Chris Urmson: Self-Driving Cars at Aurora, Google, CMU, and DARPA
技术与编程音乐与艺术商业与创业AI 与机器学习政治与社会
📋 章节目录
暂无章节信息
🔑 关键词
technologyvehicledondrivingurbanchallengegoingablevehiclessafetydriverdrivemilesautonomousroaddoingsensordidnreallidar
💬 精彩语录
暂无语录
🎙️ 完整对话（901 条）
Lex Fridman (00:00.000)
The following is a conversation with Chris Sampson.
以下是与克里斯·桑普森的对话。
Lex Fridman (00:03.120)
He was a CTO of the Google self driving car team,
他曾担任谷歌自动驾驶汽车团队的CTO，
Lex Fridman (00:06.000)
a key engineer and leader behind the Carnegie Mellon
卡内基梅隆大学背后的关键工程师和领导者
Lex Fridman (00:08.880)
University autonomous vehicle entries in the DARPA Grand
DARPA Grand 中的大学自动驾驶车辆参赛作品
Lex Fridman (00:12.000)
Challenges and the winner of the DARPA Urban Challenge.
DARPA 城市挑战赛的挑战和获胜者。
Chris Urmson (00:16.160)
Today, he's the CEO of Aurora Innovation, an autonomous
如今，他是 Aurora Innovation 的首席执行官，该公司是一家自主
Lex Fridman (00:20.100)
vehicle software company.
车辆软件公司。
Chris Urmson (00:21.360)
He started with Sterling Anderson,
他从斯特林·安德森开始，
Lex Fridman (00:23.600)
who was the former director of Tesla Autopilot,
谁是特斯拉自动驾驶仪的前总监，
Lex Fridman (00:25.960)
and drew back now, Uber's former autonomy and perception lead.
优步以前的自主权和认知领先地位现在退缩了。
Lex Fridman (00:30.120)
Chris is one of the top roboticists and autonomous
克里斯是顶尖的机器人专家和自主者之一
Chris Urmson (00:32.880)
vehicle experts in the world, and a longtime voice
全球汽车专家，长期发声
Lex Fridman (00:36.320)
of reason in a space that is shrouded
被笼罩的空间中的理性
Chris Urmson (00:38.840)
in both mystery and hype.
无论是神秘还是炒作。
Lex Fridman (00:41.320)
He both acknowledges the incredible challenges
他都承认面临着令人难以置信的挑战
Chris Urmson (00:43.600)
involved in solving the problem of autonomous driving
参与解决自动驾驶问题
Lex Fridman (00:46.480)
and is working hard to solve it.
并正在努力解决它。
Chris Urmson (00:49.760)
This is the Artificial Intelligence podcast.
这是人工智能播客。
Lex Fridman (00:52.400)
If you enjoy it, subscribe on YouTube,
如果您喜欢，请在 YouTube 上订阅，
Chris Urmson (00:54.720)
give it five stars on iTunes, support it on Patreon,
在 iTunes 上给它五颗星，在 Patreon 上支持它，
Lex Fridman (00:57.920)
or simply connect with me on Twitter
Chris Urmson (00:59.720)
at Lex Friedman, spelled F R I D M A N.
Lex Fridman (01:03.240)
And now, here's my conversation with Chris Sampson.
Chris Urmson (01:09.120)
You were part of both the DARPA Grand Challenge
Lex Fridman (01:11.960)
and the DARPA Urban Challenge teams
Chris Urmson (01:13.880)
at CMU with Red Whitaker.
Lex Fridman (01:17.040)
What technical or philosophical things
Lex Fridman (01:19.720)
have you learned from these races?
Lex Fridman (01:22.240)
I think the high order bit was that it could be done.
Chris Urmson (01:26.600)
I think that was the thing that was
Lex Fridman (01:30.200)
incredible about the first of the Grand Challenges,
Chris Urmson (01:34.880)
that I remember I was a grad student at Carnegie Mellon,
Lex Fridman (01:38.160)
and there was kind of this dichotomy of it
Chris Urmson (01:45.360)
seemed really hard, so that would
Lex Fridman (01:46.720)
be cool and interesting.
Lex Fridman (01:48.800)
But at the time, we were the only robotics institute around,
Lex Fridman (01:52.800)
and so if we went into it and fell on our faces,
Chris Urmson (01:55.560)
that would be embarrassing.
Lex Fridman (01:58.360)
So I think just having the will to go do it,
Chris Urmson (02:01.120)
to try to do this thing that at the time
Lex Fridman (02:02.880)
was marked as darn near impossible,
Lex Fridman (02:05.000)
and then after a couple of tries,
Lex Fridman (02:06.960)
be able to actually make it happen,
Chris Urmson (02:08.420)
I think that was really exciting.
Lex Fridman (02:12.320)
But at which point did you believe it was possible?
Lex Fridman (02:15.040)
Did you from the very beginning?
Lex Fridman (02:16.960)
Did you personally?
Chris Urmson (02:18.000)
Because you're one of the lead engineers.
Lex Fridman (02:19.800)
You actually had to do a lot of the work.
Chris Urmson (02:21.800)
Yeah, I was the technical director there,
Lex Fridman (02:23.880)
and did a lot of the work, along with a bunch
Chris Urmson (02:26.120)
of other really good people.
Lex Fridman (02:28.420)
Did I believe it could be done?
Chris Urmson (02:29.760)
Yeah, of course.
Lex Fridman (02:31.080)
Why would you go do something you thought
Lex Fridman (02:32.760)
was completely impossible?
Lex Fridman (02:34.800)
We thought it was going to be hard.
Chris Urmson (02:36.260)
We didn't know how we were going to be able to do it.
Lex Fridman (02:37.800)
We didn't know if we'd be able to do it the first time.
Chris Urmson (02:42.880)
Turns out we couldn't.
Lex Fridman (02:45.960)
That, yeah, I guess you have to.
Lex Fridman (02:48.400)
I think there's a certain benefit to naivete, right?
Lex Fridman (02:52.960)
That if you don't know how hard something really is,
Chris Urmson (02:55.440)
you try different things, and it gives you an opportunity
Lex Fridman (02:59.600)
that others who are wiser maybe don't have.
Lex Fridman (03:04.120)
What were the biggest pain points?
Lex Fridman (03:05.720)
Mechanical, sensors, hardware, software,
Chris Urmson (03:08.880)
algorithms for mapping, localization,
Lex Fridman (03:11.800)
just general perception, control?
Lex Fridman (03:13.680)
Like hardware, software, first of all?
Lex Fridman (03:15.320)
I think that's the joy of this field, is that it's all hard
Lex Fridman (03:20.120)
and that you have to be good at each part of it.
Lex Fridman (03:25.360)
So for the urban challenges, if I look back at it from today,
Chris Urmson (03:32.360)
it should be easy today, that it was a static world.
Lex Fridman (03:38.960)
There weren't other actors moving through it,
Chris Urmson (03:40.800)
is what that means.
Lex Fridman (03:42.480)
It was out in the desert, so you get really good GPS.
Lex Fridman (03:47.080)
So that went, and we could map it roughly.
Lex Fridman (03:51.400)
And so in retrospect now, it's within the realm of things
Chris Urmson (03:55.160)
we could do back then.
Lex Fridman (03:57.840)
Just actually getting the vehicle and the,
Chris Urmson (03:59.720)
there's a bunch of engineering work
Lex Fridman (04:00.680)
to get the vehicle so that we could control it and drive it.
Chris Urmson (04:04.760)
That's still a pain today, but it was even more so back then.
Lex Fridman (04:09.600)
And then the uncertainty of exactly what they wanted us to do
Chris Urmson (04:14.280)
was part of the challenge as well.
Lex Fridman (04:17.040)
Right, you didn't actually know the track heading in here.
Chris Urmson (04:19.440)
You knew approximately, but you didn't actually
Lex Fridman (04:21.480)
know the route that was going to be taken.
Chris Urmson (04:23.520)
That's right, we didn't know the route.
Lex Fridman (04:24.920)
We didn't even really, the way the rules had been described,
Chris Urmson (04:28.600)
you had to kind of guess.
