back to indexChris Urmson: Self-Driving Cars at Aurora, Google, CMU, and DARPA | Lex Fridman Podcast #28
link |
The following is a conversation with Chris Sampson.
link |
He was a CTO of the Google self driving car team,
link |
a key engineer and leader behind the Carnegie Mellon
link |
University autonomous vehicle entries in the DARPA Grand
link |
Challenges and the winner of the DARPA Urban Challenge.
link |
Today, he's the CEO of Aurora Innovation, an autonomous
link |
vehicle software company.
link |
He started with Sterling Anderson,
link |
who was the former director of Tesla Autopilot,
link |
and drew back now, Uber's former autonomy and perception lead.
link |
Chris is one of the top roboticists and autonomous
link |
vehicle experts in the world, and a longtime voice
link |
of reason in a space that is shrouded
link |
in both mystery and hype.
link |
He both acknowledges the incredible challenges
link |
involved in solving the problem of autonomous driving
link |
and is working hard to solve it.
link |
This is the Artificial Intelligence podcast.
link |
If you enjoy it, subscribe on YouTube,
link |
give it five stars on iTunes, support it on Patreon,
link |
or simply connect with me on Twitter
link |
at Lex Friedman, spelled F R I D M A N.
link |
And now, here's my conversation with Chris Sampson.
link |
You were part of both the DARPA Grand Challenge
link |
and the DARPA Urban Challenge teams
link |
at CMU with Red Whitaker.
link |
What technical or philosophical things
link |
have you learned from these races?
link |
I think the high order bit was that it could be done.
link |
I think that was the thing that was
link |
incredible about the first of the Grand Challenges,
link |
that I remember I was a grad student at Carnegie Mellon,
link |
and there was kind of this dichotomy of it
link |
seemed really hard, so that would
link |
be cool and interesting.
link |
But at the time, we were the only robotics institute around,
link |
and so if we went into it and fell on our faces,
link |
that would be embarrassing.
link |
So I think just having the will to go do it,
link |
to try to do this thing that at the time
link |
was marked as darn near impossible,
link |
and then after a couple of tries,
link |
be able to actually make it happen,
link |
I think that was really exciting.
link |
But at which point did you believe it was possible?
link |
Did you from the very beginning?
link |
Did you personally?
link |
Because you're one of the lead engineers.
link |
You actually had to do a lot of the work.
link |
Yeah, I was the technical director there,
link |
and did a lot of the work, along with a bunch
link |
of other really good people.
link |
Did I believe it could be done?
link |
Why would you go do something you thought
link |
was completely impossible?
link |
We thought it was going to be hard.
link |
We didn't know how we were going to be able to do it.
link |
We didn't know if we'd be able to do it the first time.
link |
Turns out we couldn't.
link |
That, yeah, I guess you have to.
link |
I think there's a certain benefit to naivete, right?
link |
That if you don't know how hard something really is,
link |
you try different things, and it gives you an opportunity
link |
that others who are wiser maybe don't have.
link |
What were the biggest pain points?
link |
Mechanical, sensors, hardware, software,
link |
algorithms for mapping, localization,
link |
just general perception, control?
link |
Like hardware, software, first of all?
link |
I think that's the joy of this field, is that it's all hard
link |
and that you have to be good at each part of it.
link |
So for the urban challenges, if I look back at it from today,
link |
it should be easy today, that it was a static world.
link |
There weren't other actors moving through it,
link |
is what that means.
link |
It was out in the desert, so you get really good GPS.
link |
So that went, and we could map it roughly.
link |
And so in retrospect now, it's within the realm of things
link |
we could do back then.
link |
Just actually getting the vehicle and the,
link |
there's a bunch of engineering work
link |
to get the vehicle so that we could control it and drive it.
link |
That's still a pain today, but it was even more so back then.
link |
And then the uncertainty of exactly what they wanted us to do
link |
was part of the challenge as well.
link |
Right, you didn't actually know the track heading in here.
link |
You knew approximately, but you didn't actually
link |
know the route that was going to be taken.
link |
That's right, we didn't know the route.
link |
We didn't even really, the way the rules had been described,
link |
you had to kind of guess.
link |
So if you think back to that challenge,
link |
the idea was that the government would give us,
link |
the DARPA would give us a set of waypoints
link |
and kind of the width that you had to stay within
link |
between the line that went between each of those waypoints.
