Hugging Face's logo

Datasets:
shubhamg2208
/
lexicap

Tasks:
Text Classification
Text Generation
Sub-tasks:
sentiment-analysis
dialogue-modeling
language-modeling
Languages:
English
Multilinguality:
monolingual
Size Categories:
n<1K
Language Creators:
found
Source Datasets:
original
Tags:
karpathy,whisper,openai
Dataset card Files Community
lexicap / vtt /episode_028_large.vtt
Shubham Gupta
Add files to lfs
323a418
raw history blame contribute delete
No virus
60.9 kB
WEBVTT
00:00.000 --> 00:03.120
The following is a conversation with Chris Sampson.
00:03.120 --> 00:06.000
He was a CTO of the Google self driving car team,
00:06.000 --> 00:08.880
a key engineer and leader behind the Carnegie Mellon
00:08.880 --> 00:12.000
University autonomous vehicle entries in the DARPA Grand
00:12.000 --> 00:16.160
Challenges and the winner of the DARPA Urban Challenge.
00:16.160 --> 00:20.100
Today, he's the CEO of Aurora Innovation, an autonomous
00:20.100 --> 00:21.360
vehicle software company.
00:21.360 --> 00:23.600
He started with Sterling Anderson,
00:23.600 --> 00:25.960
who was the former director of Tesla Autopilot,
00:25.960 --> 00:30.120
and drew back now, Uber's former autonomy and perception lead.
00:30.120 --> 00:32.880
Chris is one of the top roboticists and autonomous
00:32.880 --> 00:36.320
vehicle experts in the world, and a longtime voice
00:36.320 --> 00:38.840
of reason in a space that is shrouded
00:38.840 --> 00:41.320
in both mystery and hype.
00:41.320 --> 00:43.600
He both acknowledges the incredible challenges
00:43.600 --> 00:46.480
involved in solving the problem of autonomous driving
00:46.480 --> 00:49.760
and is working hard to solve it.
00:49.760 --> 00:52.400
This is the Artificial Intelligence podcast.
00:52.400 --> 00:54.720
If you enjoy it, subscribe on YouTube,
00:54.720 --> 00:57.920
give it five stars on iTunes, support it on Patreon,
00:57.920 --> 00:59.720
or simply connect with me on Twitter
00:59.720 --> 01:03.240
at Lex Friedman, spelled F R I D M A N.
01:03.240 --> 01:09.120
And now, here's my conversation with Chris Sampson.
01:09.120 --> 01:11.960
You were part of both the DARPA Grand Challenge
01:11.960 --> 01:13.880
and the DARPA Urban Challenge teams
01:13.880 --> 01:17.040
at CMU with Red Whitaker.
01:17.040 --> 01:19.720
What technical or philosophical things
01:19.720 --> 01:22.240
have you learned from these races?
01:22.240 --> 01:26.600
I think the high order bit was that it could be done.
01:26.600 --> 01:30.200
I think that was the thing that was
01:30.200 --> 01:34.880
incredible about the first of the Grand Challenges,
01:34.880 --> 01:38.160
that I remember I was a grad student at Carnegie Mellon,
01:38.160 --> 01:45.360
and there was kind of this dichotomy of it
01:45.360 --> 01:46.720
seemed really hard, so that would
01:46.720 --> 01:48.800
be cool and interesting.
01:48.800 --> 01:52.800
But at the time, we were the only robotics institute around,
01:52.800 --> 01:55.560
and so if we went into it and fell on our faces,
01:55.560 --> 01:58.360
that would be embarrassing.
01:58.360 --> 02:01.120
So I think just having the will to go do it,
02:01.120 --> 02:02.880
to try to do this thing that at the time
02:02.880 --> 02:05.000
was marked as darn near impossible,
02:05.000 --> 02:06.960
and then after a couple of tries,
02:06.960 --> 02:08.420
be able to actually make it happen,
02:08.420 --> 02:12.320
I think that was really exciting.
02:12.320 --> 02:15.040
But at which point did you believe it was possible?
02:15.040 --> 02:16.960
Did you from the very beginning?
02:16.960 --> 02:18.000
Did you personally?
02:18.000 --> 02:19.800
Because you're one of the lead engineers.
02:19.800 --> 02:21.800
You actually had to do a lot of the work.
02:21.800 --> 02:23.880
Yeah, I was the technical director there,
02:23.880 --> 02:26.120
and did a lot of the work, along with a bunch
02:26.120 --> 02:28.420
of other really good people.
02:28.420 --> 02:29.760
Did I believe it could be done?
02:29.760 --> 02:31.080
Yeah, of course.
02:31.080 --> 02:32.760
Why would you go do something you thought
02:32.760 --> 02:34.800
was completely impossible?
02:34.800 --> 02:36.260
We thought it was going to be hard.
02:36.260 --> 02:37.800
We didn't know how we were going to be able to do it.
02:37.800 --> 02:42.880
We didn't know if we'd be able to do it the first time.
02:42.880 --> 02:45.960
Turns out we couldn't.
02:45.960 --> 02:48.400
That, yeah, I guess you have to.
02:48.400 --> 02:52.960
I think there's a certain benefit to naivete, right?
02:52.960 --> 02:55.440
That if you don't know how hard something really is,
02:55.440 --> 02:59.600
you try different things, and it gives you an opportunity
02:59.600 --> 03:04.120
that others who are wiser maybe don't have.
03:04.120 --> 03:05.720
What were the biggest pain points?
03:05.720 --> 03:08.880
Mechanical, sensors, hardware, software,
03:08.880 --> 03:11.800
algorithms for mapping, localization,
03:11.800 --> 03:13.680
just general perception, control?
03:13.680 --> 03:15.320
Like hardware, software, first of all?
03:15.320 --> 03:20.120
I think that's the joy of this field, is that it's all hard
03:20.120 --> 03:25.360
and that you have to be good at each part of it.
03:25.360 --> 03:32.360
So for the urban challenges, if I look back at it from today,
03:32.360 --> 03:38.960
it should be easy today, that it was a static world.
03:38.960 --> 03:40.800
There weren't other actors moving through it,
03:40.800 --> 03:42.480
is what that means.
03:42.480 --> 03:47.080
It was out in the desert, so you get really good GPS.
03:47.080 --> 03:51.400
So that went, and we could map it roughly.
03:51.400 --> 03:55.160
And so in retrospect now, it's within the realm of things
03:55.160 --> 03:57.840
we could do back then.
03:57.840 --> 03:59.720
Just actually getting the vehicle and the,
03:59.720 --> 04:00.680
there's a bunch of engineering work
04:00.680 --> 04:04.760
to get the vehicle so that we could control it and drive it.
04:04.760 --> 04:09.600
That's still a pain today, but it was even more so back then.
04:09.600 --> 04:14.280
And then the uncertainty of exactly what they wanted us to do
04:14.280 --> 04:17.040
was part of the challenge as well.
04:17.040 --> 04:19.440
Right, you didn't actually know the track heading in here.
04:19.440 --> 04:21.480
You knew approximately, but you didn't actually
04:21.480 --> 04:23.520
know the route that was going to be taken.
04:23.520 --> 04:24.920
That's right, we didn't know the route.
04:24.920 --> 04:28.600
We didn't even really, the way the rules had been described,
04:28.600 --> 04:29.800
you had to kind of guess.
04:29.800 --> 04:33.360
So if you think back to that challenge,
04:33.360 --> 04:36.960
the idea was that the government would give us,
04:36.960 --> 04:40.320
the DARPA would give us a set of waypoints
04:40.320 --> 04:43.520
and kind of the width that you had to stay within
04:43.520 --> 04:46.800
between the line that went between each of those waypoints.
04:46.800 --> 04:49.280
And so the most devious thing they could have done
04:49.280 --> 04:53.280
is set a kilometer wide corridor across a field
04:53.280 --> 04:58.280
of scrub brush and rocks and said, go figure it out.