Lex Fridman (04:29.800)
So if you think back to that challenge,
Chris Urmson (04:33.360)
the idea was that the government would give us,
Lex Fridman (04:36.960)
the DARPA would give us a set of waypoints
Lex Fridman (04:40.320)
and kind of the width that you had to stay within
Lex Fridman (04:43.520)
between the line that went between each of those waypoints.
Lex Fridman (04:46.800)
And so the most devious thing they could have done
Lex Fridman (04:49.280)
is set a kilometer wide corridor across a field
Chris Urmson (04:53.280)
of scrub brush and rocks and said, go figure it out.
Lex Fridman (04:58.520)
Fortunately, it really, it turned into basically driving
Chris Urmson (05:01.920)
along a set of trails, which is much more relevant
Lex Fridman (05:05.000)
to the application they were looking for.
Lex Fridman (05:08.760)
But no, it was a hell of a thing back in the day.
Lex Fridman (05:12.080)
So the legend, Red, was kind of leading that effort
Chris Urmson (05:16.640)
in terms of just broadly speaking.
Lex Fridman (05:19.120)
So you're a leader now.
Lex Fridman (05:22.040)
What have you learned from Red about leadership?
Lex Fridman (05:25.000)
I think there's a couple things.
Chris Urmson (05:26.200)
One is go and try those really hard things.
Lex Fridman (05:31.080)
That's where there is an incredible opportunity.
Chris Urmson (05:34.480)
I think the other big one, though,
Lex Fridman (05:36.560)
is to see people for who they can be, not who they are.
Chris Urmson (05:41.720)
It's one of the things that I actually,
Lex Fridman (05:43.720)
one of the deepest lessons I learned from Red
Chris Urmson (05:46.080)
was that he would look at undergraduates
Lex Fridman (05:50.200)
or graduate students and empower them to be leaders,
Chris Urmson (05:56.120)
to have responsibility, to do great things
Lex Fridman (06:00.320)
that I think another person might look at them
Lex Fridman (06:04.480)
and think, oh, well, that's just an undergraduate student.
Lex Fridman (06:06.600)
What could they know?
Lex Fridman (06:08.680)
And so I think that kind of trust but verify,
Lex Fridman (06:12.720)
have confidence in what people can become,
Chris Urmson (06:14.480)
I think is a really powerful thing.
Lex Fridman (06:16.680)
So through that, let's just fast forward through the history.
Lex Fridman (06:20.440)
Can you maybe talk through the technical evolution
Lex Fridman (06:24.160)
of autonomous vehicle systems
Chris Urmson (06:26.200)
from the first two Grand Challenges to the Urban Challenge
Lex Fridman (06:29.960)
to today, are there major shifts in your mind
Lex Fridman (06:33.560)
or is it the same kind of technology just made more robust?
Lex Fridman (06:37.240)
I think there's been some big, big steps.
Lex Fridman (06:40.880)
So for the Grand Challenge,
Lex Fridman (06:43.720)
the real technology that unlocked that was HD mapping.
Chris Urmson (06:51.400)
Prior to that, a lot of the off road robotics work
Lex Fridman (06:55.160)
had been done without any real prior model
Chris Urmson (06:58.480)
of what the vehicle was going to encounter.
Lex Fridman (07:01.400)
And so that innovation that the fact that we could get
Chris Urmson (07:05.960)
decimeter resolution models was really a big deal.
Lex Fridman (07:13.440)
And that allowed us to kind of bound the complexity
Chris Urmson (07:18.200)
of the driving problem the vehicle had
Lex Fridman (07:19.680)
and allowed it to operate at speed
Chris Urmson (07:21.040)
because we could assume things about the environment
Lex Fridman (07:23.800)
that it was going to encounter.
Lex Fridman (07:25.360)
So that was the big step there.
Lex Fridman (07:31.280)
For the Urban Challenge,
Chris Urmson (07:37.240)
one of the big technological innovations there
Lex Fridman (07:39.280)
was the multi beam LIDAR
Lex Fridman (07:41.960)
and being able to generate high resolution,
Lex Fridman (07:45.760)
mid to long range 3D models of the world
Lex Fridman (07:48.680)
and use that for understanding the world around the vehicle.
Lex Fridman (07:53.680)
And that was really kind of a game changing technology.
Chris Urmson (07:58.600)
In parallel with that,
Lex Fridman (08:00.000)
we saw a bunch of other technologies
Chris Urmson (08:04.360)
that had been kind of converging
Lex Fridman (08:06.120)
half their day in the sun.
Lex Fridman (08:08.440)
So Bayesian estimation had been,
Lex Fridman (08:12.560)
SLAM had been a big field in robotics.
Chris Urmson (08:17.840)
You would go to a conference a couple of years before that
Lex Fridman (08:20.760)
and every paper would effectively have SLAM somewhere in it.
Lex Fridman (08:24.880)
And so seeing that the Bayesian estimation techniques
Lex Fridman (08:30.720)
play out on a very visible stage,
Chris Urmson (08:33.400)
I thought that was pretty exciting to see.
Lex Fridman (08:38.080)
And mostly SLAM was done based on LIDAR at that time.
Chris Urmson (08:41.560)
Yeah, and in fact, we weren't really doing SLAM per se
Lex Fridman (08:45.600)
in real time because we had a model ahead of time,
Chris Urmson (08:47.480)
we had a roadmap, but we were doing localization.
Lex Fridman (08:51.040)
And we were using the LIDAR or the cameras
Chris Urmson (08:53.560)
depending on who exactly was doing it
Lex Fridman (08:55.400)
to localize to a model of the world.
Lex Fridman (08:57.560)
And I thought that was a big step
Lex Fridman (09:00.160)
from kind of naively trusting GPS, INS before that.
Lex Fridman (09:06.640)
And again, lots of work had been going on in this field.
Lex Fridman (09:09.840)
Certainly this was not doing anything
Chris Urmson (09:13.040)
particularly innovative in SLAM or in localization,
Lex Fridman (09:16.840)
but it was seeing that technology necessary
Chris Urmson (09:20.200)
in a real application on a big stage,
Lex Fridman (09:21.800)
I thought was very cool.
Lex Fridman (09:23.080)
So for the urban challenge,
Lex Fridman (09:24.000)
those are already maps constructed offline in general.
Lex Fridman (09:28.600)
And did people do that individually,
Lex Fridman (09:30.920)
did individual teams do it individually
Lex Fridman (09:33.600)
so they had their own different approaches there
Lex Fridman (09:36.440)
or did everybody kind of share that information
Lex Fridman (09:41.720)
at least intuitively?
Lex Fridman (09:42.880)
So DARPA gave all the teams a model of the world, a map.
Lex Fridman (09:49.640)
And then one of the things that we had to figure out
Lex Fridman (09:53.240)
back then was, and it's still one of these things
Chris Urmson (09:56.080)
that trips people up today
Lex Fridman (09:57.280)
is actually the coordinate system.
Lex Fridman (10:00.280)
So you get a latitude longitude
Lex Fridman (10:03.080)
and to so many decimal places,
Chris Urmson (10:05.040)
you don't really care about kind of the ellipsoid
Lex Fridman (10:07.360)
of the earth that's being used.
Lex Fridman (10:09.560)
But when you want to get to 10 centimeter
Lex Fridman (10:12.240)
or centimeter resolution,
Chris Urmson (10:14.400)
you care whether the coordinate system is NADS 83
Lex Fridman (10:18.520)
or WGS 84 or these are different ways to describe
Chris Urmson (10:24.200)
both the kind of non sphericalness of the earth,
Lex Fridman (10:26.760)
but also kind of the, I think,
Chris Urmson (10:31.080)
I can't remember which one,
Lex Fridman (10:32.080)
the tectonic shifts that are happening
Lex Fridman (10:33.600)
and how to transform the global datum as a function of that.
Lex Fridman (10:37.000)
So getting a map and then actually matching it to reality
Chris Urmson (10:41.020)
to centimeter resolution, that was kind of interesting
Lex Fridman (10:42.880)
and fun back then.
