那个场景里，你看，就是从像素到……对。
You know, in in that space, just, you know, pixels to Yep.
轨迹。
Trajectories.
这样我们就能进行额外的安全验证了，没错。
It allows us to have additional safety validation Yep.
实时叠加这些层级，同时DIP还为我们提供了额外机制来设定奖励函数，你知道，就是用来评估批评者或者...所以你看，这又绕回来了，我们差不多完成了一个完整的循环。
Layers in real time, and it also allows us you know, a DIP gives us additional mechanism to specify the reward function, you know, for evaluation of critic or, you know, for So this is, again, like, we've gone kinda full circle of it.
这是意图吗？
Is it intent?
是的。
Yes.
没错。
It is.
对。
Yes.
如果你想大规模实现完全自动驾驶，就得加上这些额外的东西。
Then if you wanna do it at scale for full autonomy, it's augmented with all of this other stuff.
关于模拟这一点，真的很有趣。
That's very interesting on the simulating point.
对端到端模型来说，模拟实在太难了，因为处理中间表示比还原一个像素级的完整世界要容易得多。
It's just very hard to simulate for an end to end model because it's easier to deal in end to end or it's easier to deal in intermediate representations rather than coming up with a pixel perfect view of the world.
两者都需要。
You need both.
没错。
Yeah.
对。
Yeah.
所以呢，有个经过增强的天线架构，确实很有用。
So, you know, having an antenna architecture that's augmented Yes.
有了这种结构，你就能在两种模式之间自由切换。
With that structure allows you to kind of play in both of those worlds.
没错。
Yeah.
是啊。
Yeah.
当然。
Yeah.
你做自动驾驶汽车，最终想达成什么目标？
What are you looking to do as a self driving car?
听起来有点怪，但很多人可能没意识到，你得解决好多不同的问题：既要让人安全到达目的地，还得尽量快地送到——没错。
I mean, it sounds funny, but I think people maybe don't realize that there are many different things that you're looking to solve for, where you're looking to get the person to their destination, you're looking to get them there reasonably promptly Yeah.
还要开得特别平稳
But also drive quite smoothly
对
Mhmm.
还要有多重安全保障
And also have many lines of safety Yeah.
也不能惹其他司机烦，老被按喇叭，这事儿也得考虑
And also not annoy other drivers and get honked at, and you and and and.
是的
Yeah.
那你们在优化哪些奖励函数，或者还有哪些不太为人所知的目标？
So what are some of the reward functions or kind of things you're optimizing for that maybe are not obvious to people?
安全永远是第一位的
So safety is the primary focus.
对吧
Right?
但当然，我们也希望车开得平稳，不管是对车里的人还是其他道路使用者来说。
But, of course, we also want to be a smooth driver so that, you know, for both people in the car and other actors.
没错。
Yep.
我们还希望车子行为可预测、规矩守序，这样才够友好。
And we also want to be, you know, a predictable, well behaved one so that it can, like, nicely Yes.
能很好地融入整个道路系统的社交生态。
Fit into the whole social ecosystem of our our roadways.
看起来，自动驾驶一出来就面临一个问题：人们就是没法好好相处，或者说，不是每个人都对机器人客气。
It seems like one of the issues that has quickly emerged with self driving is the fact that people can't have nice things or, you know, not everyone is nice to the robots.
所以，不管你是开过一个乱糟糟的地段，被别人堵住，还是干脆说‘我不在这儿送你了’。
And so, you know, whether you're, you know, driving through a dodgy area or getting blocked or, you know, maybe I'm not gonna drop you off here.
可能我会绕一圈，找个人少点、安全点的地方再把你放下。
Maybe I'm gonna go around the block and, you know, drop you somewhere better.
但正如你所说，这些全是人类带来的问题，你们打算怎么解决呢？
But all of these, as you say, kind of other human issues, how do go about solving those?
你提到的这些问题，其实都是我们得好好改进的地方。
A lot of the ones that you mentioned are just things that, you know, we need to work on Yep.
而且说实话，如果我们没在你想要的地方下车，那肯定是我们没做到位。
And understanding, honestly, you know, said that if we're not dropping you off, we're exactly where you want it to be dropped off Yep.
或者，我们没给你一个好用的界面，让你能清楚表达需求——这责任在我们。
Or, you know, we don't give you kind of a good interface Yeah.
这事儿我们得担着。
To tell us, that's on us.
对吧？
Right?
我们就是想把它做得更好。
You just, you know, we wanna make it better.
是的。
Yeah.
对吧？
Right?
我觉得下车这一环节其实是自动驾驶旅程中特别讲究的一部分。
It feels like the drop off is actually a pretty nuanced part of the the self driving journey.
比如高速上还好，还有那些……
Like, the the highway stuff and the Yeah.