link |
And so the most devious thing they could have done
link |
is set a kilometer wide corridor across a field
link |
of scrub brush and rocks and said, go figure it out.
link |
Fortunately, it really, it turned into basically driving
link |
along a set of trails, which is much more relevant
link |
to the application they were looking for.
link |
But no, it was a hell of a thing back in the day.
link |
So the legend, Red, was kind of leading that effort
link |
in terms of just broadly speaking.
link |
So you're a leader now.
link |
What have you learned from Red about leadership?
link |
I think there's a couple things.
link |
One is go and try those really hard things.
link |
That's where there is an incredible opportunity.
link |
I think the other big one, though,
link |
is to see people for who they can be, not who they are.
link |
It's one of the things that I actually,
link |
one of the deepest lessons I learned from Red
link |
was that he would look at undergraduates
link |
or graduate students and empower them to be leaders,
link |
to have responsibility, to do great things
link |
that I think another person might look at them
link |
and think, oh, well, that's just an undergraduate student.
link |
What could they know?
link |
And so I think that kind of trust but verify,
link |
have confidence in what people can become,
link |
I think is a really powerful thing.
link |
So through that, let's just fast forward through the history.
link |
Can you maybe talk through the technical evolution
link |
of autonomous vehicle systems
link |
from the first two Grand Challenges to the Urban Challenge
link |
to today, are there major shifts in your mind
link |
or is it the same kind of technology just made more robust?
link |
I think there's been some big, big steps.
link |
So for the Grand Challenge,
link |
the real technology that unlocked that was HD mapping.
link |
Prior to that, a lot of the off road robotics work
link |
had been done without any real prior model
link |
of what the vehicle was going to encounter.
link |
And so that innovation that the fact that we could get
link |
decimeter resolution models was really a big deal.
link |
And that allowed us to kind of bound the complexity
link |
of the driving problem the vehicle had
link |
and allowed it to operate at speed
link |
because we could assume things about the environment
link |
that it was going to encounter.
link |
So that was the big step there.
link |
For the Urban Challenge,
link |
one of the big technological innovations there
link |
was the multi beam LIDAR
link |
and being able to generate high resolution,
link |
mid to long range 3D models of the world
link |
and use that for understanding the world around the vehicle.
link |
And that was really kind of a game changing technology.
link |
In parallel with that,
link |
we saw a bunch of other technologies
link |
that had been kind of converging
link |
half their day in the sun.
link |
So Bayesian estimation had been,
link |
SLAM had been a big field in robotics.
link |
You would go to a conference a couple of years before that
link |
and every paper would effectively have SLAM somewhere in it.
link |
And so seeing that the Bayesian estimation techniques
link |
play out on a very visible stage,
link |
I thought that was pretty exciting to see.
link |
And mostly SLAM was done based on LIDAR at that time.
link |
Yeah, and in fact, we weren't really doing SLAM per se
link |
in real time because we had a model ahead of time,
link |
we had a roadmap, but we were doing localization.
link |
And we were using the LIDAR or the cameras
link |
depending on who exactly was doing it
link |
to localize to a model of the world.
link |
And I thought that was a big step
link |
from kind of naively trusting GPS, INS before that.
link |
And again, lots of work had been going on in this field.
link |
Certainly this was not doing anything
link |
particularly innovative in SLAM or in localization,
link |
but it was seeing that technology necessary
link |
in a real application on a big stage,
link |
I thought was very cool.
link |
So for the urban challenge,
link |
those are already maps constructed offline in general.
link |
And did people do that individually,
link |
did individual teams do it individually
link |
so they had their own different approaches there
link |
or did everybody kind of share that information
link |
at least intuitively?
link |
So DARPA gave all the teams a model of the world, a map.
link |
And then one of the things that we had to figure out
link |
back then was, and it's still one of these things
link |
that trips people up today
link |
is actually the coordinate system.
link |
So you get a latitude longitude
link |
and to so many decimal places,
link |
you don't really care about kind of the ellipsoid
link |
of the earth that's being used.
link |
But when you want to get to 10 centimeter
link |
or centimeter resolution,
link |
you care whether the coordinate system is NADS 83
link |
or WGS 84 or these are different ways to describe
link |
both the kind of non sphericalness of the earth,
link |
but also kind of the, I think,
link |
I can't remember which one,
link |
the tectonic shifts that are happening
link |
and how to transform the global datum as a function of that.