Lex Fridman (10:44.040)
So how much work was the perception doing there?
Lex Fridman (10:46.760)
So how much were you relying on localization based on maps
Lex Fridman (10:52.480)
without using perception to register to the maps?
Lex Fridman (10:55.760)
And I guess the question is how advanced
Lex Fridman (10:58.000)
was perception at that point?
Lex Fridman (10:59.800)
It's certainly behind where we are today, right?
Chris Urmson (11:01.960)
We're more than a decade since the urban challenge.
Lex Fridman (11:05.840)
But the core of it was there.
Chris Urmson (11:08.640)
That we were tracking vehicles.
Lex Fridman (11:13.120)
We had to do that at 100 plus meter range
Chris Urmson (11:15.640)
because we had to merge with other traffic.
Lex Fridman (11:18.320)
We were using, again, Bayesian estimates
Chris Urmson (11:21.240)
for state of these vehicles.
Lex Fridman (11:23.860)
We had to deal with a bunch of the problems
Chris Urmson (11:25.580)
that you think of today,
Lex Fridman (11:26.920)
of predicting where that vehicle's going to be
Chris Urmson (11:29.820)
a few seconds into the future.
Lex Fridman (11:31.060)
We had to deal with the fact
Chris Urmson (11:32.380)
that there were multiple hypotheses for that
Lex Fridman (11:35.320)
because a vehicle at an intersection might be going right
Chris Urmson (11:37.660)
or it might be going straight
Lex Fridman (11:38.780)
or it might be making a left turn.
Lex Fridman (11:41.500)
And we had to deal with the challenge of the fact
Lex Fridman (11:44.120)
that our behavior was going to impact the behavior
Chris Urmson (11:47.600)
of that other operator.
Lex Fridman (11:48.960)
And we did a lot of that in relatively naive ways,
Lex Fridman (11:53.480)
but it kind of worked.
Lex Fridman (11:54.820)
Still had to have some kind of solution.
Lex Fridman (11:57.080)
And so where does that, 10 years later,
Lex Fridman (11:59.960)
where does that take us today
Chris Urmson (12:01.520)
from that artificial city construction
Lex Fridman (12:04.260)
to real cities to the urban environment?
Chris Urmson (12:07.000)
Yeah, I think the biggest thing
Lex Fridman (12:09.160)
is that the actors are truly unpredictable.
Chris Urmson (12:15.720)
That most of the time, the drivers on the road,
Lex Fridman (12:18.800)
the other road users are out there behaving well,
Lex Fridman (12:24.080)
but every once in a while they're not.
Lex Fridman (12:27.080)
The variety of other vehicles is, you have all of them.
Lex Fridman (12:32.080)
In terms of behavior, in terms of perception, or both?
Lex Fridman (12:35.840)
Both.
Chris Urmson (12:38.740)
Back then we didn't have to deal with cyclists,
Lex Fridman (12:40.520)
we didn't have to deal with pedestrians,
Chris Urmson (12:42.800)
didn't have to deal with traffic lights.
Lex Fridman (12:46.260)
The scale over which that you have to operate is now
Chris Urmson (12:49.400)
is much larger than the air base
Lex Fridman (12:51.120)
that we were thinking about back then.
Lex Fridman (12:52.720)
So what, easy question,
Lex Fridman (12:56.280)
what do you think is the hardest part about driving?
Chris Urmson (12:59.720)
Easy question.
Lex Fridman (13:00.560)
Yeah, no, I'm joking.
Chris Urmson (13:02.560)
I'm sure nothing really jumps out at you as one thing,
Lex Fridman (13:07.440)
but in the jump from the urban challenge to the real world,
Chris Urmson (13:12.920)
is there something that's a particular,
Lex Fridman (13:15.320)
you foresee as very serious, difficult challenge?
Chris Urmson (13:18.480)
I think the most fundamental difference
Lex Fridman (13:21.080)
is that we're doing it for real.
Chris Urmson (13:26.760)
That in that environment,
Lex Fridman (13:28.960)
it was both a limited complexity environment
Chris Urmson (13:31.880)
because certain actors weren't there,
Lex Fridman (13:33.240)
because the roads were maintained,
Chris Urmson (13:35.380)
there were barriers keeping people separate
Lex Fridman (13:37.360)
from robots at the time,
Lex Fridman (13:40.840)
and it only had to work for 60 miles.
Lex Fridman (13:43.300)
Which, looking at it from 2006,
Lex Fridman (13:46.160)
it had to work for 60 miles, right?
Lex Fridman (13:48.960)
Looking at it from now,
Chris Urmson (13:51.880)
we want things that will go and drive
Lex Fridman (13:53.720)
for half a million miles,
Lex Fridman (13:57.160)
and it's just a different game.
Lex Fridman (14:00.940)
So how important,
Chris Urmson (14:03.480)
you said LiDAR came into the game early on,
Lex Fridman (14:06.080)
and it's really the primary driver
Chris Urmson (14:07.880)
of autonomous vehicles today as a sensor.
Lex Fridman (14:10.240)
So how important is the role of LiDAR
Lex Fridman (14:11.920)
in the sensor suite in the near term?
Lex Fridman (14:14.800)
So I think it's essential.
Chris Urmson (14:17.920)
I believe, but I also believe that cameras are essential,
Lex Fridman (14:20.480)
and I believe the radar is essential.
Chris Urmson (14:22.120)
I think that you really need to use
Lex Fridman (14:26.280)
the composition of data from these different sensors
Chris Urmson (14:28.720)
if you want the thing to really be robust.
Lex Fridman (14:32.640)
The question I wanna ask,
Chris Urmson (14:34.360)
let's see if we can untangle it,
Lex Fridman (14:35.600)
is what are your thoughts on the Elon Musk
Chris Urmson (14:39.320)
provocative statement that LiDAR is a crutch,
Lex Fridman (14:42.340)
that it's a kind of, I guess, growing pains,
Lex Fridman (14:47.760)
and that much of the perception task
Lex Fridman (14:49.920)
can be done with cameras?
Lex Fridman (14:52.120)
So I think it is undeniable
Lex Fridman (14:55.440)
that people walk around without lasers in their foreheads,
Lex Fridman (14:59.360)
and they can get into vehicles and drive them,
Lex Fridman (15:01.880)
and so there's an existence proof
Chris Urmson (15:05.600)
that you can drive using passive vision.
Lex Fridman (15:10.880)
No doubt, can't argue with that.
Chris Urmson (15:12.720)
In terms of sensors, yeah, so there's proof.
Lex Fridman (15:14.680)
Yeah, in terms of sensors, right?
Lex Fridman (15:16.000)
So there's an example that we all go do it,
Lex Fridman (15:20.200)
many of us every day.
Chris Urmson (15:21.380)
In terms of LiDAR being a crutch, sure.
Lex Fridman (15:28.180)
But in the same way that the combustion engine
Chris Urmson (15:33.100)
was a crutch on the path to an electric vehicle,
Lex Fridman (15:35.260)
in the same way that any technology ultimately gets
Chris Urmson (15:40.840)
replaced by some superior technology in the future,
Lex Fridman (15:44.380)
and really the way that I look at this
Chris Urmson (15:47.740)
is that the way we get around on the ground,
Lex Fridman (15:51.460)
the way that we use transportation is broken,
Lex Fridman (15:55.280)
and that we have this, I think the number I saw this morning,
Lex Fridman (15:59.740)
37,000 Americans killed last year on our roads,
Lex Fridman (16:04.060)
and that's just not acceptable.
Lex Fridman (16:05.380)
And so any technology that we can bring to bear
Chris Urmson (16:09.460)
that accelerates this self driving technology
Lex Fridman (16:12.860)
coming to market and saving lives
Chris Urmson (16:14.640)
is technology we should be using.
Lex Fridman (16:18.280)
And it feels just arbitrary to say,
Chris Urmson (16:20.840)
well, I'm not okay with using lasers
Lex Fridman (16:26.240)
because that's whatever,
Lex Fridman (16:27.820)
but I am okay with using an eight megapixel camera
Lex Fridman (16:30.720)
or a 16 megapixel camera.