时速35英里的城市道路我们都搞定了，但下车体验里头的细节真不少。
You know, the 35 mile an hour roads, that is all nailed, but there's just, a lot of nuance in the drop off experience.
我觉得哪一项都不容易。
I'd say they're all hard.
你选高速和选下车点，出发点本来就不一样。
You pick freeways and you pick drop offs for different reasons.
对吧？
Right?
说到下车，你说得一点没错。
For drop offs, there's you're absolutely right.
确实有些地方，可能没那么显而易见。
There are, you know, a few things that are maybe not obvious.
你好好琢磨一下这个问题就明白了
You you know, you just think about this problem.
核心就是搞清楚你想去哪儿，对的
But it's understanding where you wanna go Yep.
然后尽最大可能给你提供方便
And making it as convenient as possible for you.
而且接送的场景其实不是完全对称的，接乘客和送乘客差别挺大的
And pickups and from dropout is not exactly symmetrical.
对吧？
Right?
但我当时也在理解具体场景——你知道，到底该停在哪里？
But then I was also understanding the context of the situation where you you know, where do you stop?
对。
Yes.
靠边停车。
Pull over.
你可不能堵住人家的车道。
You don't wanna block a driveway.
也不能随便双排停车，你知道的。
You don't wanna, you know, double park.
不过有时候要是只是停一下，可能也没那么严重。
Although in some cases where if it's just a quick one, maybe it's okay.
所以要做好这件事，其实有很多细节要考虑，得让整个体验——
So there's a lot of nuance that goes into doing that well so that it's a Yes.
对骑手来说顺畅、无摩擦，是的。
Smooth less, frictionless experience for the rider Yes.
对其他人也是这样。
As well as other folks.
是的。
Yeah.
没错。
Yeah.
高速公路上，大部分时间都挺平静的
Freeways, for most of the time
是的
Yeah.
没啥特别的事发生，你知道的
They're very you know, not much happens.
对啊
Mhmm.
设计得特别规范，没错
They're very well structured Yep.
因为我们本来就是这么设计的
Because we designed them that way.
但还是有一堆特别复杂的情况会发生，是吧
But there is still that long tail of really complicated stuff that happens Yes.
一旦出点岔子，后果可就严重了
Where the consequences of, you know, a bad event
严重得多。
Much more severe.
是吧？
Are
严重得多。
much more severe.
对吧？
Right?
速度高多了。
Speed is much higher.
所有东西的影响都跟速度的平方成正比。
Everything is, you know, quadratic in speed.
所以那边出的事可多了。
So it but we see a lot of stuff there.
你能想象到护栏从高架桥上砸下来的情景。
And you imagine grills falling off of freeways.
想象一下，人们出车祸，车子失控打转的样子。
Imagine, you know, people getting into, accidents and kind of spinning out of control.
你有没有见过
Do you see one
那种平板卡车，上面堆得满满的货物，你开车跟在后头？
those flatbed trucks with just, like, a bunch of stuff piled in it, and you're driving behind it?
我不确定。
I don't know.
我每次看到都觉得特别提心吊胆。
I always find it very nerve racking.
看着挺
Looks a
是啊，我知道。
bit I know.
对。
Yeah.
是的
Yeah.
对
Yeah.
咱们都见过，那些车开过去一路撒东西。
And we like, we've seen them, you know, leave a trail.
没错
Yes.
确实
Yes.
是啊
Yeah.
行吧
Okay.
所以问题也不一样。
So there's a different set of problems.
但我觉得，大家对Waymo的普遍看法是，自动驾驶的核心技术你们已经基本解决了，现在主要就是扩大规模，还有像极端暴雪天气这种极端边缘情况。
But it it feel I feel like the general sentiment with Waymo is that the, driving has mostly now been solved by you guys, and it's kind of a question of scaling up and maybe some super long tail stuff, really snowy conditions.
这么说吧，这真的是你们内部的看法吗？还是说实际情况比这复杂得多？
Like, is that your sense internally, or is there actually much more nuance to it than that?
我觉得吧，
I would say the, yeah.
听起来好像你们的工程问题都搞定了。
It sounds like, you know, we're done with engineering.
对。
Yeah.
没错。
Yeah.
我觉得我们已经明显从最初的科研和核心技术研发阶段，进入了加速全球部署和落地的新阶段。
I would say that we've clearly moved past the stage of scientific research and kind of deep core technology development to this new phase of, accelerated global scaling and deployment.
是的。
Yes.
所以我们还得继续努力。
So, you know, we still have work to do.
没错。
Yep.
对吧？
Right?
但我觉得，现在核心科技上已经没什么短板了。
But I don't see today any limitations or any gaps in the core technology.
开车的技术现在够好了吗？
The driving is good enough now?