link |
So getting a map and then actually matching it to reality
link |
to centimeter resolution, that was kind of interesting
link |
and fun back then.
link |
So how much work was the perception doing there?
link |
So how much were you relying on localization based on maps
link |
without using perception to register to the maps?
link |
And I guess the question is how advanced
link |
was perception at that point?
link |
It's certainly behind where we are today, right?
link |
We're more than a decade since the urban challenge.
link |
But the core of it was there.
link |
That we were tracking vehicles.
link |
We had to do that at 100 plus meter range
link |
because we had to merge with other traffic.
link |
We were using, again, Bayesian estimates
link |
for state of these vehicles.
link |
We had to deal with a bunch of the problems
link |
that you think of today,
link |
of predicting where that vehicle's going to be
link |
a few seconds into the future.
link |
We had to deal with the fact
link |
that there were multiple hypotheses for that
link |
because a vehicle at an intersection might be going right
link |
or it might be going straight
link |
or it might be making a left turn.
link |
And we had to deal with the challenge of the fact
link |
that our behavior was going to impact the behavior
link |
of that other operator.
link |
And we did a lot of that in relatively naive ways,
link |
but it kind of worked.
link |
Still had to have some kind of solution.
link |
And so where does that, 10 years later,
link |
where does that take us today
link |
from that artificial city construction
link |
to real cities to the urban environment?
link |
Yeah, I think the biggest thing
link |
is that the actors are truly unpredictable.
link |
That most of the time, the drivers on the road,
link |
the other road users are out there behaving well,
link |
but every once in a while they're not.
link |
The variety of other vehicles is, you have all of them.
link |
In terms of behavior, in terms of perception, or both?
link |
Back then we didn't have to deal with cyclists,
link |
we didn't have to deal with pedestrians,
link |
didn't have to deal with traffic lights.
link |
The scale over which that you have to operate is now
link |
is much larger than the air base
link |
that we were thinking about back then.
link |
So what, easy question,
link |
what do you think is the hardest part about driving?
link |
Yeah, no, I'm joking.
link |
I'm sure nothing really jumps out at you as one thing,
link |
but in the jump from the urban challenge to the real world,
link |
is there something that's a particular,
link |
you foresee as very serious, difficult challenge?
link |
I think the most fundamental difference
link |
is that we're doing it for real.
link |
That in that environment,
link |
it was both a limited complexity environment
link |
because certain actors weren't there,
link |
because the roads were maintained,
link |
there were barriers keeping people separate
link |
from robots at the time,
link |
and it only had to work for 60 miles.
link |
Which, looking at it from 2006,
link |
it had to work for 60 miles, right?
link |
Looking at it from now,
link |
we want things that will go and drive
link |
for half a million miles,
link |
and it's just a different game.
link |
you said LiDAR came into the game early on,
link |
and it's really the primary driver
link |
of autonomous vehicles today as a sensor.
link |
So how important is the role of LiDAR
link |
in the sensor suite in the near term?
link |
So I think it's essential.
link |
I believe, but I also believe that cameras are essential,
link |
and I believe the radar is essential.
link |
I think that you really need to use
link |
the composition of data from these different sensors
link |
if you want the thing to really be robust.
link |
The question I wanna ask,
link |
let's see if we can untangle it,
link |
is what are your thoughts on the Elon Musk
link |
provocative statement that LiDAR is a crutch,
link |
that it's a kind of, I guess, growing pains,
link |
and that much of the perception task
link |
can be done with cameras?
link |
So I think it is undeniable
link |
that people walk around without lasers in their foreheads,
link |
and they can get into vehicles and drive them,
link |
and so there's an existence proof
link |
that you can drive using passive vision.
link |
No doubt, can't argue with that.
link |
In terms of sensors, yeah, so there's proof.
link |
Yeah, in terms of sensors, right?
link |
So there's an example that we all go do it,
link |
many of us every day.
link |
In terms of LiDAR being a crutch, sure.
link |
But in the same way that the combustion engine
link |
was a crutch on the path to an electric vehicle,
link |
in the same way that any technology ultimately gets
link |
replaced by some superior technology in the future,
link |
and really the way that I look at this
link |
is that the way we get around on the ground,
link |
the way that we use transportation is broken,
link |
and that we have this, I think the number I saw this morning,
link |
37,000 Americans killed last year on our roads,
link |
and that's just not acceptable.