Chris Urmson (16:32.880)
These are just bits of technology,
Lex Fridman (16:34.640)
and we should be taking the best technology
Chris Urmson (16:36.360)
from the tool bin that allows us to go and solve a problem.
Lex Fridman (16:41.360)
The question I often talk to, well, obviously you do as well,
Chris Urmson (16:45.160)
to sort of automotive companies,
Lex Fridman (16:48.280)
and if there's one word that comes up more often
Chris Urmson (16:51.360)
than anything, it's cost, and trying to drive costs down.
Lex Fridman (16:55.280)
So while it's true that it's a tragic number, the 37,000,
Chris Urmson (17:01.400)
the question is, and I'm not the one asking this question
Lex Fridman (17:04.880)
because I hate this question,
Lex Fridman (17:05.820)
but we want to find the cheapest sensor suite
Lex Fridman (17:09.960)
that creates a safe vehicle.
Lex Fridman (17:13.280)
So in that uncomfortable trade off,
Lex Fridman (17:18.220)
do you foresee LiDAR coming down in cost in the future,
Chris Urmson (17:23.680)
or do you see a day where level four autonomy
Lex Fridman (17:26.680)
is possible without LiDAR?
Chris Urmson (17:29.880)
I see both of those, but it's really a matter of time.
Lex Fridman (17:32.880)
And I think really, maybe I would talk to the question
Chris Urmson (17:36.040)
you asked about the cheapest sensor.
Lex Fridman (17:37.840)
I don't think that's actually what you want.
Lex Fridman (17:40.360)
What you want is a sensor suite that is economically viable.
Lex Fridman (17:45.680)
And then after that, everything is about margin
Lex Fridman (17:49.440)
and driving costs out of the system.
Lex Fridman (17:52.120)
What you also want is a sensor suite that works.
Lex Fridman (17:55.360)
And so it's great to tell a story about
Lex Fridman (17:59.600)
how it would be better to have a self driving system
Chris Urmson (18:03.260)
with a $50 sensor instead of a $500 sensor.
Lex Fridman (18:08.040)
But if the $500 sensor makes it work
Lex Fridman (18:10.520)
and the $50 sensor doesn't work, who cares?
Lex Fridman (18:15.680)
So long as you can actually have an economic opportunity,
Chris Urmson (18:20.020)
there's an economic opportunity there.
Lex Fridman (18:21.520)
And the economic opportunity is important
Chris Urmson (18:23.760)
because that's how you actually have a sustainable business
Lex Fridman (18:27.760)
and that's how you can actually see this come to scale
Lex Fridman (18:31.120)
and be out in the world.
Lex Fridman (18:32.400)
And so when I look at LiDAR,
Chris Urmson (18:35.960)
I see a technology that has no underlying
Lex Fridman (18:38.880)
fundamentally expense to it, fundamental expense to it.
Chris Urmson (18:42.420)
It's going to be more expensive than an imager
Lex Fridman (18:46.080)
because CMOS processes or FAP processes
Chris Urmson (18:51.360)
are dramatically more scalable than mechanical processes.
Lex Fridman (18:56.200)
But we still should be able to drive costs down
Chris Urmson (18:58.320)
substantially on that side.
Lex Fridman (19:00.120)
And then I also do think that with the right business model
Chris Urmson (19:05.880)
you can absorb more,
Lex Fridman (19:07.560)
certainly more cost on the bill of materials.
Chris Urmson (19:09.480)
Yeah, if the sensor suite works, extra value is provided,
Lex Fridman (19:12.600)
thereby you don't need to drive costs down to zero.
Chris Urmson (19:15.480)
It's the basic economics.
Lex Fridman (19:17.100)
You've talked about your intuition
Chris Urmson (19:18.820)
that level two autonomy is problematic
Lex Fridman (19:22.200)
because of the human factor of vigilance,
Chris Urmson (19:25.920)
decrement, complacency, over trust and so on,
Lex Fridman (19:28.040)
just us being human.
Chris Urmson (19:29.600)
We over trust the system,
Lex Fridman (19:31.120)
we start doing even more so partaking
Chris Urmson (19:34.240)
in the secondary activities like smartphones and so on.
Lex Fridman (19:38.680)
Have your views evolved on this point in either direction?
Lex Fridman (19:43.000)
Can you speak to it?
Lex Fridman (19:44.800)
So, and I want to be really careful
Chris Urmson (19:47.480)
because sometimes this gets twisted in a way
Lex Fridman (19:50.380)
that I certainly didn't intend.
Lex Fridman (19:53.040)
So active safety systems are a really important technology
Lex Fridman (19:58.040)
that we should be pursuing and integrating into vehicles.
Lex Fridman (20:02.080)
And there's an opportunity in the near term
Lex Fridman (20:04.280)
to reduce accidents, reduce fatalities,
Lex Fridman (20:06.520)
and we should be pushing on that.
Lex Fridman (20:11.960)
Level two systems are systems
Chris Urmson (20:14.680)
where the vehicle is controlling two axes.
Lex Fridman (20:18.080)
So braking and throttle slash steering.
Lex Fridman (20:23.480)
And I think there are variants of level two systems
Lex Fridman (20:25.680)
that are supporting the driver.
Chris Urmson (20:27.280)
That absolutely we should encourage to be out there.
Lex Fridman (20:31.080)
Where I think there's a real challenge
Chris Urmson (20:32.880)
is in the human factors part around this
Lex Fridman (20:37.640)
and the misconception from the public
Chris Urmson (20:41.240)
around the capability set that that enables
Lex Fridman (20:43.600)
and the trust that they should have in it.
Lex Fridman (20:46.640)
And that is where I kind of,
Lex Fridman (20:50.000)
I'm actually incrementally more concerned
Chris Urmson (20:52.920)
around level three systems
Lex Fridman (20:54.440)
and how exactly a level two system is marketed and delivered
Lex Fridman (20:58.440)
and how much effort people have put into those human factors.
Lex Fridman (21:01.840)
So I still believe several things around this.
Chris Urmson (21:05.640)
One is people will overtrust the technology.
Lex Fridman (21:09.440)
We've seen over the last few weeks
Chris Urmson (21:11.440)
a spate of people sleeping in their Tesla.
Lex Fridman (21:14.920)
I watched an episode last night of Trevor Noah
Chris Urmson (21:19.920)
talking about this and him,
Lex Fridman (21:23.920)
this is a smart guy who has a lot of resources
Chris Urmson (21:26.720)
at his disposal describing a Tesla as a self driving car
Lex Fridman (21:30.720)
and that why shouldn't people be sleeping in their Tesla?
Lex Fridman (21:33.480)
And it's like, well, because it's not a self driving car
Lex Fridman (21:36.560)
and it is not intended to be
Lex Fridman (21:38.840)
and these people will almost certainly die at some point
Lex Fridman (21:46.400)
or hurt other people.
Lex Fridman (21:48.040)
And so we need to really be thoughtful
Lex Fridman (21:50.080)
about how that technology is described
Lex Fridman (21:51.840)
and brought to market.
Lex Fridman (21:54.240)
I also think that because of the economic challenges
Chris Urmson (21:59.240)
we were just talking about,
Lex Fridman (22:01.240)
that these level two driver assistance systems,
Chris Urmson (22:05.160)
that technology path will diverge
Lex Fridman (22:07.280)
from the technology path that we need to be on
Chris Urmson (22:10.200)
to actually deliver truly self driving vehicles,
Lex Fridman (22:14.080)
ones where you can get in it and drive it.
Chris Urmson (22:16.920)
Can get in it and sleep and have the equivalent
Lex Fridman (22:20.800)
or better safety than a human driver behind the wheel.
Chris Urmson (22:24.680)
Because again, the economics are very different
Lex Fridman (22:28.480)
in those two worlds and so that leads
Chris Urmson (22:30.880)
to divergent technology.
Lex Fridman (22:32.800)
So you just don't see the economics
Chris Urmson (22:34.680)
of gradually increasing from level two
Lex Fridman (22:38.560)
and doing so quickly enough
Chris Urmson (22:41.600)
to where it doesn't cause safety, critical safety concerns.