04:58.520 --> 05:01.920
Fortunately, it really, it turned into basically driving
05:01.920 --> 05:05.000
along a set of trails, which is much more relevant
05:05.000 --> 05:07.920
to the application they were looking for.
05:08.760 --> 05:12.080
But no, it was a hell of a thing back in the day.
05:12.080 --> 05:16.640
So the legend, Red, was kind of leading that effort
05:16.640 --> 05:19.120
in terms of just broadly speaking.
05:19.120 --> 05:22.040
So you're a leader now.
05:22.040 --> 05:25.000
What have you learned from Red about leadership?
05:25.000 --> 05:26.200
I think there's a couple things.
05:26.200 --> 05:31.080
One is go and try those really hard things.
05:31.080 --> 05:34.480
That's where there is an incredible opportunity.
05:34.480 --> 05:36.560
I think the other big one, though,
05:36.560 --> 05:40.680
is to see people for who they can be, not who they are.
05:41.720 --> 05:43.720
It's one of the things that I actually,
05:43.720 --> 05:46.080
one of the deepest lessons I learned from Red
05:46.080 --> 05:50.200
was that he would look at undergraduates
05:50.200 --> 05:55.200
or graduate students and empower them to be leaders,
05:56.120 --> 06:00.320
to have responsibility, to do great things
06:00.320 --> 06:04.480
that I think another person might look at them
06:04.480 --> 06:06.600
and think, oh, well, that's just an undergraduate student.
06:06.600 --> 06:07.720
What could they know?
06:08.680 --> 06:12.720
And so I think that kind of trust but verify,
06:12.720 --> 06:14.480
have confidence in what people can become,
06:14.480 --> 06:16.680
I think is a really powerful thing.
06:16.680 --> 06:20.440
So through that, let's just fast forward through the history.
06:20.440 --> 06:24.160
Can you maybe talk through the technical evolution
06:24.160 --> 06:26.200
of autonomous vehicle systems
06:26.200 --> 06:29.960
from the first two Grand Challenges to the Urban Challenge
06:29.960 --> 06:33.560
to today, are there major shifts in your mind
06:33.560 --> 06:37.240
or is it the same kind of technology just made more robust?
06:37.240 --> 06:39.840
I think there's been some big, big steps.
06:40.880 --> 06:43.720
So for the Grand Challenge,
06:43.720 --> 06:48.720
the real technology that unlocked that was HD mapping.
06:51.400 --> 06:54.200
Prior to that, a lot of the off road robotics work
06:55.160 --> 06:58.480
had been done without any real prior model
06:58.480 --> 07:01.400
of what the vehicle was going to encounter.
07:01.400 --> 07:04.880
And so that innovation that the fact that we could get
07:05.960 --> 07:10.960
decimeter resolution models was really a big deal.
07:13.440 --> 07:18.200
And that allowed us to kind of bound the complexity
07:18.200 --> 07:19.680
of the driving problem the vehicle had
07:19.680 --> 07:21.040
and allowed it to operate at speed
07:21.040 --> 07:23.800
because we could assume things about the environment
07:23.800 --> 07:25.360
that it was going to encounter.
07:25.360 --> 07:29.720
So that was the big step there.
07:31.280 --> 07:35.280
For the Urban Challenge,
07:37.240 --> 07:39.280
one of the big technological innovations there
07:39.280 --> 07:41.040
was the multi beam LIDAR
07:41.960 --> 07:45.760
and being able to generate high resolution,
07:45.760 --> 07:48.680
mid to long range 3D models of the world
07:48.680 --> 07:53.680
and use that for understanding the world around the vehicle.
07:53.680 --> 07:56.600
And that was really kind of a game changing technology.
07:58.600 --> 08:00.000
In parallel with that,
08:00.000 --> 08:04.360
we saw a bunch of other technologies
08:04.360 --> 08:06.120
that had been kind of converging
08:06.120 --> 08:08.440
half their day in the sun.
08:08.440 --> 08:12.560
So Bayesian estimation had been,
08:12.560 --> 08:17.560
SLAM had been a big field in robotics.
08:17.840 --> 08:20.760
You would go to a conference a couple of years before that
08:20.760 --> 08:24.880
and every paper would effectively have SLAM somewhere in it.
08:24.880 --> 08:29.320
And so seeing that the Bayesian estimation techniques
08:30.720 --> 08:33.400
play out on a very visible stage,
08:33.400 --> 08:36.520
I thought that was pretty exciting to see.
08:38.080 --> 08:41.560
And mostly SLAM was done based on LIDAR at that time.
08:41.560 --> 08:44.560
Yeah, and in fact, we weren't really doing SLAM per se
08:45.600 --> 08:47.480
in real time because we had a model ahead of time,
08:47.480 --> 08:51.040
we had a roadmap, but we were doing localization.
08:51.040 --> 08:53.560
And we were using the LIDAR or the cameras
08:53.560 --> 08:55.400
depending on who exactly was doing it
08:55.400 --> 08:57.560
to localize to a model of the world.
08:57.560 --> 09:00.160
And I thought that was a big step
09:00.160 --> 09:05.160
from kind of naively trusting GPS, INS before that.
09:06.640 --> 09:09.840
And again, lots of work had been going on in this field.
09:09.840 --> 09:13.040
Certainly this was not doing anything
09:13.040 --> 09:16.840
particularly innovative in SLAM or in localization,
09:16.840 --> 09:20.200
but it was seeing that technology necessary
09:20.200 --> 09:21.800
in a real application on a big stage,
09:21.800 --> 09:23.080
I thought was very cool.
09:23.080 --> 09:24.000
So for the urban challenge,
09:24.000 --> 09:28.600
those are already maps constructed offline in general.
09:28.600 --> 09:30.920
And did people do that individually,
09:30.920 --> 09:33.600
did individual teams do it individually
09:33.600 --> 09:36.440
so they had their own different approaches there
09:36.440 --> 09:41.440
or did everybody kind of share that information
09:41.720 --> 09:42.880
at least intuitively?
09:42.880 --> 09:47.880
So DARPA gave all the teams a model of the world, a map.
09:49.640 --> 09:53.240
And then one of the things that we had to figure out
09:53.240 --> 09:56.080
back then was, and it's still one of these things
09:56.080 --> 09:57.280
that trips people up today
09:57.280 --> 10:00.280
is actually the coordinate system.
10:00.280 --> 10:03.080
So you get a latitude longitude
10:03.080 --> 10:05.040
and to so many decimal places,
10:05.040 --> 10:07.360
you don't really care about kind of the ellipsoid
10:07.360 --> 10:09.560
of the earth that's being used.
10:09.560 --> 10:12.240
But when you want to get to 10 centimeter
10:12.240 --> 10:14.400
or centimeter resolution,
10:14.400 --> 10:18.520
you care whether the coordinate system is NADS 83
10:18.520 --> 10:23.520
or WGS 84 or these are different ways to describe
10:24.200 --> 10:26.760
both the kind of non sphericalness of the earth,
10:26.760 --> 10:31.080
but also kind of the, I think,
10:31.080 --> 10:32.080
I can't remember which one,
10:32.080 --> 10:33.600
the tectonic shifts that are happening
10:33.600 --> 10:37.000
and how to transform the global datum as a function of that.
10:37.000 --> 10:41.020
So getting a map and then actually matching it to reality
10:41.020 --> 10:42.880
to centimeter resolution, that was kind of interesting
10:42.880 --> 10:44.040
and fun back then.
10:44.040 --> 10:46.760
So how much work was the perception doing there?
10:46.760 --> 10:51.760
So how much were you relying on localization based on maps
10:52.480 --> 10:55.760
without using perception to register to the maps?
10:55.760 --> 10:58.000
And I guess the question is how advanced
10:58.000 --> 10:59.800
was perception at that point?
10:59.800 --> 11:01.960
It's certainly behind where we are today, right?
11:01.960 --> 11:05.840
We're more than a decade since the urban challenge.
11:05.840 --> 11:08.640
But the core of it was there.
11:08.640 --> 11:13.120
That we were tracking vehicles.