link |
And so any technology that we can bring to bear
link |
that accelerates this self driving technology
link |
coming to market and saving lives
link |
is technology we should be using.
link |
And it feels just arbitrary to say,
link |
well, I'm not okay with using lasers
link |
because that's whatever,
link |
but I am okay with using an eight megapixel camera
link |
or a 16 megapixel camera.
link |
These are just bits of technology,
link |
and we should be taking the best technology
link |
from the tool bin that allows us to go and solve a problem.
link |
The question I often talk to, well, obviously you do as well,
link |
to sort of automotive companies,
link |
and if there's one word that comes up more often
link |
than anything, it's cost, and trying to drive costs down.
link |
So while it's true that it's a tragic number, the 37,000,
link |
the question is, and I'm not the one asking this question
link |
because I hate this question,
link |
but we want to find the cheapest sensor suite
link |
that creates a safe vehicle.
link |
So in that uncomfortable trade off,
link |
do you foresee LiDAR coming down in cost in the future,
link |
or do you see a day where level four autonomy
link |
is possible without LiDAR?
link |
I see both of those, but it's really a matter of time.
link |
And I think really, maybe I would talk to the question
link |
you asked about the cheapest sensor.
link |
I don't think that's actually what you want.
link |
What you want is a sensor suite that is economically viable.
link |
And then after that, everything is about margin
link |
and driving costs out of the system.
link |
What you also want is a sensor suite that works.
link |
And so it's great to tell a story about
link |
how it would be better to have a self driving system
link |
with a $50 sensor instead of a $500 sensor.
link |
But if the $500 sensor makes it work
link |
and the $50 sensor doesn't work, who cares?
link |
So long as you can actually have an economic opportunity,
link |
there's an economic opportunity there.
link |
And the economic opportunity is important
link |
because that's how you actually have a sustainable business
link |
and that's how you can actually see this come to scale
link |
and be out in the world.
link |
And so when I look at LiDAR,
link |
I see a technology that has no underlying
link |
fundamentally expense to it, fundamental expense to it.
link |
It's going to be more expensive than an imager
link |
because CMOS processes or FAP processes
link |
are dramatically more scalable than mechanical processes.
link |
But we still should be able to drive costs down
link |
substantially on that side.
link |
And then I also do think that with the right business model
link |
you can absorb more,
link |
certainly more cost on the bill of materials.
link |
Yeah, if the sensor suite works, extra value is provided,
link |
thereby you don't need to drive costs down to zero.
link |
It's the basic economics.
link |
You've talked about your intuition
link |
that level two autonomy is problematic
link |
because of the human factor of vigilance,
link |
decrement, complacency, over trust and so on,
link |
just us being human.
link |
We over trust the system,
link |
we start doing even more so partaking
link |
in the secondary activities like smartphones and so on.
link |
Have your views evolved on this point in either direction?
link |
Can you speak to it?
link |
So, and I want to be really careful
link |
because sometimes this gets twisted in a way
link |
that I certainly didn't intend.
link |
So active safety systems are a really important technology
link |
that we should be pursuing and integrating into vehicles.
link |
And there's an opportunity in the near term
link |
to reduce accidents, reduce fatalities,
link |
and we should be pushing on that.
link |
Level two systems are systems
link |
where the vehicle is controlling two axes.
link |
So braking and throttle slash steering.
link |
And I think there are variants of level two systems
link |
that are supporting the driver.
link |
That absolutely we should encourage to be out there.
link |
Where I think there's a real challenge
link |
is in the human factors part around this
link |
and the misconception from the public
link |
around the capability set that that enables
link |
and the trust that they should have in it.
link |
And that is where I kind of,
link |
I'm actually incrementally more concerned
link |
around level three systems
link |
and how exactly a level two system is marketed and delivered
link |
and how much effort people have put into those human factors.
link |
So I still believe several things around this.
link |
One is people will overtrust the technology.
link |
We've seen over the last few weeks
link |
a spate of people sleeping in their Tesla.
link |
I watched an episode last night of Trevor Noah
link |
talking about this and him,
link |
this is a smart guy who has a lot of resources
link |
at his disposal describing a Tesla as a self driving car
link |
and that why shouldn't people be sleeping in their Tesla?