Lex Fridman (22:44.480)
You believe that it needs to diverge at this point
Chris Urmson (22:48.680)
into basically different routes.
Lex Fridman (22:50.800)
And really that comes back to what are those L2
Lex Fridman (22:55.560)
and L1 systems doing?
Lex Fridman (22:57.080)
And they are driver assistance functions
Chris Urmson (22:59.840)
where the people that are marketing that responsibly
Lex Fridman (23:04.400)
are being very clear and putting human factors in place
Chris Urmson (23:08.000)
such that the driver is actually responsible for the vehicle
Lex Fridman (23:12.440)
and that the technology is there to support the driver.
Lex Fridman (23:15.160)
And the safety cases that are built around those
Lex Fridman (23:19.880)
are dependent on that driver attention and attentiveness.
Lex Fridman (23:24.040)
And at that point, you can kind of give up
Lex Fridman (23:29.160)
to some degree for economic reasons,
Chris Urmson (23:31.240)
you can give up on say false negatives.
Lex Fridman (23:34.800)
And the way to think about this
Chris Urmson (23:36.200)
is for a four collision mitigation braking system,
Lex Fridman (23:39.320)
if it half the times the driver missed a vehicle
Chris Urmson (23:43.960)
in front of it, it hit the brakes
Lex Fridman (23:46.080)
and brought the vehicle to a stop,
Chris Urmson (23:47.680)
that would be an incredible, incredible advance
Lex Fridman (23:51.640)
in safety on our roads, right?
Chris Urmson (23:53.040)
That would be equivalent to seat belts.
Lex Fridman (23:55.000)
But it would mean that if that vehicle
Chris Urmson (23:56.600)
wasn't being monitored, it would hit one out of two cars.
Lex Fridman (24:00.600)
And so economically, that's a perfectly good solution
Chris Urmson (24:05.120)
for a driver assistance system.
Lex Fridman (24:06.280)
What you should do at that point,
Chris Urmson (24:07.240)
if you can get it to work 50% of the time,
Lex Fridman (24:09.240)
is drive the cost out of that
Lex Fridman (24:10.520)
so you can get it on as many vehicles as possible.
Lex Fridman (24:13.320)
But driving the cost out of it
Chris Urmson (24:14.760)
doesn't drive up performance on the false negative case.
Lex Fridman (24:18.800)
And so you'll continue to not have a technology
Chris Urmson (24:21.440)
that could really be available for a self driven vehicle.
Lex Fridman (24:25.680)
So clearly the communication,
Lex Fridman (24:28.440)
and this probably applies to all four vehicles as well,
Lex Fridman (24:31.600)
the marketing and communication
Chris Urmson (24:34.440)
of what the technology is actually capable of,
Lex Fridman (24:37.040)
how hard it is, how easy it is,
Chris Urmson (24:38.400)
all that kind of stuff is highly problematic.
Lex Fridman (24:41.000)
So say everybody in the world was perfectly communicated
Lex Fridman (24:45.640)
and were made to be completely aware
Lex Fridman (24:48.400)
of every single technology out there,
Lex Fridman (24:50.000)
what it's able to do.
Lex Fridman (24:52.840)
What's your intuition?
Lex Fridman (24:54.120)
And now we're maybe getting into philosophical ground.
Lex Fridman (24:56.880)
Is it possible to have a level two vehicle
Lex Fridman (25:00.000)
where we don't over trust it?
Lex Fridman (25:04.680)
I don't think so.
Chris Urmson (25:05.800)
If people truly understood the risks and internalized it,
Lex Fridman (25:11.160)
then sure, you could do that safely.
Lex Fridman (25:14.320)
But that's a world that doesn't exist.
Lex Fridman (25:16.160)
The people are going to,
Chris Urmson (25:18.720)
if the facts are put in front of them,
Lex Fridman (25:20.760)
they're gonna then combine that with their experience.
Lex Fridman (25:24.440)
And let's say they're using an L2 system
Lex Fridman (25:28.360)
and they go up and down the 101 every day
Lex Fridman (25:30.800)
and they do that for a month.
Lex Fridman (25:32.720)
And it just worked every day for a month.
Chris Urmson (25:36.200)
Like that's pretty compelling at that point,
Lex Fridman (25:39.000)
just even if you know the statistics,
Chris Urmson (25:41.800)
you're like, well, I don't know,
Lex Fridman (25:43.400)
maybe there's something funny about those.
Chris Urmson (25:44.760)
Maybe they're driving in difficult places.
Lex Fridman (25:46.920)
Like I've seen it with my own eyes, it works.
Lex Fridman (25:49.840)
And the problem is that that sample size that they have,
Lex Fridman (25:52.400)
so it's 30 miles up and down,
Lex Fridman (25:53.880)
so 60 miles times 30 days,
Lex Fridman (25:56.360)
so 60, 180, 1,800 miles.
Chris Urmson (25:58.720)
Like that's a drop in the bucket
Lex Fridman (26:03.280)
compared to the, what, 85 million miles between fatalities.
Lex Fridman (26:07.640)
And so they don't really have a true estimate
Lex Fridman (26:11.400)
based on their personal experience of the real risks,
Lex Fridman (26:14.440)
but they're gonna trust it anyway,
Lex Fridman (26:15.640)
because it's hard not to.
Lex Fridman (26:16.480)
It worked for a month, what's gonna change?
Lex Fridman (26:18.640)
So even if you start a perfect understanding of the system,
Chris Urmson (26:21.640)
your own experience will make it drift.
Lex Fridman (26:24.160)
I mean, that's a big concern.
Chris Urmson (26:25.920)
Over a year, over two years even,
Lex Fridman (26:28.160)
it doesn't have to be months.
Lex Fridman (26:29.440)
And I think that as this technology moves
Lex Fridman (26:32.920)
from what I would say is kind of the more technology savvy
Chris Urmson (26:37.760)
ownership group to the mass market,
Lex Fridman (26:42.640)
you may be able to have some of those folks
Chris Urmson (26:44.600)
who are really familiar with technology,
Lex Fridman (26:46.280)
they may be able to internalize it better.
Lex Fridman (26:48.840)
And your kind of immunization
Lex Fridman (26:50.800)
against this kind of false risk assessment
Chris Urmson (26:53.360)
might last longer,
Lex Fridman (26:54.280)
but as folks who aren't as savvy about that
Chris Urmson (26:58.680)
read the material and they compare that
Lex Fridman (27:00.880)
to their personal experience,
Chris Urmson (27:02.160)
I think there it's going to move more quickly.
Lex Fridman (27:08.160)
So your work, the program that you've created at Google
Lex Fridman (27:11.280)
and now at Aurora is focused more on the second path
Lex Fridman (27:16.600)
of creating full autonomy.
Lex Fridman (27:18.480)
So it's such a fascinating,
Lex Fridman (27:20.880)
I think it's one of the most interesting AI problems
Lex Fridman (27:24.560)
of the century, right?
Lex Fridman (27:25.600)
It's, I just talked to a lot of people,
Chris Urmson (27:28.280)
just regular people, I don't know,
Lex Fridman (27:29.440)
my mom, about autonomous vehicles,
Lex Fridman (27:31.720)
and you begin to grapple with ideas
Lex Fridman (27:34.520)
of giving your life control over to a machine.
Chris Urmson (27:38.080)
It's philosophically interesting,
Lex Fridman (27:40.040)
it's practically interesting.
Lex Fridman (27:41.760)
So let's talk about safety.
Lex Fridman (27:43.720)
How do you think we demonstrate,
Chris Urmson (27:46.240)
you've spoken about metrics in the past,
Lex Fridman (27:47.880)
how do you think we demonstrate to the world
Lex Fridman (27:51.880)
that an autonomous vehicle, an Aurora system is safe?
Lex Fridman (27:56.160)
This is one where it's difficult
Chris Urmson (27:57.320)
because there isn't a soundbite answer.