11:13.120 --> 11:15.640
We had to do that at 100 plus meter range
11:15.640 --> 11:18.320
because we had to merge with other traffic.
11:18.320 --> 11:21.240
We were using, again, Bayesian estimates
11:21.240 --> 11:23.860
for state of these vehicles.
11:23.860 --> 11:25.580
We had to deal with a bunch of the problems
11:25.580 --> 11:26.920
that you think of today,
11:26.920 --> 11:29.820
of predicting where that vehicle's going to be
11:29.820 --> 11:31.060
a few seconds into the future.
11:31.060 --> 11:32.380
We had to deal with the fact
11:32.380 --> 11:35.320
that there were multiple hypotheses for that
11:35.320 --> 11:37.660
because a vehicle at an intersection might be going right
11:37.660 --> 11:38.780
or it might be going straight
11:38.780 --> 11:40.620
or it might be making a left turn.
11:41.500 --> 11:44.120
And we had to deal with the challenge of the fact
11:44.120 --> 11:47.600
that our behavior was going to impact the behavior
11:47.600 --> 11:48.960
of that other operator.
11:48.960 --> 11:53.480
And we did a lot of that in relatively naive ways,
11:53.480 --> 11:54.820
but it kind of worked.
11:54.820 --> 11:57.080
Still had to have some kind of solution.
11:57.080 --> 11:59.960
And so where does that, 10 years later,
11:59.960 --> 12:01.520
where does that take us today
12:01.520 --> 12:04.260
from that artificial city construction
12:04.260 --> 12:07.000
to real cities to the urban environment?
12:07.000 --> 12:09.160
Yeah, I think the biggest thing
12:09.160 --> 12:14.160
is that the actors are truly unpredictable.
12:15.720 --> 12:18.800
That most of the time, the drivers on the road,
12:18.800 --> 12:23.800
the other road users are out there behaving well,
12:24.080 --> 12:25.880
but every once in a while they're not.
12:27.080 --> 12:32.080
The variety of other vehicles is, you have all of them.
12:32.080 --> 12:35.840
In terms of behavior, in terms of perception, or both?
12:35.840 --> 12:36.680
Both.
12:38.740 --> 12:40.520
Back then we didn't have to deal with cyclists,
12:40.520 --> 12:42.800
we didn't have to deal with pedestrians,
12:42.800 --> 12:44.800
didn't have to deal with traffic lights.
12:46.260 --> 12:49.400
The scale over which that you have to operate is now
12:49.400 --> 12:51.120
is much larger than the air base
12:51.120 --> 12:52.720
that we were thinking about back then.
12:52.720 --> 12:55.420
So what, easy question,
12:56.280 --> 12:59.720
what do you think is the hardest part about driving?
12:59.720 --> 13:00.560
Easy question.
13:00.560 --> 13:02.560
Yeah, no, I'm joking.
13:02.560 --> 13:07.440
I'm sure nothing really jumps out at you as one thing,
13:07.440 --> 13:12.440
but in the jump from the urban challenge to the real world,
13:12.920 --> 13:15.320
is there something that's a particular,
13:15.320 --> 13:18.480
you foresee as very serious, difficult challenge?
13:18.480 --> 13:21.080
I think the most fundamental difference
13:21.080 --> 13:25.340
is that we're doing it for real.
13:26.760 --> 13:28.960
That in that environment,
13:28.960 --> 13:31.880
it was both a limited complexity environment
13:31.880 --> 13:33.240
because certain actors weren't there,
13:33.240 --> 13:35.380
because the roads were maintained,
13:35.380 --> 13:37.360
there were barriers keeping people separate
13:37.360 --> 13:39.400
from robots at the time,
13:40.840 --> 13:43.300
and it only had to work for 60 miles.
13:43.300 --> 13:46.160
Which, looking at it from 2006,
13:46.160 --> 13:48.960
it had to work for 60 miles, right?
13:48.960 --> 13:50.940
Looking at it from now,
13:51.880 --> 13:53.720
we want things that will go and drive
13:53.720 --> 13:57.160
for half a million miles,
13:57.160 --> 14:00.020
and it's just a different game.
14:00.940 --> 14:03.480
So how important,
14:03.480 --> 14:06.080
you said LiDAR came into the game early on,
14:06.080 --> 14:07.880
and it's really the primary driver
14:07.880 --> 14:10.240
of autonomous vehicles today as a sensor.
14:10.240 --> 14:11.920
So how important is the role of LiDAR
14:11.920 --> 14:14.800
in the sensor suite in the near term?
14:14.800 --> 14:16.740
So I think it's essential.
14:17.920 --> 14:20.480
I believe, but I also believe that cameras are essential,
14:20.480 --> 14:22.120
and I believe the radar is essential.
14:22.120 --> 14:26.280
I think that you really need to use
14:26.280 --> 14:28.720
the composition of data from these different sensors
14:28.720 --> 14:32.640
if you want the thing to really be robust.
14:32.640 --> 14:34.360
The question I wanna ask,
14:34.360 --> 14:35.600
let's see if we can untangle it,
14:35.600 --> 14:39.320
is what are your thoughts on the Elon Musk
14:39.320 --> 14:42.340
provocative statement that LiDAR is a crutch,
14:42.340 --> 14:47.340
that it's a kind of, I guess, growing pains,
14:47.760 --> 14:49.920
and that much of the perception task
14:49.920 --> 14:52.120
can be done with cameras?
14:52.120 --> 14:55.440
So I think it is undeniable
14:55.440 --> 14:59.360
that people walk around without lasers in their foreheads,
14:59.360 --> 15:01.880
and they can get into vehicles and drive them,
15:01.880 --> 15:05.600
and so there's an existence proof
15:05.600 --> 15:09.600
that you can drive using passive vision.
15:10.880 --> 15:12.720
No doubt, can't argue with that.
15:12.720 --> 15:14.680
In terms of sensors, yeah, so there's proof.
15:14.680 --> 15:16.000
Yeah, in terms of sensors, right?
15:16.000 --> 15:20.200
So there's an example that we all go do it,
15:20.200 --> 15:21.380
many of us every day.
15:21.380 --> 15:26.380
In terms of LiDAR being a crutch, sure.
15:28.180 --> 15:33.100
But in the same way that the combustion engine
15:33.100 --> 15:35.260
was a crutch on the path to an electric vehicle,
15:35.260 --> 15:39.300
in the same way that any technology ultimately gets
15:40.840 --> 15:44.380
replaced by some superior technology in the future,
15:44.380 --> 15:47.740
and really the way that I look at this
15:47.740 --> 15:51.460
is that the way we get around on the ground,
15:51.460 --> 15:53.920
the way that we use transportation is broken,
15:55.280 --> 15:59.740
and that we have this, I think the number I saw this morning,
15:59.740 --> 16:04.060
37,000 Americans killed last year on our roads,
16:04.060 --> 16:05.380
and that's just not acceptable.
16:05.380 --> 16:09.460
And so any technology that we can bring to bear
16:09.460 --> 16:12.860
that accelerates this self driving technology
16:12.860 --> 16:14.640
coming to market and saving lives
16:14.640 --> 16:17.320
is technology we should be using.
16:18.280 --> 16:20.840
And it feels just arbitrary to say,
16:20.840 --> 16:25.840
well, I'm not okay with using lasers
16:26.240 --> 16:27.820
because that's whatever,
16:27.820 --> 16:30.720
but I am okay with using an eight megapixel camera
16:30.720 --> 16:32.880
or a 16 megapixel camera.
16:32.880 --> 16:34.640
These are just bits of technology,
16:34.640 --> 16:36.360
and we should be taking the best technology
16:36.360 --> 16:41.360
from the tool bin that allows us to go and solve a problem.
16:41.360 --> 16:45.160
The question I often talk to, well, obviously you do as well,
16:45.160 --> 16:48.280
to sort of automotive companies,
16:48.280 --> 16:51.360
and if there's one word that comes up more often
16:51.360 --> 16:55.280
than anything, it's cost, and trying to drive costs down.