link |
And it's like, well, because it's not a self driving car
link |
and it is not intended to be
link |
and these people will almost certainly die at some point
link |
or hurt other people.
link |
And so we need to really be thoughtful
link |
about how that technology is described
link |
and brought to market.
link |
I also think that because of the economic challenges
link |
we were just talking about,
link |
that these level two driver assistance systems,
link |
that technology path will diverge
link |
from the technology path that we need to be on
link |
to actually deliver truly self driving vehicles,
link |
ones where you can get in it and drive it.
link |
Can get in it and sleep and have the equivalent
link |
or better safety than a human driver behind the wheel.
link |
Because again, the economics are very different
link |
in those two worlds and so that leads
link |
to divergent technology.
link |
So you just don't see the economics
link |
of gradually increasing from level two
link |
and doing so quickly enough
link |
to where it doesn't cause safety, critical safety concerns.
link |
You believe that it needs to diverge at this point
link |
into basically different routes.
link |
And really that comes back to what are those L2
link |
and L1 systems doing?
link |
And they are driver assistance functions
link |
where the people that are marketing that responsibly
link |
are being very clear and putting human factors in place
link |
such that the driver is actually responsible for the vehicle
link |
and that the technology is there to support the driver.
link |
And the safety cases that are built around those
link |
are dependent on that driver attention and attentiveness.
link |
And at that point, you can kind of give up
link |
to some degree for economic reasons,
link |
you can give up on say false negatives.
link |
And the way to think about this
link |
is for a four collision mitigation braking system,
link |
if it half the times the driver missed a vehicle
link |
in front of it, it hit the brakes
link |
and brought the vehicle to a stop,
link |
that would be an incredible, incredible advance
link |
in safety on our roads, right?
link |
That would be equivalent to seat belts.
link |
But it would mean that if that vehicle
link |
wasn't being monitored, it would hit one out of two cars.
link |
And so economically, that's a perfectly good solution
link |
for a driver assistance system.
link |
What you should do at that point,
link |
if you can get it to work 50% of the time,
link |
is drive the cost out of that
link |
so you can get it on as many vehicles as possible.
link |
But driving the cost out of it
link |
doesn't drive up performance on the false negative case.
link |
And so you'll continue to not have a technology
link |
that could really be available for a self driven vehicle.
link |
So clearly the communication,
link |
and this probably applies to all four vehicles as well,
link |
the marketing and communication
link |
of what the technology is actually capable of,
link |
how hard it is, how easy it is,
link |
all that kind of stuff is highly problematic.
link |
So say everybody in the world was perfectly communicated
link |
and were made to be completely aware
link |
of every single technology out there,
link |
what it's able to do.
link |
What's your intuition?
link |
And now we're maybe getting into philosophical ground.
link |
Is it possible to have a level two vehicle
link |
where we don't over trust it?
link |
If people truly understood the risks and internalized it,
link |
then sure, you could do that safely.
link |
But that's a world that doesn't exist.
link |
The people are going to,
link |
if the facts are put in front of them,
link |
they're gonna then combine that with their experience.
link |
And let's say they're using an L2 system
link |
and they go up and down the 101 every day
link |
and they do that for a month.
link |
And it just worked every day for a month.
link |
Like that's pretty compelling at that point,
link |
just even if you know the statistics,
link |
you're like, well, I don't know,
link |
maybe there's something funny about those.
link |
Maybe they're driving in difficult places.
link |
Like I've seen it with my own eyes, it works.
link |
And the problem is that that sample size that they have,
link |
so it's 30 miles up and down,
link |
so 60 miles times 30 days,
link |
so 60, 180, 1,800 miles.
link |
Like that's a drop in the bucket
link |
compared to the, what, 85 million miles between fatalities.
link |
And so they don't really have a true estimate
link |
based on their personal experience of the real risks,
link |
but they're gonna trust it anyway,
link |
because it's hard not to.
link |
It worked for a month, what's gonna change?
link |
So even if you start a perfect understanding of the system,
link |
your own experience will make it drift.
link |
I mean, that's a big concern.
link |
Over a year, over two years even,
link |
it doesn't have to be months.
link |
And I think that as this technology moves
link |
from what I would say is kind of the more technology savvy
link |
ownership group to the mass market,
link |
you may be able to have some of those folks
link |
who are really familiar with technology,
link |
they may be able to internalize it better.