Lex Fridman (27:59.280)
That we have to show a combination of work
Chris Urmson (28:05.960)
that was done diligently and thoughtfully,
Lex Fridman (28:08.360)
and this is where something like a functional safety process
Chris Urmson (28:10.840)
is part of that.
Lex Fridman (28:11.680)
It's like here's the way we did the work,
Chris Urmson (28:15.280)
that means that we were very thorough.
Lex Fridman (28:17.160)
So if you believe that what we said
Chris Urmson (28:20.040)
about this is the way we did it,
Lex Fridman (28:21.440)
then you can have some confidence
Chris Urmson (28:22.720)
that we were thorough in the engineering work
Lex Fridman (28:25.200)
we put into the system.
Lex Fridman (28:26.920)
And then on top of that,
Lex Fridman (28:28.920)
to kind of demonstrate that we weren't just thorough,
Chris Urmson (28:32.000)
we were actually good at what we did,
Lex Fridman (28:35.280)
there'll be a kind of a collection of evidence
Chris Urmson (28:38.200)
in terms of demonstrating that the capabilities
Lex Fridman (28:40.440)
worked the way we thought they did,
Chris Urmson (28:42.920)
statistically and to whatever degree
Lex Fridman (28:45.320)
we can demonstrate that,
Chris Urmson (28:48.160)
both in some combination of simulations,
Lex Fridman (28:50.320)
some combination of unit testing
Lex Fridman (28:53.080)
and decomposition testing,
Lex Fridman (28:54.640)
and then some part of it will be on road data.
Lex Fridman (28:58.160)
And I think the way we'll ultimately
Lex Fridman (29:02.680)
convey this to the public
Chris Urmson (29:04.000)
is there'll be clearly some conversation
Lex Fridman (29:06.760)
with the public about it,
Lex Fridman (29:08.200)
but we'll kind of invoke the kind of the trusted nodes
Lex Fridman (29:12.040)
and that we'll spend more time
Chris Urmson (29:13.880)
being able to go into more depth with folks like NHTSA
Lex Fridman (29:17.280)
and other federal and state regulatory bodies
Lex Fridman (29:19.720)
and kind of given that they are
Lex Fridman (29:22.080)
operating in the public interest and they're trusted,
Chris Urmson (29:26.240)
that if we can show enough work to them
Lex Fridman (29:28.640)
that they're convinced,
Chris Urmson (29:30.000)
then I think we're in a pretty good place.
Lex Fridman (29:33.800)
That means you work with people
Chris Urmson (29:35.000)
that are essentially experts at safety
Lex Fridman (29:36.920)
to try to discuss and show.
Lex Fridman (29:39.000)
Do you think, the answer's probably no,
Lex Fridman (29:41.720)
but just in case,
Lex Fridman (29:42.920)
do you think there exists a metric?
Lex Fridman (29:44.360)
So currently people have been using
Chris Urmson (29:46.320)
number of disengagements.
Lex Fridman (29:48.200)
And it quickly turns into a marketing scheme
Chris Urmson (29:50.120)
to sort of you alter the experiments you run to adjust.
Lex Fridman (29:54.280)
I think you've spoken that you don't like.
Chris Urmson (29:56.280)
Don't love it.
Lex Fridman (29:57.120)
No, in fact, I was on the record telling DMV
Chris Urmson (29:59.680)
that I thought this was not a great metric.
Lex Fridman (30:01.960)
Do you think it's possible to create a metric,
Chris Urmson (30:05.280)
a number that could demonstrate safety
Lex Fridman (30:09.440)
outside of fatalities?
Lex Fridman (30:12.320)
So I do.
Lex Fridman (30:13.440)
And I think that it won't be just one number.
Lex Fridman (30:17.600)
So as we are internally grappling with this,
Lex Fridman (30:21.280)
and at some point we'll be able to talk
Chris Urmson (30:23.560)
more publicly about it,
Lex Fridman (30:25.040)
is how do we think about human performance
Chris Urmson (30:28.520)
in different tasks,
Lex Fridman (30:29.840)
say detecting traffic lights
Lex Fridman (30:32.160)
or safely making a left turn across traffic?
Lex Fridman (30:37.680)
And what do we think the failure rates are
Lex Fridman (30:40.080)
for those different capabilities for people?
Lex Fridman (30:42.520)
And then demonstrating to ourselves
Lex Fridman (30:44.760)
and then ultimately folks in the regulatory role
Lex Fridman (30:48.480)
and then ultimately the public
Chris Urmson (30:50.760)
that we have confidence that our system
Lex Fridman (30:52.400)
will work better than that.
Lex Fridman (30:54.760)
And so these individual metrics
Lex Fridman (30:57.040)
will kind of tell a compelling story ultimately.
Chris Urmson (31:01.760)
I do think at the end of the day
Lex Fridman (31:03.920)
what we care about in terms of safety
Chris Urmson (31:06.640)
is life saved and injuries reduced.
Lex Fridman (31:12.160)
And then ultimately kind of casualty dollars
Chris Urmson (31:16.440)
that people aren't having to pay to get their car fixed.
Lex Fridman (31:19.360)
And I do think that in aviation
Chris Urmson (31:22.680)
they look at a kind of an event pyramid
Lex Fridman (31:25.880)
where a crash is at the top of that
Lex Fridman (31:28.600)
and that's the worst event obviously
Lex Fridman (31:30.440)
and then there's injuries and near miss events and whatnot
Lex Fridman (31:34.240)
and violation of operating procedures
Lex Fridman (31:37.320)
and you kind of build a statistical model
Chris Urmson (31:40.160)
of the relevance of the low severity things
Lex Fridman (31:44.440)
or the high severity things.
Lex Fridman (31:45.280)
And I think that's something
Lex Fridman (31:46.120)
where we'll be able to look at as well
Chris Urmson (31:48.200)
because an event per 85 million miles
Lex Fridman (31:51.840)
is statistically a difficult thing
Chris Urmson (31:54.440)
even at the scale of the U.S.
Lex Fridman (31:56.800)
to kind of compare directly.
Lex Fridman (31:59.360)
And that event fatality that's connected
Lex Fridman (32:02.240)
to an autonomous vehicle is significantly
Chris Urmson (32:07.440)
at least currently magnified
Lex Fridman (32:09.160)
in the amount of attention it gets.
Lex Fridman (32:12.320)
So that speaks to public perception.
Lex Fridman (32:15.080)
I think the most popular topic
Chris Urmson (32:16.720)
about autonomous vehicles in the public
Lex Fridman (32:19.480)
is the trolley problem formulation, right?
Chris Urmson (32:23.080)
Which has, let's not get into that too much
Lex Fridman (32:27.000)
but is misguided in many ways.
Lex Fridman (32:29.600)
But it speaks to the fact that people are grappling
Lex Fridman (32:32.320)
with this idea of giving control over to a machine.
Lex Fridman (32:36.160)
So how do you win the hearts and minds of the people
Lex Fridman (32:41.560)
that autonomy is something that could be a part
Lex Fridman (32:44.600)
of their lives?
Lex Fridman (32:45.520)
I think you let them experience it, right?
Chris Urmson (32:47.640)
I think it's right.
Lex Fridman (32:50.440)
I think people should be skeptical.
Chris Urmson (32:52.800)
I think people should ask questions.
Lex Fridman (32:55.680)
I think they should doubt
Chris Urmson (32:57.000)
because this is something new and different.
Lex Fridman (33:00.120)
They haven't touched it yet.
Lex Fridman (33:01.880)
And I think that's perfectly reasonable.
Lex Fridman (33:03.640)
And, but at the same time,
Chris Urmson (33:07.320)
it's clear there's an opportunity to make the road safer.
Lex Fridman (33:09.320)
It's clear that we can improve access to mobility.
Chris Urmson (33:12.440)
It's clear that we can reduce the cost of mobility.
Lex Fridman (33:16.640)
And that once people try that
Lex Fridman (33:19.480)
and understand that it's safe
Lex Fridman (33:22.720)
and are able to use in their daily lives,
Chris Urmson (33:24.440)
I think it's one of these things
Lex Fridman (33:25.280)
that will just be obvious.