核心科技我觉得已经够用了，实在想不出还有什么驾驶场景是目前技术解决不了的。
Well, the the core technology, I think, is good enough that I can't, you know, think of any, you know, aspect of driving Yep.
都靠这套基础技术撑得住。
That is not supported by the fundamental technology.
不过话说回来，专门化方面要做的事儿还多着呢。
Now that said, there is, you know, a lot of work to do in, you know, specialization Mhmm.
在部署之前，我们还得做充分的验证，这样才能稳妥推进。
And in validation before, you know, we can, you know, deploy responsibly.
对吧？
Right?
我们可不是在全球各地都开车呢。
We're not driving everywhere in the world.
我们计划今年先在伦敦开始运营。
You know, we are planning to start operating in London Mhmm.
还有东京，今年也会上线。
And in Tokyo this year.
你是不是觉得，我们在旧金山用的这套系统，直接搬到伦敦就能用？其实不行。
And, you know, are we just do we have a driver that, you know, you're using today in San Francisco that we can just plop down in London and No.
是的。
Yeah.
对吧？
Right?
但我们看到的是，从核心技术是否成熟的角度来说，情况非常令人鼓舞。
But what we're seeing is incredibly encouraging from the perspective of, like, is the core technology there?
对吧？
Right?
现在关键是收集数据、做些本地化调整和验证，毕竟伦敦和东京的交通标志都不一样。
Now it's a matter of, you know, collecting the data, doing some specialization and validation, and you can your signs are different, you know, in both of those places.
人家开车是靠左行驶，不过说实话，这对电脑来说真不算难。
People drive on the other side of the road, But, you know, that's actually not that hard for computers.
对吧？
Right?
核心技术的适应性其实很强，但还是得做不少本地化工作。
The core technology generalizes really well, but there's still work that you have to do.
最不适应的是什么？
What generalizes least well?
我们越来越发现，尤其是——
Increasingly, we're finding, especially, you know,
现在我们能把Waymo的AI和数字世界的AI、视觉语言模型连起来，直接继承它的通用知识。
now that we're able to kind of hook the Waymo AI to the AI in the digital world, the VLMs, and kinda inherit the Yes.
借助视觉语言模型的通用世界知识。
General world knowledge from VLMs.
因为有了这些通用知识，我们在零样本或小样本学习上取得了非常不错的效果。
We're seeing really strong results from, like, zero shot or, you few shot learning because of that general knowledge that we bring in.
但有些情况，比如严寒天气，会直接影响整个系统。
But there are a few things like, say, cold weather, cold winter weather, where it affects the entire stack.
对吧？
Right?
这不光是AI的问题。
So it's not just, you know, the AI.
我们还得考虑。
We actually have to.
硬件。
Hardware.
是的
Yeah.
你得有硬件，还得有合适的清洁液，里面得加加热元件。
You know, you have need the hardware, you need to have the proper cleaning solution, you know, heating elements in it.
没错
Yes.
再想想那些计算机完全能解决的问题，比如运动控制和湿滑路面。
And then, you know, think about, things that are completely solvable by computers, like motion control and slippery surfaces.
对吧？
Right?
这得花不少功夫。
So that takes a bunch of work.
光靠引入外部知识是解决不了的。
You don't get that for free from just, you know, pulling in
是的
Yeah.
有些是那个，VLM，没错。
Some, you know, v l VLM Yeah.
解码器。
Decoder.
我的印象是，早期在旧金山或者凤凰城这些早期市场，可能有很多本地化的工作，比如地图绘制之类的，但你们似乎是通过通用化解决了这个问题，或者只是提升了处理每个城市特定需求的能力？
Wasn't the case I mean, my impression, not knowing anything, is that in the early days, there was maybe a lot of San Francisco specific work or Phoenix specific work in the early markets, whether it be mapping or something else, and that you guys seem to either have solved that in generalizing it or just scaled up your ability to do the city specific work?
是什么让你们能这么快地扩展到新城市？
What enabled the kind of the rapid city expansion?
我们通常更关注Waymo自动驾驶系统的能力和部署，而不是直接在那个领域里，没错。
We usually think about it in know, the capability of the Waymo driver as well as deployment not primarily and directly, you know, in that space of Yeah.
城市，是的。
You know, cities Yep.
或者邮编。
Or ZIP codes.
想想它的运行范围。
Think about the operating domain.
是的
Mhmm.
对吧？
Right?
然后这就只是
And then that's just the
高速公路和寒冷天气是
Freeways and cold weather are
总是高速公路、寒冷天气、下雪、下雨、大雾、密度这些因素等等。
always Freeways, cold weather, snow, rain, fog, density, etcetera, etcetera.
而我们正在打造的，正是这些东西。
And then that, like, that's what we are building.