link |
And your kind of immunization
link |
against this kind of false risk assessment
link |
might last longer,
link |
but as folks who aren't as savvy about that
link |
read the material and they compare that
link |
to their personal experience,
link |
I think there it's going to move more quickly.
link |
So your work, the program that you've created at Google
link |
and now at Aurora is focused more on the second path
link |
of creating full autonomy.
link |
So it's such a fascinating,
link |
I think it's one of the most interesting AI problems
link |
of the century, right?
link |
It's, I just talked to a lot of people,
link |
just regular people, I don't know,
link |
my mom, about autonomous vehicles,
link |
and you begin to grapple with ideas
link |
of giving your life control over to a machine.
link |
It's philosophically interesting,
link |
it's practically interesting.
link |
So let's talk about safety.
link |
How do you think we demonstrate,
link |
you've spoken about metrics in the past,
link |
how do you think we demonstrate to the world
link |
that an autonomous vehicle, an Aurora system is safe?
link |
This is one where it's difficult
link |
because there isn't a soundbite answer.
link |
That we have to show a combination of work
link |
that was done diligently and thoughtfully,
link |
and this is where something like a functional safety process
link |
It's like here's the way we did the work,
link |
that means that we were very thorough.
link |
So if you believe that what we said
link |
about this is the way we did it,
link |
then you can have some confidence
link |
that we were thorough in the engineering work
link |
we put into the system.
link |
And then on top of that,
link |
to kind of demonstrate that we weren't just thorough,
link |
we were actually good at what we did,
link |
there'll be a kind of a collection of evidence
link |
in terms of demonstrating that the capabilities
link |
worked the way we thought they did,
link |
statistically and to whatever degree
link |
we can demonstrate that,
link |
both in some combination of simulations,
link |
some combination of unit testing
link |
and decomposition testing,
link |
and then some part of it will be on road data.
link |
And I think the way we'll ultimately
link |
convey this to the public
link |
is there'll be clearly some conversation
link |
with the public about it,
link |
but we'll kind of invoke the kind of the trusted nodes
link |
and that we'll spend more time
link |
being able to go into more depth with folks like NHTSA
link |
and other federal and state regulatory bodies
link |
and kind of given that they are
link |
operating in the public interest and they're trusted,
link |
that if we can show enough work to them
link |
that they're convinced,
link |
then I think we're in a pretty good place.
link |
That means you work with people
link |
that are essentially experts at safety
link |
to try to discuss and show.
link |
Do you think, the answer's probably no,
link |
do you think there exists a metric?
link |
So currently people have been using
link |
number of disengagements.
link |
And it quickly turns into a marketing scheme
link |
to sort of you alter the experiments you run to adjust.
link |
I think you've spoken that you don't like.
link |
No, in fact, I was on the record telling DMV
link |
that I thought this was not a great metric.
link |
Do you think it's possible to create a metric,
link |
a number that could demonstrate safety
link |
outside of fatalities?
link |
And I think that it won't be just one number.
link |
So as we are internally grappling with this,
link |
and at some point we'll be able to talk
link |
more publicly about it,
link |
is how do we think about human performance
link |
in different tasks,
link |
say detecting traffic lights
link |
or safely making a left turn across traffic?
link |
And what do we think the failure rates are
link |
for those different capabilities for people?
link |
And then demonstrating to ourselves
link |
and then ultimately folks in the regulatory role
link |
and then ultimately the public
link |
that we have confidence that our system
link |
will work better than that.
link |
And so these individual metrics
link |
will kind of tell a compelling story ultimately.
link |
I do think at the end of the day
link |
what we care about in terms of safety
link |
is life saved and injuries reduced.
link |
And then ultimately kind of casualty dollars
link |
that people aren't having to pay to get their car fixed.
link |
And I do think that in aviation
link |
they look at a kind of an event pyramid
link |
where a crash is at the top of that
link |
and that's the worst event obviously
link |
and then there's injuries and near miss events and whatnot
link |
and violation of operating procedures
link |
and you kind of build a statistical model
link |
of the relevance of the low severity things
link |
or the high severity things.
link |
And I think that's something
link |
where we'll be able to look at as well
link |
because an event per 85 million miles
link |
is statistically a difficult thing
link |
even at the scale of the U.S.
link |
to kind of compare directly.