Lex Fridman (33:28.040)
And I've seen this practically in demonstrations
Lex Fridman (33:32.240)
that I've given where I've had people come in
Lex Fridman (33:35.560)
and they're very skeptical.
Lex Fridman (33:38.840)
Again, in a vehicle, my favorite one
Chris Urmson (33:40.440)
is taking somebody out on the freeway
Lex Fridman (33:42.560)
and we're on the 101 driving at 65 miles an hour.
Lex Fridman (33:46.000)
And after 10 minutes, they kind of turn and ask,
Lex Fridman (33:48.400)
is that all it does?
Lex Fridman (33:49.480)
And you're like, it's a self driving car.
Lex Fridman (33:52.080)
I'm not sure exactly what you thought it would do, right?
Lex Fridman (33:54.840)
But it becomes mundane,
Lex Fridman (33:58.840)
which is exactly what you want a technology
Lex Fridman (34:01.480)
like this to be, right?
Lex Fridman (34:02.720)
We don't really, when I turn the light switch on in here,
Chris Urmson (34:07.280)
I don't think about the complexity of those electrons
Lex Fridman (34:12.000)
being pushed down a wire from wherever it was
Lex Fridman (34:14.200)
and being generated.
Lex Fridman (34:15.240)
It's like, I just get annoyed if it doesn't work, right?
Lex Fridman (34:19.080)
And what I value is the fact
Lex Fridman (34:21.400)
that I can do other things in this space.
Chris Urmson (34:23.080)
I can see my colleagues.
Lex Fridman (34:24.560)
I can read stuff on a paper.
Chris Urmson (34:26.160)
I can not be afraid of the dark.
Lex Fridman (34:30.360)
And I think that's what we want this technology to be like
Chris Urmson (34:33.320)
is it's in the background
Lex Fridman (34:34.640)
and people get to have those life experiences
Lex Fridman (34:37.120)
and do so safely.
Lex Fridman (34:38.440)
So putting this technology in the hands of people
Lex Fridman (34:42.160)
speaks to scale of deployment, right?
Lex Fridman (34:46.320)
So what do you think the dreaded question about the future
Chris Urmson (34:50.880)
because nobody can predict the future,
Lex Fridman (34:53.560)
but just maybe speak poetically
Chris Urmson (34:57.240)
about when do you think we'll see a large scale deployment
Lex Fridman (35:00.880)
of autonomous vehicles, 10,000, those kinds of numbers?
Chris Urmson (35:06.680)
We'll see that within 10 years.
Lex Fridman (35:09.240)
I'm pretty confident.
Lex Fridman (35:14.040)
What's an impressive scale?
Lex Fridman (35:16.040)
What moment, so you've done the DARPA challenge
Chris Urmson (35:19.200)
where there's one vehicle.
Lex Fridman (35:20.440)
At which moment does it become, wow, this is serious scale?
Lex Fridman (35:23.960)
So I think the moment it gets serious
Lex Fridman (35:26.520)
is when we really do have a driverless vehicle
Chris Urmson (35:32.240)
operating on public roads
Lex Fridman (35:35.000)
and that we can do that kind of continuously.
Chris Urmson (35:37.960)
Without a safety driver.
Lex Fridman (35:38.880)
Without a safety driver in the vehicle.
Chris Urmson (35:40.440)
I think at that moment,
Lex Fridman (35:41.560)
we've kind of crossed the zero to one threshold.
Lex Fridman (35:45.920)
And then it is about how do we continue to scale that?
Lex Fridman (35:50.200)
How do we build the right business models?
Lex Fridman (35:53.960)
How do we build the right customer experience around it
Lex Fridman (35:56.320)
so that it is actually a useful product out in the world?
Lex Fridman (36:00.960)
And I think that is really,
Lex Fridman (36:03.600)
at that point it moves from
Lex Fridman (36:05.920)
what is this kind of mixed science engineering project
Lex Fridman (36:09.200)
into engineering and commercialization
Lex Fridman (36:12.360)
and really starting to deliver on the value
Lex Fridman (36:15.840)
that we all see here and actually making that real in the world.
Lex Fridman (36:20.680)
What do you think that deployment looks like?
Lex Fridman (36:22.240)
Where do we first see the inkling of no safety driver,
Lex Fridman (36:26.440)
one or two cars here and there?
Lex Fridman (36:28.600)
Is it on the highway?
Lex Fridman (36:29.800)
Is it in specific routes in the urban environment?
Lex Fridman (36:33.160)
I think it's gonna be urban, suburban type environments.
Chris Urmson (36:37.880)
Yeah, with Aurora, when we thought about how to tackle this,
Lex Fridman (36:41.560)
it was kind of in vogue to think about trucking
Chris Urmson (36:46.040)
as opposed to urban driving.
Lex Fridman (36:47.800)
And again, the human intuition around this
Chris Urmson (36:51.280)
is that freeways are easier to drive on
Lex Fridman (36:57.080)
because everybody's kind of going in the same direction
Lex Fridman (36:59.280)
and lanes are a little wider, et cetera.
Lex Fridman (37:01.560)
And I think that that intuition is pretty good,
Chris Urmson (37:03.320)
except we don't really care about most of the time.
Lex Fridman (37:06.040)
We care about all of the time.
Lex Fridman (37:08.400)
And when you're driving on a freeway with a truck,
Lex Fridman (37:10.880)
say 70 miles an hour,
Lex Fridman (37:14.600)
and you've got 70,000 pound load with you,
Lex Fridman (37:16.240)
that's just an incredible amount of kinetic energy.
Lex Fridman (37:18.880)
And so when that goes wrong, it goes really wrong.
Lex Fridman (37:22.640)
And those challenges that you see occur more rarely,
Lex Fridman (37:27.800)
so you don't get to learn as quickly.
Lex Fridman (37:31.120)
And they're incrementally more difficult than urban driving,
Lex Fridman (37:34.720)
but they're not easier than urban driving.
Lex Fridman (37:37.440)
And so I think this happens in moderate speed
Chris Urmson (37:41.640)
urban environments because if two vehicles crash
Lex Fridman (37:45.280)
at 25 miles per hour, it's not good,
Lex Fridman (37:48.120)
but probably everybody walks away.
Lex Fridman (37:51.080)
And those events where there's the possibility
Chris Urmson (37:53.720)
for that occurring happen frequently.
Lex Fridman (37:55.800)
So we get to learn more rapidly.
Chris Urmson (37:58.000)
We get to do that with lower risk for everyone.
Lex Fridman (38:02.520)
And then we can deliver value to people
Chris Urmson (38:04.360)
that need to get from one place to another.
Lex Fridman (38:05.880)
And once we've got that solved,
Chris Urmson (38:08.160)
then the freeway driving part of this just falls out.
Lex Fridman (38:11.320)
But we're able to learn more safely,
Chris Urmson (38:13.080)
more quickly in the urban environment.
Lex Fridman (38:15.200)
So 10 years and then scale 20, 30 year,
Chris Urmson (38:18.760)
who knows if a sufficiently compelling experience
Lex Fridman (38:22.040)
is created, it could be faster and slower.
Lex Fridman (38:24.400)
Do you think there could be breakthroughs
Lex Fridman (38:27.160)
and what kind of breakthroughs might there be
Lex Fridman (38:29.920)
that completely change that timeline?
Lex Fridman (38:32.400)
Again, not only am I asking you to predict the future,
Chris Urmson (38:35.360)
I'm asking you to predict breakthroughs
Lex Fridman (38:37.360)
that haven't happened yet.
Lex Fridman (38:38.360)
So what's the, I think another way to ask that
Lex Fridman (38:41.440)
would be if I could wave a magic wand,
Lex Fridman (38:44.320)
what part of the system would I make work today
Lex Fridman (38:46.720)
to accelerate it as quickly as possible?
Chris Urmson (38:52.120)
Don't say infrastructure, please don't say infrastructure.
Lex Fridman (38:54.200)
No, it's definitely not infrastructure.
Chris Urmson (38:56.320)
It's really that perception forecasting capability.