我们就是根据这些来评估的，然后对应到具体的城市。
That's where we're evaluating, and then that maps to a city.
比如某个特定城市，没错。
Like, particular city Yep.
是在运营范围内还是超出范围了。
Be within the operating domain or outside of it.
对。
Yep.
对吧？
Right?
所以呢，你看，如果我们回顾一下历史，我们最初是在哪里首次推出完全自动驾驶的商业服务的
So what, where, you know, if we could provide history a little bit, our initial deployment in where we started offering a fully autonomous commercial service for the first time
对的。
Yeah.
这事发生在2020年，没错。
Was in 2020 Yep.
地点是亚利桑那州的钱德勒市。
In Chandler, Arizona.
那次测试用的就是我们所说的第四代Waymo自动驾驶系统。
So and that was on, what we call the fourth generation of the Waymo Driver.
这个你们还记得吧，就是那批太平洋小型巴士，硬件和软件都不一样。
This was the if you remember, the Pacifica minivans with, you know, different hardware, different software.
那时候我们一心一意要把整个系统从头到尾做完整。
There, you know, we were super focused on kind of doing the whole thing end to end.
对。
Mhmm.
我们得学会怎么造这个自动驾驶系统，怎么测试、怎么稳定运营、24小时不间断服务客户，从客户反馈里学习，而且当时特别聚焦在奇andler这一块区域。
You know, learn how to build the driver, evaluate it, deploy regularly, operate it end to end twenty twenty four seven with customers, learn from the customers, and then we're very focused on that operating domain of, know, mostly Chandler.
选了个中等偏低复杂度的地方起步。
See a medium, low complexity one.
等我们升级到第五代系统——也就是现在车上用的这套——我们就想彻底打破原来的运营范围限制。
Then when we made the jump to the fifth generation of our system, this is you know, what's on the I basis today, we really wanted to take a huge bite out of that operating domain.
我们把数据收集遍了全美各个州、各个城市。
And we collected data all over The United States, all different states, different cities.
当我们决定在旧金山和凤凰城最复杂的地段上线时，硬件上做了大升级，更重要的是软件和AI层面的飞跃，我觉得这才是真正的质变。
When we chose to deploy in the hardest parts of San Francisco, hardest parts of Phoenix, we made a big jump on the hardware side and, most importantly, on the software, the AI side, and I would say that was the big discontinuous, jump.
现在你们看到的这些，都是因为我们把自动驾驶的整个研发和部署流程全面升级、反复迭代之后的结果，所以现在我们才能底气十足地说‘行’。
And that that's what you're seeing now after we've, you know, scaled up and, you know, iterated on the, you know, all of the aspects of building and deploying driver, this is now why you're seeing us kinda, you know, go Yes.
同时，我们也在全美范围内加速扩张。
In parallel and scaling, you know, in The US.
所以第五代自动驾驶系统，比第四代通用性强多了？
And So driver version five was just a much more generalizable stack than version four?
具体强在哪儿呢？
And what was it about?
是因为它用的数据集范围更广吗？
Just that it had been trained on a much wider dataset?
关键是我们当时在AI上押了一把大的。
It it was when we made this big bet on AI.
对。
Yeah.
第四代的时候，其实里面塞了挺多小的AI和机器学习模型的。
That was I think there was a lot more, you know, kind of little AI models and ML models in the fourth generation.
我们下了更大的赌注，直接把AI作为第五代的核心架构。
Got We made a much bigger bet and jump to kind of AI as the backbone for the fifth generation.
AI是核心架构。
AI is the backbone
你是说第四代有很多零散的小型AI子系统，对吧？
as the core engine, as in you're saying that Gen four had lots of small little AI subsystems for okay.
对。
Yeah.
是的。
Yeah.
我们完成了这个转变，之后一直在迭代和优化这个模型。
And that's been so we've kind of made that jump, and we've been, you know, iterating and improving the model since then.
咱们能聊聊硬件吗？
Can we go to hardware a second?
关于硬件的问题可多了。
So lots of hardware questions.
但有一点，这个领域里的每个人都有一个特别有魅力的演示车，而且是专门定制的
But one is maybe, everyone in this space has a very charismatic demo of a vehicle that is custom made
是的
Yeah.
专门为自动驾驶设计的
For self driving.
所以呢，经常看到那种没有方向盘、座椅面对面的面包车
And so, you know, it's often the van with the, you know, no steering wheel, seats facing in both directions.
你们自己也有这样的车
You know, you guys have one.
特斯拉有那个没有方向盘的Cyber Cab
Tesla has the steering wheel less cyber cab.
Cruise有他们的Cruise Origin
You know, Cruise had the Cruise Origin.
而我们呢，还在开捷豹，对吧
And, we're still driving in Jaguars that Yeah.