16:55.280 --> 17:00.280
So while it's true that it's a tragic number, the 37,000,
17:01.400 --> 17:04.880
the question is, and I'm not the one asking this question
17:04.880 --> 17:05.820
because I hate this question,
17:05.820 --> 17:09.960
but we want to find the cheapest sensor suite
17:09.960 --> 17:13.280
that creates a safe vehicle.
17:13.280 --> 17:18.220
So in that uncomfortable trade off,
17:18.220 --> 17:23.220
do you foresee LiDAR coming down in cost in the future,
17:23.680 --> 17:26.680
or do you see a day where level four autonomy
17:26.680 --> 17:29.880
is possible without LiDAR?
17:29.880 --> 17:32.880
I see both of those, but it's really a matter of time.
17:32.880 --> 17:36.040
And I think really, maybe I would talk to the question
17:36.040 --> 17:37.840
you asked about the cheapest sensor.
17:37.840 --> 17:40.360
I don't think that's actually what you want.
17:40.360 --> 17:45.360
What you want is a sensor suite that is economically viable.
17:45.680 --> 17:49.440
And then after that, everything is about margin
17:49.440 --> 17:52.120
and driving costs out of the system.
17:52.120 --> 17:55.360
What you also want is a sensor suite that works.
17:55.360 --> 17:58.200
And so it's great to tell a story about
17:59.600 --> 18:03.260
how it would be better to have a self driving system
18:03.260 --> 18:08.040
with a $50 sensor instead of a $500 sensor.
18:08.040 --> 18:10.520
But if the $500 sensor makes it work
18:10.520 --> 18:14.760
and the $50 sensor doesn't work, who cares?
18:15.680 --> 18:20.020
So long as you can actually have an economic opportunity,
18:20.020 --> 18:21.520
there's an economic opportunity there.
18:21.520 --> 18:23.760
And the economic opportunity is important
18:23.760 --> 18:27.760
because that's how you actually have a sustainable business
18:27.760 --> 18:31.120
and that's how you can actually see this come to scale
18:31.120 --> 18:32.400
and be out in the world.
18:32.400 --> 18:34.780
And so when I look at LiDAR,
18:35.960 --> 18:38.880
I see a technology that has no underlying
18:38.880 --> 18:42.420
fundamentally expense to it, fundamental expense to it.
18:42.420 --> 18:46.080
It's going to be more expensive than an imager
18:46.080 --> 18:50.360
because CMOS processes or FAP processes
18:51.360 --> 18:55.080
are dramatically more scalable than mechanical processes.
18:56.200 --> 18:58.320
But we still should be able to drive costs down
18:58.320 --> 19:00.120
substantially on that side.
19:00.120 --> 19:04.840
And then I also do think that with the right business model
19:05.880 --> 19:07.560
you can absorb more,
19:07.560 --> 19:09.480
certainly more cost on the bill of materials.
19:09.480 --> 19:12.600
Yeah, if the sensor suite works, extra value is provided,
19:12.600 --> 19:15.480
thereby you don't need to drive costs down to zero.
19:15.480 --> 19:17.100
It's the basic economics.
19:17.100 --> 19:18.820
You've talked about your intuition
19:18.820 --> 19:22.200
that level two autonomy is problematic
19:22.200 --> 19:25.920
because of the human factor of vigilance,
19:25.920 --> 19:28.040
decrement, complacency, over trust and so on,
19:28.040 --> 19:29.600
just us being human.
19:29.600 --> 19:31.120
We over trust the system,
19:31.120 --> 19:34.240
we start doing even more so partaking
19:34.240 --> 19:37.180
in the secondary activities like smartphones and so on.
19:38.680 --> 19:43.000
Have your views evolved on this point in either direction?
19:43.000 --> 19:44.800
Can you speak to it?
19:44.800 --> 19:47.480
So, and I want to be really careful
19:47.480 --> 19:50.380
because sometimes this gets twisted in a way
19:50.380 --> 19:53.040
that I certainly didn't intend.
19:53.040 --> 19:58.040
So active safety systems are a really important technology
19:58.040 --> 20:00.680
that we should be pursuing and integrating into vehicles.
20:02.080 --> 20:04.280
And there's an opportunity in the near term
20:04.280 --> 20:06.520
to reduce accidents, reduce fatalities,
20:06.520 --> 20:10.320
and we should be pushing on that.
20:11.960 --> 20:14.680
Level two systems are systems
20:14.680 --> 20:18.080
where the vehicle is controlling two axes.
20:18.080 --> 20:21.720
So braking and throttle slash steering.
20:23.480 --> 20:25.680
And I think there are variants of level two systems
20:25.680 --> 20:27.280
that are supporting the driver.
20:27.280 --> 20:31.080
That absolutely we should encourage to be out there.
20:31.080 --> 20:32.880
Where I think there's a real challenge
20:32.880 --> 20:37.640
is in the human factors part around this
20:37.640 --> 20:41.240
and the misconception from the public
20:41.240 --> 20:43.600
around the capability set that that enables
20:43.600 --> 20:45.640
and the trust that they should have in it.
20:46.640 --> 20:50.000
And that is where I kind of,
20:50.000 --> 20:52.920
I'm actually incrementally more concerned
20:52.920 --> 20:54.440
around level three systems
20:54.440 --> 20:58.440
and how exactly a level two system is marketed and delivered
20:58.440 --> 21:01.840
and how much effort people have put into those human factors.
21:01.840 --> 21:05.640
So I still believe several things around this.
21:05.640 --> 21:09.440
One is people will overtrust the technology.
21:09.440 --> 21:11.440
We've seen over the last few weeks
21:11.440 --> 21:14.040
a spate of people sleeping in their Tesla.
21:14.920 --> 21:19.920
I watched an episode last night of Trevor Noah
21:19.920 --> 21:23.920
talking about this and him,
21:23.920 --> 21:26.720
this is a smart guy who has a lot of resources
21:26.720 --> 21:30.720
at his disposal describing a Tesla as a self driving car
21:30.720 --> 21:33.480
and that why shouldn't people be sleeping in their Tesla?
21:33.480 --> 21:36.560
And it's like, well, because it's not a self driving car
21:36.560 --> 21:38.840
and it is not intended to be
21:38.840 --> 21:43.840
and these people will almost certainly die at some point
21:46.400 --> 21:48.040
or hurt other people.
21:48.040 --> 21:50.080
And so we need to really be thoughtful
21:50.080 --> 21:51.840
about how that technology is described
21:51.840 --> 21:53.280
and brought to market.
21:54.240 --> 21:59.240
I also think that because of the economic challenges
21:59.240 --> 22:01.240
we were just talking about,
22:01.240 --> 22:05.160
that these level two driver assistance systems,
22:05.160 --> 22:07.280
that technology path will diverge
22:07.280 --> 22:10.200
from the technology path that we need to be on
22:10.200 --> 22:14.080
to actually deliver truly self driving vehicles,
22:14.080 --> 22:16.920
ones where you can get in it and drive it.
22:16.920 --> 22:20.800
Can get in it and sleep and have the equivalent
22:20.800 --> 22:24.680
or better safety than a human driver behind the wheel.
22:24.680 --> 22:27.520
Because again, the economics are very different
22:28.480 --> 22:30.880
in those two worlds and so that leads
22:30.880 --> 22:32.800
to divergent technology.
22:32.800 --> 22:34.680
So you just don't see the economics
22:34.680 --> 22:38.560
of gradually increasing from level two
22:38.560 --> 22:41.600
and doing so quickly enough
22:41.600 --> 22:44.480
to where it doesn't cause safety, critical safety concerns.
22:44.480 --> 22:47.680
You believe that it needs to diverge at this point
22:48.680 --> 22:50.800
into basically different routes.
22:50.800 --> 22:55.560
And really that comes back to what are those L2
22:55.560 --> 22:57.080
and L1 systems doing?