Lex Fridman (39:00.600)
So if tomorrow you could give me a perfect model
Chris Urmson (39:04.840)
of what's happened, what is happening
Lex Fridman (39:06.960)
and what will happen for the next five seconds
Chris Urmson (39:10.360)
around a vehicle on the roadway,
Lex Fridman (39:13.040)
that would accelerate things pretty dramatically.
Chris Urmson (39:15.360)
Are you, in terms of staying up at night,
Lex Fridman (39:17.600)
are you mostly bothered by cars, pedestrians or cyclists?
Lex Fridman (39:21.760)
So I worry most about the vulnerable road users
Lex Fridman (39:25.960)
about the combination of cyclists and cars, right?
Chris Urmson (39:28.480)
Or cyclists and pedestrians because they're not in armor.
Lex Fridman (39:31.960)
The cars, they're bigger, they've got protection
Chris Urmson (39:36.480)
for the people and so the ultimate risk is lower there.
Lex Fridman (39:41.080)
Whereas a pedestrian or a cyclist,
Chris Urmson (39:43.240)
they're out on the road and they don't have any protection
Lex Fridman (39:46.480)
and so we need to pay extra attention to that.
Lex Fridman (39:49.720)
Do you think about a very difficult technical challenge
Lex Fridman (39:55.720)
of the fact that pedestrians,
Chris Urmson (39:58.520)
if you try to protect pedestrians
Lex Fridman (40:00.240)
by being careful and slow, they'll take advantage of that.
Lex Fridman (40:04.560)
So the game theoretic dance, does that worry you
Lex Fridman (40:09.040)
of how, from a technical perspective, how we solve that?
Chris Urmson (40:12.480)
Because as humans, the way we solve that
Lex Fridman (40:14.560)
is kind of nudge our way through the pedestrians
Chris Urmson (40:17.240)
which doesn't feel, from a technical perspective,
Lex Fridman (40:20.000)
as a appropriate algorithm.
Lex Fridman (40:23.200)
But do you think about how we solve that problem?
Lex Fridman (40:25.920)
Yeah, I think there's two different concepts there.
Lex Fridman (40:31.360)
So one is, am I worried that because these vehicles
Lex Fridman (40:35.820)
are self driving, people will kind of step in the road
Lex Fridman (40:37.600)
and take advantage of them?
Lex Fridman (40:38.640)
And I've heard this and I don't really believe it
Chris Urmson (40:43.760)
because if I'm driving down the road
Lex Fridman (40:45.960)
and somebody steps in front of me, I'm going to stop.
Chris Urmson (40:50.600)
Even if I'm annoyed, I'm not gonna just drive
Lex Fridman (40:53.660)
through a person stood in the road.
Lex Fridman (40:56.400)
And so I think today people can take advantage of this
Lex Fridman (41:00.400)
and you do see some people do it.
Chris Urmson (41:02.560)
I guess there's an incremental risk
Lex Fridman (41:04.180)
because maybe they have lower confidence
Chris Urmson (41:05.880)
that I'm gonna see them than they might have
Lex Fridman (41:07.720)
for an automated vehicle and so maybe that shifts
Chris Urmson (41:10.400)
it a little bit.
Lex Fridman (41:12.040)
But I think people don't wanna get hit by cars.
Lex Fridman (41:14.360)
And so I think that I'm not that worried
Lex Fridman (41:17.080)
about people walking out of the 101
Lex Fridman (41:18.760)
and creating chaos more than they would today.
Lex Fridman (41:24.400)
Regarding kind of the nudging through a big stream
Chris Urmson (41:27.040)
of pedestrians leaving a concert or something,
Lex Fridman (41:30.040)
I think that is further down the technology pipeline.
Chris Urmson (41:33.520)
I think that you're right, that's tricky.
Lex Fridman (41:36.960)
I don't think it's necessarily,
Chris Urmson (41:40.360)
I think the algorithm people use for this is pretty simple.
Lex Fridman (41:43.600)
It's kind of just move forward slowly
Lex Fridman (41:44.800)
and if somebody's really close then stop.
Lex Fridman (41:46.800)
And I think that that probably can be replicated
Chris Urmson (41:50.880)
pretty easily and particularly given that
Lex Fridman (41:54.040)
you don't do this at 30 miles an hour,
Chris Urmson (41:55.720)
you do it at one, that even in those situations
Lex Fridman (41:59.080)
the risk is relatively minimal.
Lex Fridman (42:01.200)
But it's not something we're thinking about
Lex Fridman (42:03.640)
in any serious way.
Lex Fridman (42:04.560)
And probably that's less an algorithm problem
Lex Fridman (42:07.920)
and more creating a human experience.
Lex Fridman (42:10.160)
So the HCI people that create a visual display
Lex Fridman (42:14.300)
that you're pleasantly as a pedestrian
Chris Urmson (42:16.260)
nudged out of the way, that's an experience problem,
Lex Fridman (42:20.760)
not an algorithm problem.
Lex Fridman (42:22.880)
Who's the main competitor to Aurora today?
Lex Fridman (42:25.480)
And how do you outcompete them in the long run?
Lex Fridman (42:28.640)
So we really focus a lot on what we're doing here.
Lex Fridman (42:31.200)
I think that, I've said this a few times,
Chris Urmson (42:34.480)
that this is a huge difficult problem
Lex Fridman (42:37.960)
and it's great that a bunch of companies are tackling it
Chris Urmson (42:40.320)
because I think it's so important for society
Lex Fridman (42:42.320)
that somebody gets there.
Lex Fridman (42:43.800)
So we don't spend a whole lot of time
Lex Fridman (42:49.120)
thinking tactically about who's out there
Lex Fridman (42:51.600)
and how do we beat that person individually.
Lex Fridman (42:55.240)
What are we trying to do to go faster ultimately?
Chris Urmson (42:59.760)
Well part of it is the leadership team we have
Lex Fridman (43:02.640)
has got pretty tremendous experience.
Lex Fridman (43:04.200)
And so we kind of understand the landscape
Lex Fridman (43:06.440)
and understand where the cul de sacs are to some degree
Lex Fridman (43:09.160)
and we try and avoid those.
Lex Fridman (43:10.980)
I think there's a part of it,
Chris Urmson (43:14.260)
just this great team we've built.
Lex Fridman (43:16.260)
People, this is a technology and a company
Chris Urmson (43:19.080)
that people believe in the mission of
Lex Fridman (43:22.320)
and so it allows us to attract
Chris Urmson (43:23.740)
just awesome people to go work.
Lex Fridman (43:26.800)
We've got a culture I think that people appreciate
Chris Urmson (43:29.320)
that allows them to focus,
Lex Fridman (43:30.460)
allows them to really spend time solving problems.
Lex Fridman (43:33.120)
And I think that keeps them energized.
Lex Fridman (43:35.900)
And then we've invested hard,
Chris Urmson (43:38.940)
invested heavily in the infrastructure
Lex Fridman (43:43.500)
and architectures that we think will ultimately accelerate us.
Lex Fridman (43:46.540)
So because of the folks we're able to bring in early on,
Lex Fridman (43:50.660)
because of the great investors we have,
Chris Urmson (43:53.540)
we don't spend all of our time doing demos
Lex Fridman (43:56.780)
and kind of leaping from one demo to the next.
Chris Urmson (43:58.660)
We've been given the freedom to invest in
Lex Fridman (44:03.940)
infrastructure to do machine learning,
Chris Urmson (44:05.500)
infrastructure to pull data from our on road testing,
Lex Fridman (44:08.600)
infrastructure to use that to accelerate engineering.
Lex Fridman (44:11.500)
And I think that early investment
Lex Fridman (44:14.480)
and continuing investment in those kind of tools
Chris Urmson (44:17.340)
will ultimately allow us to accelerate
Lex Fridman (44:19.780)
and do something pretty incredible.
Chris Urmson (44:21.940)
Chris, beautifully put.
Lex Fridman (44:23.420)
It's a good place to end.
Chris Urmson (44:24.660)
Thank you so much for talking today.
Lex Fridman (44:26.500)
Thank you very much. Really enjoyed it.
🔗 相关节目