前面还装着方向盘，跟普通家用车差不多。
Have a steering wheel in the front and are pretty similar to consumer cars.
我觉得挺有意思的是，要是十年前聊这个，我们可能会说，嗯，造一辆专属车其实没那么难。
And it's interesting to me because, know, if we were, talking about this ten years ago, we might say, well, yeah, developing a custom car, like, that's relatively straightforward.
我们知道怎么在新车上装一堆传感器，但软件得花很长时间才能搞定。
We know how to put a bunch of sensors on a new car, but the software will take a long time.
有意思的是，软件我们进步巨大，但车子还是基于大家平时开的那些车型改的。
And what's interesting is we've made huge progress in the software, but interestingly, the cars are still derivatives of, you know, cars that people are driving.
所以我很好奇，到2026年了，为啥定制硬件还没真正落地？
And so I'm curious why you just think the custom hardware has not happened as of 2026.
跟Waymo那种颠覆性进步比，这确实算个小改进，但奇怪的是，它到现在还没发生。
It's obviously it's it's a small improvement compared to, you know, Waymo is the big improvement, but it's just interesting that it still has to happen.
好吧，就说我们车的第六代吧，对。
Well, let's say our sixth generation of the vehicle Mhmm.
而这个驾驶员，就是我们自己的版本。
And the driver, is our version of that.
哦，天哪。
Oh, no.
确实如此。
It is.
就是那个Ohike平台，你知道的。
The Ohike, you know, platform.
对吧？
Right?
所以呢，它还是保留了那个……是的。
So that is you know, it still has the Yes.
咱们可以聊聊，比如座椅要不要朝后放，你知道的。
You know, we can talk about, you know, whether you want to have the seats Sure.
座椅朝后还是朝前，这事儿。
Pointed backwards or not.
说实话，demo里看着挺酷，但实际用起来……是的。
I actually, you know, think it's you know, it looks nice in a demo, but practically speaking Yes.
可能这不是个好办法。
Maybe not the way to go.
但这是个定制设计的方案。
But that is, it is a custom designed Yes.
这辆车我们花了很多心思，想摆脱那种以司机为中心的设计。
Vehicle, and it is we put a lot of thought into, you know, moving away from a car that's designed around Yes.
司机。
Driver Yes.
转向以乘客为中心的设计。
To a car that's designed around Yes.
乘客。
Passenger.
空间大得多，而且这事儿已经在进行了。
And it's, you know, much more spacious, like but and it's it's happening.
目前还没对公众开放，你知道的。
It's you know, we're not it's not open to the, you know, the public yet.
不过前几天我坐了一趟，全程自动驾驶，今年就能实现了。
But, you know, I took a ride in it the other day, fully autonomously, and that's coming this year.
对。
Yes.
作为乘客，体验好多少啊？
How much better is it as a passenger experience?
你亲自试试就知道了。
You'll tell me once you give it a try.
我超爱这个。
I love it.
好吧。
Okay.
说到底，就是空间大，没错。
It's, so it's, yeah, it's all about the space Yeah.
还有上下车的方便，还有那些屏幕和后排的交互界面。
And the convenience of, you know, ingress and egress and the the screens and the interface of the pasture.
所以我们对它的每一个细节都花了很多心思。
So we put a lot of thought into every aspect of it.
没错。
Yes.
它装了滑动门。
So it has sliding doors.
上下车特别方便。
It's very easy to get in.
地板是全平的。
It has a flat floor.
是啊。
It is yeah.
坐在后排的话，整个人都能完全伸展开。
If you sit on the back, you can, like, fully stretch out.
对。
Mhmm.
那儿空间大得不得了。
And there's so much space there.
从外面看，它确实挺大的，对吧？
And it it looks you know, from the outside, you know, it looks fairly big.
是啊。
Yes.
没错吧？
Right?
但它的实际占地面积其实没大多少。
But the actual footprint of that is It's not bigger.
就比I-PACE大一点点。
Barely barely barely larger than the I PACE.
对啊。
Mhmm.
所以你一走进去，感觉就像进了客厅一样，简直不可思议。
So it's kind of amazing that, you know, you walk in, just it feels like you're in the living room.
是的。
Yes.
我想问的是，Waymo现在每年用捷豹I-PACE跑大约两千五百万单，差不多这个量级。
I guess my question is just, you know, Waymo does, you know, 25,000,000, rides a year run rate ish with the Jaguar I PACE.
挺有意思的是，自动驾驶到目前为止的规模化，居然主要靠的是这种老车改造的方式。
And it's interesting that so much scaling has happened with self driving so far on the old, you know, retrofit yeah.
也许这其实挺正常的。
Maybe that's to be expected.
嗯，
Well,
这跟高端车型的定位是匹配的，我觉得这也不是理所当然的。
it matches the high I don't think, you know, it's a given.