22:57.080 --> 22:59.840
And they are driver assistance functions
22:59.840 --> 23:04.400
where the people that are marketing that responsibly
23:04.400 --> 23:08.000
are being very clear and putting human factors in place
23:08.000 --> 23:12.440
such that the driver is actually responsible for the vehicle
23:12.440 --> 23:15.160
and that the technology is there to support the driver.
23:15.160 --> 23:19.880
And the safety cases that are built around those
23:19.880 --> 23:24.040
are dependent on that driver attention and attentiveness.
23:24.040 --> 23:28.000
And at that point, you can kind of give up
23:29.160 --> 23:31.240
to some degree for economic reasons,
23:31.240 --> 23:33.480
you can give up on say false negatives.
23:34.800 --> 23:36.200
And the way to think about this
23:36.200 --> 23:39.320
is for a four collision mitigation braking system,
23:39.320 --> 23:43.960
if it half the times the driver missed a vehicle
23:43.960 --> 23:46.080
in front of it, it hit the brakes
23:46.080 --> 23:47.680
and brought the vehicle to a stop,
23:47.680 --> 23:51.640
that would be an incredible, incredible advance
23:51.640 --> 23:53.040
in safety on our roads, right?
23:53.040 --> 23:55.000
That would be equivalent to seat belts.
23:55.000 --> 23:56.600
But it would mean that if that vehicle
23:56.600 --> 23:59.440
wasn't being monitored, it would hit one out of two cars.
24:00.600 --> 24:05.120
And so economically, that's a perfectly good solution
24:05.120 --> 24:06.280
for a driver assistance system.
24:06.280 --> 24:07.240
What you should do at that point,
24:07.240 --> 24:09.240
if you can get it to work 50% of the time,
24:09.240 --> 24:10.520
is drive the cost out of that
24:10.520 --> 24:13.320
so you can get it on as many vehicles as possible.
24:13.320 --> 24:14.760
But driving the cost out of it
24:14.760 --> 24:18.800
doesn't drive up performance on the false negative case.
24:18.800 --> 24:21.440
And so you'll continue to not have a technology
24:21.440 --> 24:25.680
that could really be available for a self driven vehicle.
24:25.680 --> 24:28.440
So clearly the communication,
24:28.440 --> 24:31.600
and this probably applies to all four vehicles as well,
24:31.600 --> 24:34.440
the marketing and communication
24:34.440 --> 24:37.040
of what the technology is actually capable of,
24:37.040 --> 24:38.400
how hard it is, how easy it is,
24:38.400 --> 24:41.000
all that kind of stuff is highly problematic.
24:41.000 --> 24:45.640
So say everybody in the world was perfectly communicated
24:45.640 --> 24:48.400
and were made to be completely aware
24:48.400 --> 24:50.000
of every single technology out there,
24:50.000 --> 24:52.840
what it's able to do.
24:52.840 --> 24:54.120
What's your intuition?
24:54.120 --> 24:56.880
And now we're maybe getting into philosophical ground.
24:56.880 --> 25:00.000
Is it possible to have a level two vehicle
25:00.000 --> 25:03.280
where we don't over trust it?
25:04.680 --> 25:05.800
I don't think so.
25:05.800 --> 25:10.800
If people truly understood the risks and internalized it,
25:11.160 --> 25:14.320
then sure, you could do that safely.
25:14.320 --> 25:16.160
But that's a world that doesn't exist.
25:16.160 --> 25:17.520
The people are going to,
25:18.720 --> 25:20.760
if the facts are put in front of them,
25:20.760 --> 25:24.440
they're gonna then combine that with their experience.
25:24.440 --> 25:28.360
And let's say they're using an L2 system
25:28.360 --> 25:30.800
and they go up and down the 101 every day
25:30.800 --> 25:32.720
and they do that for a month.
25:32.720 --> 25:36.200
And it just worked every day for a month.
25:36.200 --> 25:39.000
Like that's pretty compelling at that point,
25:39.000 --> 25:41.800
just even if you know the statistics,
25:41.800 --> 25:43.400
you're like, well, I don't know,
25:43.400 --> 25:44.760
maybe there's something funny about those.
25:44.760 --> 25:46.920
Maybe they're driving in difficult places.
25:46.920 --> 25:49.840
Like I've seen it with my own eyes, it works.
25:49.840 --> 25:52.400
And the problem is that that sample size that they have,
25:52.400 --> 25:53.880
so it's 30 miles up and down,
25:53.880 --> 25:56.360
so 60 miles times 30 days,
25:56.360 --> 25:58.720
so 60, 180, 1,800 miles.
25:58.720 --> 26:03.280
Like that's a drop in the bucket
26:03.280 --> 26:07.640
compared to the, what, 85 million miles between fatalities.
26:07.640 --> 26:11.400
And so they don't really have a true estimate
26:11.400 --> 26:14.440
based on their personal experience of the real risks,
26:14.440 --> 26:15.640
but they're gonna trust it anyway,
26:15.640 --> 26:16.480
because it's hard not to.
26:16.480 --> 26:18.640
It worked for a month, what's gonna change?
26:18.640 --> 26:21.640
So even if you start a perfect understanding of the system,
26:21.640 --> 26:24.160
your own experience will make it drift.
26:24.160 --> 26:25.920
I mean, that's a big concern.
26:25.920 --> 26:28.160
Over a year, over two years even,
26:28.160 --> 26:29.440
it doesn't have to be months.
26:29.440 --> 26:32.920
And I think that as this technology moves
26:32.920 --> 26:37.760
from what I would say is kind of the more technology savvy
26:37.760 --> 26:40.880
ownership group to the mass market,
26:42.640 --> 26:44.600
you may be able to have some of those folks
26:44.600 --> 26:46.280
who are really familiar with technology,
26:46.280 --> 26:48.840
they may be able to internalize it better.
26:48.840 --> 26:50.800
And your kind of immunization
26:50.800 --> 26:53.360
against this kind of false risk assessment
26:53.360 --> 26:54.280
might last longer,
26:54.280 --> 26:58.680
but as folks who aren't as savvy about that
26:58.680 --> 27:00.880
read the material and they compare that
27:00.880 --> 27:02.160
to their personal experience,
27:02.160 --> 27:07.160
I think there it's going to move more quickly.
27:08.160 --> 27:11.280
So your work, the program that you've created at Google
27:11.280 --> 27:16.280
and now at Aurora is focused more on the second path
27:16.600 --> 27:18.480
of creating full autonomy.
27:18.480 --> 27:20.880
So it's such a fascinating,
27:20.880 --> 27:24.560
I think it's one of the most interesting AI problems
27:24.560 --> 27:25.600
of the century, right?
27:25.600 --> 27:28.280
It's, I just talked to a lot of people,
27:28.280 --> 27:29.440
just regular people, I don't know,
27:29.440 --> 27:31.720
my mom, about autonomous vehicles,
27:31.720 --> 27:34.520
and you begin to grapple with ideas
27:34.520 --> 27:38.080
of giving your life control over to a machine.
27:38.080 --> 27:40.040
It's philosophically interesting,
27:40.040 --> 27:41.760
it's practically interesting.
27:41.760 --> 27:43.720
So let's talk about safety.
27:43.720 --> 27:46.240
How do you think we demonstrate,
27:46.240 --> 27:47.880
you've spoken about metrics in the past,
27:47.880 --> 27:51.880
how do you think we demonstrate to the world
27:51.880 --> 27:56.160
that an autonomous vehicle, an Aurora system is safe?
27:56.160 --> 27:57.320
This is one where it's difficult
27:57.320 --> 27:59.280
because there isn't a soundbite answer.
27:59.280 --> 28:04.280
That we have to show a combination of work
28:05.960 --> 28:08.360
that was done diligently and thoughtfully,
28:08.360 --> 28:10.840
and this is where something like a functional safety process
28:10.840 --> 28:11.680
is part of that.
28:11.680 --> 28:14.360
It's like here's the way we did the work,
28:15.280 --> 28:17.160
that means that we were very thorough.