link |
And that event fatality that's connected
link |
to an autonomous vehicle is significantly
link |
at least currently magnified
link |
in the amount of attention it gets.
link |
So that speaks to public perception.
link |
I think the most popular topic
link |
about autonomous vehicles in the public
link |
is the trolley problem formulation, right?
link |
Which has, let's not get into that too much
link |
but is misguided in many ways.
link |
But it speaks to the fact that people are grappling
link |
with this idea of giving control over to a machine.
link |
So how do you win the hearts and minds of the people
link |
that autonomy is something that could be a part
link |
I think you let them experience it, right?
link |
I think it's right.
link |
I think people should be skeptical.
link |
I think people should ask questions.
link |
I think they should doubt
link |
because this is something new and different.
link |
They haven't touched it yet.
link |
And I think that's perfectly reasonable.
link |
And, but at the same time,
link |
it's clear there's an opportunity to make the road safer.
link |
It's clear that we can improve access to mobility.
link |
It's clear that we can reduce the cost of mobility.
link |
And that once people try that
link |
and understand that it's safe
link |
and are able to use in their daily lives,
link |
I think it's one of these things
link |
that will just be obvious.
link |
And I've seen this practically in demonstrations
link |
that I've given where I've had people come in
link |
and they're very skeptical.
link |
Again, in a vehicle, my favorite one
link |
is taking somebody out on the freeway
link |
and we're on the 101 driving at 65 miles an hour.
link |
And after 10 minutes, they kind of turn and ask,
link |
is that all it does?
link |
And you're like, it's a self driving car.
link |
I'm not sure exactly what you thought it would do, right?
link |
But it becomes mundane,
link |
which is exactly what you want a technology
link |
like this to be, right?
link |
We don't really, when I turn the light switch on in here,
link |
I don't think about the complexity of those electrons
link |
being pushed down a wire from wherever it was
link |
and being generated.
link |
It's like, I just get annoyed if it doesn't work, right?
link |
And what I value is the fact
link |
that I can do other things in this space.
link |
I can see my colleagues.
link |
I can read stuff on a paper.
link |
I can not be afraid of the dark.
link |
And I think that's what we want this technology to be like
link |
is it's in the background
link |
and people get to have those life experiences
link |
So putting this technology in the hands of people
link |
speaks to scale of deployment, right?
link |
So what do you think the dreaded question about the future
link |
because nobody can predict the future,
link |
but just maybe speak poetically
link |
about when do you think we'll see a large scale deployment
link |
of autonomous vehicles, 10,000, those kinds of numbers?
link |
We'll see that within 10 years.
link |
I'm pretty confident.
link |
What's an impressive scale?
link |
What moment, so you've done the DARPA challenge
link |
where there's one vehicle.
link |
At which moment does it become, wow, this is serious scale?
link |
So I think the moment it gets serious
link |
is when we really do have a driverless vehicle
link |
operating on public roads
link |
and that we can do that kind of continuously.
link |
Without a safety driver.
link |
Without a safety driver in the vehicle.
link |
I think at that moment,
link |
we've kind of crossed the zero to one threshold.
link |
And then it is about how do we continue to scale that?
link |
How do we build the right business models?
link |
How do we build the right customer experience around it
link |
so that it is actually a useful product out in the world?
link |
And I think that is really,
link |
at that point it moves from
link |
what is this kind of mixed science engineering project
link |
into engineering and commercialization
link |
and really starting to deliver on the value
link |
that we all see here and actually making that real in the world.
link |
What do you think that deployment looks like?
link |
Where do we first see the inkling of no safety driver,
link |
one or two cars here and there?
link |
Is it on the highway?
link |
Is it in specific routes in the urban environment?
link |
I think it's gonna be urban, suburban type environments.
link |
Yeah, with Aurora, when we thought about how to tackle this,
link |
it was kind of in vogue to think about trucking
link |
as opposed to urban driving.
link |
And again, the human intuition around this
link |
is that freeways are easier to drive on
link |
because everybody's kind of going in the same direction
link |
and lanes are a little wider, et cetera.
link |
And I think that that intuition is pretty good,
link |
except we don't really care about most of the time.
link |
We care about all of the time.
link |
And when you're driving on a freeway with a truck,
link |
say 70 miles an hour,
link |
and you've got 70,000 pound load with you,
link |
that's just an incredible amount of kinetic energy.