你说得对。
You're you're right.
我觉得吧，如果你想想它的价值主张——
I think I but if you think about the value proposition Mhmm.
对，当然还有安全性这个问题。
Right, of course, there is the safety of it.
没错。
Yep.
你得担心这个。
You have to worry about.
还有隐私问题。
There's also the, privacy.
没错。
Yep.
一个人坐车，或者跟别人一起，但不用跟任何人打交道。
Being in the car by yourself, maybe, you know, with other folks, but not having to share a touch base with another human.
对吧？
Right?
也许吧，不行。
Maybe No.
Waymo的车
Waymo has
产品真不错。
great products.
是的。
Yeah.
我想这也就是为什么我们看到自动驾驶表现这么稳定，开得特别靠谱。
But I guess this is why we're seeing such, you know, consistency in card, you know, drives well.
对。
Yes.
特别可预测。
You know, very predictable.
而且你还能做得更进一步。
And, you know, you can go beyond that.
对吧？
Right?
而且你还进一步细化，让体验变得更神奇，是的。
And you specialize even more to make the experience even more magical Yes.
围绕乘客本身来打造。
Around the rider.
但我想，要是没有这种专属车型，估计会挺让人失望的，我觉得我肯定会惊讶的，对。
But I guess it's, you know, it would have been disappointing if, you know, without the specialized car and I I think I would have been surprised Mhmm.
如果我们停留在某个低得多的水平上，是的。
If we leveled off, you know, at some other much lower Yes.
客户采纳率要是没这么高，那车本身看起来只是个优化，真正核心的价值其实来自那些其他因素。
Level of customer adoption because, yeah, a car seems like, you know, more of an optimization improvement, but the core of the value proposition comes from those other factors.
对。
Yes.
没错。
Yes.
我想，还是得一次只冒险一件事。
I guess it just take risk on one thing at a time.
我们先从软件层开始，然后再去开发专门的汽车之类的。
We'll start by, you know, doing the software layer, and then we'll build a specialized car or something like that.
没错。
That's right.
对，就是这样。
That's Yeah.
是的。
Yeah.
而且它还
It's also
嗯，我
well, I
我的意思是，就像你说的，这需要一大笔投入。
mean, as you said, it's a big investment.
对。
Yes.
所以你得先把基础部分的风险降下来。
So you have to, like you derisk the fundamentals.
对。
Yes.
而且，说真的，我们一直以来都特别注重给公司设定最大的目标。
And, you know, throughout our history, we were very focused on setting the most you know, the biggest goal for the company
没错。
Mhmm.
就是把最关键的问题先解决掉。
To derisk the most important questions.
对吧？
Right?
我们之前聊过第三代，那时候我们想做一套端到端的系统。
We talked about, you know, the third generation where, you know, we wanted to deploy something and go end to end.
我们还讨论过第四代的目标，哦，抱歉。
We talked about the what was the goal with the fourth generation and oh, sorry.
第五代。
The fifth generation.
对。
Yep.
然后就是第六代了。
And then there's the sixth generation.
对吧？
Right?
所以到第六代的时候，才值得投入这么多精力去做定制化的东西。
So it was the sixth generation where it made sense to go out spend all this, you know, effort into the custom
第六代既是定制化的车辆。
And sixth generation is both the custom vehicle.
它也是新一代的驾驶系统吗？
Is it also a new generation of the the driving stack?
是的。
Yeah.
这是全新的硬件。
It is the new hardware.
对。
Yep.
传感器啊，硬件啊，还有车上用的软硬件，全都升级了。
The sensors, you know, the hardware the, you know, software and hardware they're putting on the the OHAI vehicle Yeah.
这就是第六代。
Is the sixth generation.
对。
Yep.
跟第五代比，差别可大了。
It is very different from the fifth generation.
更简单了。
It is simpler.
能力却更强了。
It is more capable.
成本低多了。
It is much lower cost.
成本只有以前的一小部分，但性能却差不了多少。
It's, you know, a fraction of the cost that's, you know, comparable to what you would get Yes.
就像现在那些高端的驾驶辅助系统一样。
Like a fancy ADAS system, nowadays, the driver assist system.
是的。
Yeah.
对。
Yeah.
软件基本没变。
The software is pretty much the same.
没错。
Mhmm.
所以说到Waymo司机的通用性，就是这样。
So that's another so when we talk about generalizability of the Waymo drivers Yes.
是的。
Yes.
我们常说天气条件。
You know, we talk about weather conditions.
我们谈城市，但这套系统在不同车型和不同配置上也能很好适应，没错。
We talk about cities, but, it also generalizes well to different vehicle platforms and different Yes.
传感器配置上也一样。
Sensor configurations.
明白了。
Okay.