28:17.160 --> 28:20.040
So if you believe that what we said
28:20.040 --> 28:21.440
about this is the way we did it,
28:21.440 --> 28:22.720
then you can have some confidence
28:22.720 --> 28:25.200
that we were thorough in the engineering work
28:25.200 --> 28:26.920
we put into the system.
28:26.920 --> 28:28.920
And then on top of that,
28:28.920 --> 28:32.000
to kind of demonstrate that we weren't just thorough,
28:32.000 --> 28:33.800
we were actually good at what we did,
28:35.280 --> 28:38.200
there'll be a kind of a collection of evidence
28:38.200 --> 28:40.440
in terms of demonstrating that the capabilities
28:40.440 --> 28:42.920
worked the way we thought they did,
28:42.920 --> 28:45.320
statistically and to whatever degree
28:45.320 --> 28:47.280
we can demonstrate that,
28:48.160 --> 28:50.320
both in some combination of simulations,
28:50.320 --> 28:53.080
some combination of unit testing
28:53.080 --> 28:54.640
and decomposition testing,
28:54.640 --> 28:57.000
and then some part of it will be on road data.
28:58.160 --> 29:02.680
And I think the way we'll ultimately
29:02.680 --> 29:04.000
convey this to the public
29:04.000 --> 29:06.760
is there'll be clearly some conversation
29:06.760 --> 29:08.200
with the public about it,
29:08.200 --> 29:12.040
but we'll kind of invoke the kind of the trusted nodes
29:12.040 --> 29:13.880
and that we'll spend more time
29:13.880 --> 29:17.280
being able to go into more depth with folks like NHTSA
29:17.280 --> 29:19.720
and other federal and state regulatory bodies
29:19.720 --> 29:22.080
and kind of given that they are
29:22.080 --> 29:25.200
operating in the public interest and they're trusted,
29:26.240 --> 29:28.640
that if we can show enough work to them
29:28.640 --> 29:30.000
that they're convinced,
29:30.000 --> 29:33.800
then I think we're in a pretty good place.
29:33.800 --> 29:35.000
That means you work with people
29:35.000 --> 29:36.920
that are essentially experts at safety
29:36.920 --> 29:39.000
to try to discuss and show.
29:39.000 --> 29:41.720
Do you think, the answer's probably no,
29:41.720 --> 29:42.920
but just in case,
29:42.920 --> 29:44.360
do you think there exists a metric?
29:44.360 --> 29:46.320
So currently people have been using
29:46.320 --> 29:48.200
number of disengagements.
29:48.200 --> 29:50.120
And it quickly turns into a marketing scheme
29:50.120 --> 29:54.280
to sort of you alter the experiments you run to adjust.
29:54.280 --> 29:56.280
I think you've spoken that you don't like.
29:56.280 --> 29:57.120
Don't love it.
29:57.120 --> 29:59.680
No, in fact, I was on the record telling DMV
29:59.680 --> 30:01.960
that I thought this was not a great metric.
30:01.960 --> 30:05.280
Do you think it's possible to create a metric,
30:05.280 --> 30:09.440
a number that could demonstrate safety
30:09.440 --> 30:12.320
outside of fatalities?
30:12.320 --> 30:13.440
So I do.
30:13.440 --> 30:16.560
And I think that it won't be just one number.
30:17.600 --> 30:21.280
So as we are internally grappling with this,
30:21.280 --> 30:23.560
and at some point we'll be able to talk
30:23.560 --> 30:25.040
more publicly about it,
30:25.040 --> 30:28.520
is how do we think about human performance
30:28.520 --> 30:29.840
in different tasks,
30:29.840 --> 30:32.160
say detecting traffic lights
30:32.160 --> 30:36.200
or safely making a left turn across traffic?
30:37.680 --> 30:40.080
And what do we think the failure rates are
30:40.080 --> 30:42.520
for those different capabilities for people?
30:42.520 --> 30:44.760
And then demonstrating to ourselves
30:44.760 --> 30:48.480
and then ultimately folks in the regulatory role
30:48.480 --> 30:50.760
and then ultimately the public
30:50.760 --> 30:52.400
that we have confidence that our system
30:52.400 --> 30:54.760
will work better than that.
30:54.760 --> 30:57.040
And so these individual metrics
30:57.040 --> 31:00.720
will kind of tell a compelling story ultimately.
31:01.760 --> 31:03.920
I do think at the end of the day
31:03.920 --> 31:06.640
what we care about in terms of safety
31:06.640 --> 31:11.320
is life saved and injuries reduced.
31:12.160 --> 31:15.280
And then ultimately kind of casualty dollars
31:16.440 --> 31:19.360
that people aren't having to pay to get their car fixed.
31:19.360 --> 31:22.680
And I do think that in aviation
31:22.680 --> 31:25.880
they look at a kind of an event pyramid
31:25.880 --> 31:28.600
where a crash is at the top of that
31:28.600 --> 31:30.440
and that's the worst event obviously
31:30.440 --> 31:34.240
and then there's injuries and near miss events and whatnot
31:34.240 --> 31:37.320
and violation of operating procedures
31:37.320 --> 31:40.160
and you kind of build a statistical model
31:40.160 --> 31:44.440
of the relevance of the low severity things
31:44.440 --> 31:45.280
or the high severity things.
31:45.280 --> 31:46.120
And I think that's something
31:46.120 --> 31:48.200
where we'll be able to look at as well
31:48.200 --> 31:51.840
because an event per 85 million miles
31:51.840 --> 31:54.440
is statistically a difficult thing
31:54.440 --> 31:56.800
even at the scale of the U.S.
31:56.800 --> 31:59.360
to kind of compare directly.
31:59.360 --> 32:02.240
And that event fatality that's connected
32:02.240 --> 32:07.240
to an autonomous vehicle is significantly
32:07.440 --> 32:09.160
at least currently magnified
32:09.160 --> 32:12.320
in the amount of attention it gets.
32:12.320 --> 32:15.080
So that speaks to public perception.
32:15.080 --> 32:16.720
I think the most popular topic
32:16.720 --> 32:19.480
about autonomous vehicles in the public
32:19.480 --> 32:23.080
is the trolley problem formulation, right?
32:23.080 --> 32:27.000
Which has, let's not get into that too much
32:27.000 --> 32:29.600
but is misguided in many ways.
32:29.600 --> 32:32.320
But it speaks to the fact that people are grappling
32:32.320 --> 32:36.160
with this idea of giving control over to a machine.
32:36.160 --> 32:41.160
So how do you win the hearts and minds of the people
32:41.560 --> 32:44.600
that autonomy is something that could be a part
32:44.600 --> 32:45.520
of their lives?
32:45.520 --> 32:47.640
I think you let them experience it, right?
32:47.640 --> 32:50.440
I think it's right.
32:50.440 --> 32:52.800
I think people should be skeptical.
32:52.800 --> 32:55.680
I think people should ask questions.
32:55.680 --> 32:57.000
I think they should doubt
32:57.000 --> 33:00.120
because this is something new and different.
33:00.120 --> 33:01.880
They haven't touched it yet.
33:01.880 --> 33:03.640
And I think that's perfectly reasonable.
33:03.640 --> 33:07.320
And, but at the same time,
33:07.320 --> 33:09.320
it's clear there's an opportunity to make the road safer.
33:09.320 --> 33:12.440
It's clear that we can improve access to mobility.
33:12.440 --> 33:14.960
It's clear that we can reduce the cost of mobility.
33:16.640 --> 33:19.480
And that once people try that
33:19.480 --> 33:22.720
and understand that it's safe
33:22.720 --> 33:24.440
and are able to use in their daily lives,
33:24.440 --> 33:25.280
I think it's one of these things
33:25.280 --> 33:28.040
that will just be obvious.
33:28.040 --> 33:32.240
And I've seen this practically in demonstrations
33:32.240 --> 33:35.560
that I've given where I've had people come in
33:35.560 --> 33:38.840
and they're very skeptical.
33:38.840 --> 33:40.440
Again, in a vehicle, my favorite one
33:40.440 --> 33:42.560
is taking somebody out on the freeway
33:42.560 --> 33:46.000
and we're on the 101 driving at 65 miles an hour.
33:46.000 --> 33:48.400
And after 10 minutes, they kind of turn and ask,
33:48.400 --> 33:49.480
is that all it does?
33:49.480 --> 33:52.080
And you're like, it's a self driving car.
33:52.080 --> 33:54.840
I'm not sure exactly what you thought it would do, right?
33:54.840 --> 33:57.920
But it becomes mundane,
33:58.840 --> 34:01.480
which is exactly what you want a technology
34:01.480 --> 34:02.720
like this to be, right?
34:02.720 --> 34:07.280
We don't really, when I turn the light switch on in here,
34:07.280 --> 34:12.000
I don't think about the complexity of those electrons
34:12.000 --> 34:14.200
being pushed down a wire from wherever it was
34:14.200 --> 34:15.240
and being generated.
34:15.240 --> 34:19.080
It's like, I just get annoyed if it doesn't work, right?
34:19.080 --> 34:21.400
And what I value is the fact
34:21.400 --> 34:23.080
that I can do other things in this space.
34:23.080 --> 34:24.560
I can see my colleagues.
34:24.560 --> 34:26.160
I can read stuff on a paper.
34:26.160 --> 34:29.200
I can not be afraid of the dark.
34:30.360 --> 34:33.320
And I think that's what we want this technology to be like
34:33.320 --> 34:34.640
is it's in the background
34:34.640 --> 34:37.120
and people get to have those life experiences
34:37.120 --> 34:38.440
and do so safely.
34:38.440 --> 34:42.160
So putting this technology in the hands of people
34:42.160 --> 34:46.320
speaks to scale of deployment, right?
34:46.320 --> 34:50.880
So what do you think the dreaded question about the future
34:50.880 --> 34:53.560
because nobody can predict the future,
34:53.560 --> 34:57.240
but just maybe speak poetically
34:57.240 --> 35:00.880
about when do you think we'll see a large scale deployment
35:00.880 --> 35:05.880
of autonomous vehicles, 10,000, those kinds of numbers?
35:06.680 --> 35:08.240
We'll see that within 10 years.
35:09.240 --> 35:10.240
I'm pretty confident.
35:14.040 --> 35:16.040
What's an impressive scale?
35:16.040 --> 35:19.200
What moment, so you've done the DARPA challenge
35:19.200 --> 35:20.440
where there's one vehicle.
35:20.440 --> 35:23.960
At which moment does it become, wow, this is serious scale?
35:23.960 --> 35:26.520
So I think the moment it gets serious
35:26.520 --> 35:31.520
is when we really do have a driverless vehicle
35:32.240 --> 35:34.120
operating on public roads
35:35.000 --> 35:37.960
and that we can do that kind of continuously.
35:37.960 --> 35:38.880
Without a safety driver.
35:38.880 --> 35:40.440
Without a safety driver in the vehicle.
35:40.440 --> 35:41.560
I think at that moment,
35:41.560 --> 35:44.400
we've kind of crossed the zero to one threshold.
35:45.920 --> 35:50.200
And then it is about how do we continue to scale that?
35:50.200 --> 35:53.960
How do we build the right business models?
35:53.960 --> 35:56.320
How do we build the right customer experience around it
35:56.320 --> 35:59.960
so that it is actually a useful product out in the world?
36:00.960 --> 36:03.600
And I think that is really,
36:03.600 --> 36:05.920
at that point it moves from
36:05.920 --> 36:09.200
what is this kind of mixed science engineering project
36:09.200 --> 36:12.360
into engineering and commercialization
36:12.360 --> 36:15.840
and really starting to deliver on the value
36:15.840 --> 36:20.680
that we all see here and actually making that real in the world.
36:20.680 --> 36:22.240
What do you think that deployment looks like?
36:22.240 --> 36:26.440
Where do we first see the inkling of no safety driver,
36:26.440 --> 36:28.600
one or two cars here and there?
36:28.600 --> 36:29.800
Is it on the highway?
36:29.800 --> 36:33.160
Is it in specific routes in the urban environment?
36:33.160 --> 36:36.920
I think it's gonna be urban, suburban type environments.
36:37.880 --> 36:41.560
Yeah, with Aurora, when we thought about how to tackle this,
36:41.560 --> 36:45.040
it was kind of in vogue to think about trucking
36:46.040 --> 36:47.800
as opposed to urban driving.
36:47.800 --> 36:51.280
And again, the human intuition around this
36:51.280 --> 36:55.400
is that freeways are easier to drive on
36:57.080 --> 36:59.280
because everybody's kind of going in the same direction
36:59.280 --> 37:01.560
and lanes are a little wider, et cetera.
37:01.560 --> 37:03.320
And I think that that intuition is pretty good,
37:03.320 --> 37:06.040
except we don't really care about most of the time.
37:06.040 --> 37:08.400
We care about all of the time.
37:08.400 --> 37:10.880
And when you're driving on a freeway with a truck,
37:10.880 --> 37:13.440
say 70 miles an hour,
37:14.600 --> 37:16.240
and you've got 70,000 pound load with you,
37:16.240 --> 37:18.880
that's just an incredible amount of kinetic energy.
37:18.880 --> 37:21.440
And so when that goes wrong, it goes really wrong.
37:22.640 --> 37:27.640
And those challenges that you see occur more rarely,
37:27.800 --> 37:31.120
so you don't get to learn as quickly.
37:31.120 --> 37:34.720
And they're incrementally more difficult than urban driving,
37:34.720 --> 37:37.440
but they're not easier than urban driving.
37:37.440 --> 37:41.640
And so I think this happens in moderate speed
37:41.640 --> 37:45.280
urban environments because if two vehicles crash
37:45.280 --> 37:48.120
at 25 miles per hour, it's not good,
37:48.120 --> 37:50.120
but probably everybody walks away.
37:51.080 --> 37:53.720
And those events where there's the possibility
37:53.720 --> 37:55.800
for that occurring happen frequently.
37:55.800 --> 37:58.000
So we get to learn more rapidly.
37:58.000 --> 38:01.360
We get to do that with lower risk for everyone.
38:02.520 --> 38:04.360
And then we can deliver value to people
38:04.360 --> 38:05.880
that need to get from one place to another.
38:05.880 --> 38:08.160
And once we've got that solved,
38:08.160 --> 38:11.320
then the freeway driving part of this just falls out.
38:11.320 --> 38:13.080
But we're able to learn more safely,
38:13.080 --> 38:15.200
more quickly in the urban environment.
38:15.200 --> 38:18.760
So 10 years and then scale 20, 30 year,
38:18.760 --> 38:22.040
who knows if a sufficiently compelling experience
38:22.040 --> 38:24.400
is created, it could be faster and slower.
38:24.400 --> 38:27.160
Do you think there could be breakthroughs
38:27.160 --> 38:29.920
and what kind of breakthroughs might there be
38:29.920 --> 38:32.400
that completely change that timeline?
38:32.400 --> 38:35.360
Again, not only am I asking you to predict the future,
38:35.360 --> 38:37.360
I'm asking you to predict breakthroughs
38:37.360 --> 38:38.360
that haven't happened yet.
38:38.360 --> 38:41.440
So what's the, I think another way to ask that
38:41.440 --> 38:44.320
would be if I could wave a magic wand,
38:44.320 --> 38:46.720
what part of the system would I make work today
38:46.720 --> 38:49.480
to accelerate it as quickly as possible?
38:52.120 --> 38:54.200
Don't say infrastructure, please don't say infrastructure.
38:54.200 --> 38:56.320
No, it's definitely not infrastructure.
38:56.320 --> 39:00.600
It's really that perception forecasting capability.