第六代是个新车，传感器配置也全新了，但软件差不多，这简直像在重复TikTok的循环。
So Gen six is a, new vehicle and a new sensor stack, a similar so it's almost a TikTok cycle happening here.
软件几乎是同一套。
It's a similar software.
没错。
That's right.
没错。
That's right.
然后我们会部署第六代Waymo Driver，是的。
And then we're gonna put the, sixth generation Waymo Driver Yep.
比如即将推出的现代IONIQ等其他车型平台，是的。
On other vehicle platforms, like the Hyundai Ioniq that's coming, you know Yep.
今年晚些时候。
Later in the year.
第六代硬件堆栈有什么不同，你们是怎么做到降低成本的？
What is different about the sixth generation hardware stack, and how did you make it cheaper?
它仍然保留了三种相同的感知模式，但我们在所有三个方面都做了重大优化。
So it still has the same three sensing modalities, but we've made significant optimizations in, all three.
对
Yep.
所以就是统一化和简化，嗯
So unification, simplification Mhmm.
还有就是，那种随手写写的 Yeah。
And there's just, you know, the kind of just writing the Yeah.
是不是因为制造规模扩大了，我们接下来要？
Is it a case of, you know, manufacturing scale where we're gonna
会做得越来越多？嗯，规模还没完全上来，但所有这些——你想想供应链、整个行业，确实是这样。
be doing more and Well, scale hasn't fully come into place, but all of those, you know, if you think about the kind of the the supply chains, the industries Yes.
对吧？
Right?
摄像头这块已经很成熟了。
Cameras is pretty mature.
是啊。
Mhmm.
雷达也是，Yeah。
Radars Yeah.
很多年前，雷达又大又复杂，还特别贵。
Way, you know, many years ago used to be bulky, complex, very expensive.
对。
Yep.
你想想，以前我们把它们装在飞机上的时候，对吧？
You know, when we were putting them in planes and you know?
但后来我们开始把它们装到车上。
But then we started putting them on cars.
现在你花个几万美元，就能买到不错的车载雷达了。
Now you can get a, you know, decent automotive radar for, you know, tons of dollars.
没错。
Mhmm.
车载雷达其实有个变种，叫成像雷达，能提供更丰富的数据。
There is, you know, a variant of the automotive radar, and it's called imaging radar, and it gives you a richer something.
这种成像雷达的成本也大幅下降了，只是稍微慢了一点。
So that is also, you know, has come down in cost drastically, but it's a little bit behind Mhmm.
相比普通的车载雷达。
Your standard automotive radars.
激光雷达的发展路径也是一样，非常可预测。
LiDARs are following the same, very predictable Yes.
这是个广为人知的趋势。
Very well known trend.
所以我们现在不仅在做记录，还从上一代技术里吸取经验，不断改进、简化和优化。
So we're, you know, writing that, and we're also, you know, learning from the previous generation to just make improvements and simplifications and optimizations.
我问个有点傻的问题。
So a very silly question.
在自动驾驶里，激光雷达和雷达各自擅长什么？
What are LiDARs versus radars better at in a self driving context?
激光雷达啊，
LiDAR Or
它们是互补的吗？
are they complementary?
它们完全是互补的。
They're very complementary.
是的
Yeah.
你知道，这全是不断发射的信号。
You know, it it's all blasting.
对
Mhmm.
明白吧？
You know?
环保的。
Ecological.
有效地。
Effectively.
就是说，不停地往外发射光脉冲，然后碰到东西就反弹回来。
Like, you know, blasting, you know, photos out there, and then, they bounce off of something.
再返回来。
They come back.
你知道吧，你测的是返回来的信号。
You know, you measure what comes back.
这些频率差别可大了。
The frequencies are very different.
没错。
Yep.
所以激光能提供极其精细的分辨率。
So laser, gives you it's, very, very high resolution.
你可以想象，一束激光射出去，然后不停地旋转。
So you can, you know, think of it as like a laser beam that goes out, you know, spins around.
它每秒能发射数百万个激光脉冲。
It, you know, shoots out millions of these laser pulses, you know, per second.
每个脉冲都反射回来，你就相当于用超高精度在采样整个三维世界。
And then each one comes back, and you can, you know, kind of you're kind of sampling the three d structure of the world with very high resolution.
LiDAR 用于超精细的测绘。
The LiDAR for very fine grained mapping.
没错。
That's right.
对。
Yeah.
雷达的分辨率低得多，但因为它本身的物理特性，确实如此。
Radar has much lower resolution, but because of, know, the physics of it Yes.
在恶劣天气下，它的表现要好得多。
It can, it degrades much better in, adverse weather conditions.
比如大雾、下雪，还有暴雨，所以它不会
So fog, snow, you know, heavy rain So it's
被遮挡，完全没错。
not gonna be occluded by Exactly.