39:00.600 --> 39:04.840
So if tomorrow you could give me a perfect model
39:04.840 --> 39:06.960
of what's happened, what is happening
39:06.960 --> 39:09.200
and what will happen for the next five seconds
39:10.360 --> 39:13.040
around a vehicle on the roadway,
39:13.040 --> 39:15.360
that would accelerate things pretty dramatically.
39:15.360 --> 39:17.600
Are you, in terms of staying up at night,
39:17.600 --> 39:21.760
are you mostly bothered by cars, pedestrians or cyclists?
39:21.760 --> 39:25.960
So I worry most about the vulnerable road users
39:25.960 --> 39:28.480
about the combination of cyclists and cars, right?
39:28.480 --> 39:31.960
Or cyclists and pedestrians because they're not in armor.
39:31.960 --> 39:36.480
The cars, they're bigger, they've got protection
39:36.480 --> 39:39.440
for the people and so the ultimate risk is lower there.
39:41.080 --> 39:43.240
Whereas a pedestrian or a cyclist,
39:43.240 --> 39:46.480
they're out on the road and they don't have any protection
39:46.480 --> 39:49.720
and so we need to pay extra attention to that.
39:49.720 --> 39:54.120
Do you think about a very difficult technical challenge
39:55.720 --> 39:58.520
of the fact that pedestrians,
39:58.520 --> 40:00.240
if you try to protect pedestrians
40:00.240 --> 40:04.560
by being careful and slow, they'll take advantage of that.
40:04.560 --> 40:09.040
So the game theoretic dance, does that worry you
40:09.040 --> 40:12.480
of how, from a technical perspective, how we solve that?
40:12.480 --> 40:14.560
Because as humans, the way we solve that
40:14.560 --> 40:17.240
is kind of nudge our way through the pedestrians
40:17.240 --> 40:20.000
which doesn't feel, from a technical perspective,
40:20.000 --> 40:22.300
as a appropriate algorithm.
40:23.200 --> 40:25.920
But do you think about how we solve that problem?
40:25.920 --> 40:30.920
Yeah, I think there's two different concepts there.
40:31.360 --> 40:35.820
So one is, am I worried that because these vehicles
40:35.820 --> 40:37.600
are self driving, people will kind of step in the road
40:37.600 --> 40:38.640
and take advantage of them?
40:38.640 --> 40:43.640
And I've heard this and I don't really believe it
40:43.760 --> 40:45.960
because if I'm driving down the road
40:45.960 --> 40:48.400
and somebody steps in front of me, I'm going to stop.
40:50.600 --> 40:53.660
Even if I'm annoyed, I'm not gonna just drive
40:53.660 --> 40:56.400
through a person stood in the road.
40:56.400 --> 41:00.400
And so I think today people can take advantage of this
41:00.400 --> 41:02.560
and you do see some people do it.
41:02.560 --> 41:04.180
I guess there's an incremental risk
41:04.180 --> 41:05.880
because maybe they have lower confidence
41:05.880 --> 41:07.720
that I'm gonna see them than they might have
41:07.720 --> 41:10.400
for an automated vehicle and so maybe that shifts
41:10.400 --> 41:12.040
it a little bit.
41:12.040 --> 41:14.360
But I think people don't wanna get hit by cars.
41:14.360 --> 41:17.080
And so I think that I'm not that worried
41:17.080 --> 41:18.760
about people walking out of the 101
41:18.760 --> 41:23.760
and creating chaos more than they would today.
41:24.400 --> 41:27.040
Regarding kind of the nudging through a big stream
41:27.040 --> 41:30.040
of pedestrians leaving a concert or something,
41:30.040 --> 41:33.520
I think that is further down the technology pipeline.
41:33.520 --> 41:36.960
I think that you're right, that's tricky.
41:36.960 --> 41:38.620
I don't think it's necessarily,
41:40.360 --> 41:43.600
I think the algorithm people use for this is pretty simple.
41:43.600 --> 41:44.800
It's kind of just move forward slowly
41:44.800 --> 41:46.800
and if somebody's really close then stop.
41:46.800 --> 41:50.880
And I think that that probably can be replicated
41:50.880 --> 41:54.040
pretty easily and particularly given that
41:54.040 --> 41:55.720
you don't do this at 30 miles an hour,
41:55.720 --> 41:59.080
you do it at one, that even in those situations
41:59.080 --> 42:01.200
the risk is relatively minimal.
42:01.200 --> 42:03.640
But it's not something we're thinking about
42:03.640 --> 42:04.560
in any serious way.
42:04.560 --> 42:07.920
And probably that's less an algorithm problem
42:07.920 --> 42:10.160
and more creating a human experience.
42:10.160 --> 42:14.300
So the HCI people that create a visual display
42:14.300 --> 42:16.260
that you're pleasantly as a pedestrian
42:16.260 --> 42:20.760
nudged out of the way, that's an experience problem,
42:20.760 --> 42:22.000
not an algorithm problem.
42:22.880 --> 42:25.480
Who's the main competitor to Aurora today?
42:25.480 --> 42:28.640
And how do you outcompete them in the long run?
42:28.640 --> 42:31.200
So we really focus a lot on what we're doing here.
42:31.200 --> 42:34.480
I think that, I've said this a few times,
42:34.480 --> 42:37.960
that this is a huge difficult problem
42:37.960 --> 42:40.320
and it's great that a bunch of companies are tackling it
42:40.320 --> 42:42.320
because I think it's so important for society
42:42.320 --> 42:43.800
that somebody gets there.
42:43.800 --> 42:48.800
So we don't spend a whole lot of time
42:49.120 --> 42:51.600
thinking tactically about who's out there
42:51.600 --> 42:55.240
and how do we beat that person individually.
42:55.240 --> 42:58.720
What are we trying to do to go faster ultimately?
42:59.760 --> 43:02.640
Well part of it is the leadership team we have
43:02.640 --> 43:04.200
has got pretty tremendous experience.
43:04.200 --> 43:06.440
And so we kind of understand the landscape
43:06.440 --> 43:09.160
and understand where the cul de sacs are to some degree
43:09.160 --> 43:10.980
and we try and avoid those.
43:10.980 --> 43:14.260
I think there's a part of it,
43:14.260 --> 43:16.260
just this great team we've built.
43:16.260 --> 43:19.080
People, this is a technology and a company
43:19.080 --> 43:22.320
that people believe in the mission of
43:22.320 --> 43:23.740
and so it allows us to attract
43:23.740 --> 43:25.740
just awesome people to go work.
43:26.800 --> 43:29.320
We've got a culture I think that people appreciate
43:29.320 --> 43:30.460
that allows them to focus,
43:30.460 --> 43:33.120
allows them to really spend time solving problems.
43:33.120 --> 43:35.900
And I think that keeps them energized.
43:35.900 --> 43:38.940
And then we've invested hard,
43:38.940 --> 43:43.500
invested heavily in the infrastructure
43:43.500 --> 43:46.540
and architectures that we think will ultimately accelerate us.
43:46.540 --> 43:50.660
So because of the folks we're able to bring in early on,
43:50.660 --> 43:53.540
because of the great investors we have,
43:53.540 --> 43:56.780
we don't spend all of our time doing demos
43:56.780 --> 43:58.660
and kind of leaping from one demo to the next.
43:58.660 --> 44:02.820
We've been given the freedom to invest in
44:03.940 --> 44:05.500
infrastructure to do machine learning,
44:05.500 --> 44:08.600
infrastructure to pull data from our on road testing,
44:08.600 --> 44:11.500
infrastructure to use that to accelerate engineering.
44:11.500 --> 44:14.480
And I think that early investment
44:14.480 --> 44:17.340
and continuing investment in those kind of tools
44:17.340 --> 44:19.780
will ultimately allow us to accelerate
44:19.780 --> 44:21.940
and do something pretty incredible.
44:21.940 --> 44:23.420
Chris, beautifully put.
44:23.420 --> 44:24.660
It's a good place to end.
44:24.660 --> 44:26.500
Thank you so much for talking today.
44:26.500 --> 44:47.940
Thank you very much. Really enjoyed it.