在雷达和目标之间。
Between them and the target.
想象一下，你在浓雾里开车。
So imagine driving in super dense fog.
是的
Yes.
咱们快到旧金山了，所以可能没那么难应付。
We're close to San Francisco, so probably don't have to think, you that hard.
视线真的会很差。
It can be really hard to see.
对啊
Mhmm.
摄像头在恶劣天气下表现会变差。
So cameras degrade.
没错。
Yes.
激光雷达嘛，看颗粒物大小，有时候比摄像头好，有时候更差。
Laser, you know, depending on kind of the the size of the particulates, can, you degrade better or worse than camera.
雷达就不怎么受影响。
Radar is not well affected.
你可以想象一下，在高速公路上开车，雷达能特别清楚地探测到那些在摄像头里完全看不见的车辆。
So, you know, you can imagine driving on a freeway, then radar will give you really good returns for, you know, cars that are absolutely, you know, invisible in the, you know, in the camera space.
哦，这挺有意思的。
Oh, that's interesting.
那是不是有些情况下，你会更依赖雷达？
So so does that mean there are some environments where you'll be relying significantly more on radar?
但它的表现
It's But the performance
够用吗？
is good enough?
其实呢，是靠多种传感器配合使用的。
Well, so it's it's a combination of the sensors.
对吧？
Right?
每个传感器其实都有点噪声，我们得靠它们互相补充。
So we we rely on you know, each one is noisy.
对吧？
Right?
不同环境下，噪声的表现方式不一样，但我们并不是在这些传感器之间来回切换。
How the noise characteristic show up in different environments, you know, is different, but it is I mean, it's not like we switch from one to another.
我们不是先分别用摄像头、雷达和激光雷达去估算世界发生了什么，然后再去比对。
It's not like we know we estimate, you know, what's happening with the world through cameras and through radars and through LiDAR, and then we compare.
不是这样的。
No.
摄像头有自己的一套编码器。
They're like, there's an encoder for camera.
激光雷达也有自己的编码器。
There's an encoder for LiDAR.
它们各自有处理模块，所有数据都汇入同一个系统，共同给你一个最准确的全局视角。
There's a comer and they all go into the, you know, the the system, that gives you jointly the best view of what's happening, in the world.
如果天气晴朗、阳光充足，摄像头的作用就特别大。
So if you were, you know, if it's a nice, bright, sunny day, cameras are, you know, very valuable.
要是天黑得伸手不见五指，或者阳光直射眼睛，又被对向车灯晃得睁不开眼，摄像头就完蛋了。
If, you know, it's pitch dark or you have, like, sun in your face or you're blinded by the headlights from, you know, oncoming car, then camera will degrade.
还是会有点信号，但效果肯定大打折扣。
Like, there's still some, you know, noisy signal, but it will degrade.
对。
Yes.
雷达和激光雷达完全不受影响。
And radar light LiDAR is completely unaffected.
对吧？
Right?
有没有哪些技术难题是你们一直想啃下来的硬骨头？或者有什么特别想解决的问题？
Are there technical problems that are your white whale or you're just you're still chasing or you are particularly interested in solving?
哪怕是一些比较小众的场景，比如真下大雪的时候开车、旧金山那些陡坡，你们是不是特别想搞定？
Even if they're kinda niche for the know, we just we really want to have, you know, driving when it's actually snowing nailed or steep hills in San Francisco or you know?
历史上或者现在，有没有哪些问题你们一直特别关注？
Are there problems you've been very interested in historically or still are?
我在听。
I listen.
现在我特别兴奋，因为我们的全球扩张正在加速。
Super excited right now about the accelerating global expansion.
对。
Mhmm.
美国更多城市正在加入，还要拓展到海外。
More cities in The United States and going internationally.
没错。
Yes.
所以，我有点没搞明白。
So being, I don't understand.
我刚才没回答你关于技术的问题吗？
I'm not answering your your question about the technology?
你等会儿，我回头再聊这个。
I'll, you come back to that.
但说真的，这事儿现在最让我兴奋。
But, really, that's the thing that I'm, you know, today most excited about.
就是说，你懂的，以后不管飞到哪个大城市，下了飞机直接坐Waymo，想去哪儿都行。
Just, you know, go be know, getting to a place where any major, you know, metropolitan area you can fly into the airport and then take a Waymo and go anywhere you wanna go.
对，就是这样。
Like, that is Yes.
没错。
Yes.
这事儿现在真让我热血沸腾。
Insanely exciting to me right now.
所以，从技术上讲，最让我激动的就是AI的飞速发展。
So then, you know, technically, what, I'm most excited about is, all of the rapid progress in AI.
对。
Mhmm.
还有那些世界模型、基础模型的突破。
And the the world models, the foundational model work.
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。