link |
And so when that goes wrong, it goes really wrong.
link |
And those challenges that you see occur more rarely,
link |
so you don't get to learn as quickly.
link |
And they're incrementally more difficult than urban driving,
link |
but they're not easier than urban driving.
link |
And so I think this happens in moderate speed
link |
urban environments because if two vehicles crash
link |
at 25 miles per hour, it's not good,
link |
but probably everybody walks away.
link |
And those events where there's the possibility
link |
for that occurring happen frequently.
link |
So we get to learn more rapidly.
link |
We get to do that with lower risk for everyone.
link |
And then we can deliver value to people
link |
that need to get from one place to another.
link |
And once we've got that solved,
link |
then the freeway driving part of this just falls out.
link |
But we're able to learn more safely,
link |
more quickly in the urban environment.
link |
So 10 years and then scale 20, 30 year,
link |
who knows if a sufficiently compelling experience
link |
is created, it could be faster and slower.
link |
Do you think there could be breakthroughs
link |
and what kind of breakthroughs might there be
link |
that completely change that timeline?
link |
Again, not only am I asking you to predict the future,
link |
I'm asking you to predict breakthroughs
link |
that haven't happened yet.
link |
So what's the, I think another way to ask that
link |
would be if I could wave a magic wand,
link |
what part of the system would I make work today
link |
to accelerate it as quickly as possible?
link |
Don't say infrastructure, please don't say infrastructure.
link |
No, it's definitely not infrastructure.
link |
It's really that perception forecasting capability.
link |
So if tomorrow you could give me a perfect model
link |
of what's happened, what is happening
link |
and what will happen for the next five seconds
link |
around a vehicle on the roadway,
link |
that would accelerate things pretty dramatically.
link |
Are you, in terms of staying up at night,
link |
are you mostly bothered by cars, pedestrians or cyclists?
link |
So I worry most about the vulnerable road users
link |
about the combination of cyclists and cars, right?
link |
Or cyclists and pedestrians because they're not in armor.
link |
The cars, they're bigger, they've got protection
link |
for the people and so the ultimate risk is lower there.
link |
Whereas a pedestrian or a cyclist,
link |
they're out on the road and they don't have any protection
link |
and so we need to pay extra attention to that.
link |
Do you think about a very difficult technical challenge
link |
of the fact that pedestrians,
link |
if you try to protect pedestrians
link |
by being careful and slow, they'll take advantage of that.
link |
So the game theoretic dance, does that worry you
link |
of how, from a technical perspective, how we solve that?
link |
Because as humans, the way we solve that
link |
is kind of nudge our way through the pedestrians
link |
which doesn't feel, from a technical perspective,
link |
as a appropriate algorithm.
link |
But do you think about how we solve that problem?
link |
Yeah, I think there's two different concepts there.
link |
So one is, am I worried that because these vehicles
link |
are self driving, people will kind of step in the road
link |
and take advantage of them?
link |
And I've heard this and I don't really believe it
link |
because if I'm driving down the road
link |
and somebody steps in front of me, I'm going to stop.
link |
Even if I'm annoyed, I'm not gonna just drive
link |
through a person stood in the road.
link |
And so I think today people can take advantage of this
link |
and you do see some people do it.
link |
I guess there's an incremental risk
link |
because maybe they have lower confidence
link |
that I'm gonna see them than they might have
link |
for an automated vehicle and so maybe that shifts
link |
But I think people don't wanna get hit by cars.
link |
And so I think that I'm not that worried
link |
about people walking out of the 101
link |
and creating chaos more than they would today.
link |
Regarding kind of the nudging through a big stream
link |
of pedestrians leaving a concert or something,
link |
I think that is further down the technology pipeline.
link |
I think that you're right, that's tricky.
link |
I don't think it's necessarily,
link |
I think the algorithm people use for this is pretty simple.
link |
It's kind of just move forward slowly
link |
and if somebody's really close then stop.
link |
And I think that that probably can be replicated
link |
pretty easily and particularly given that
link |
you don't do this at 30 miles an hour,
link |
you do it at one, that even in those situations
link |
the risk is relatively minimal.
link |
But it's not something we're thinking about
link |
in any serious way.
link |
And probably that's less an algorithm problem
link |
and more creating a human experience.
link |
So the HCI people that create a visual display
link |
that you're pleasantly as a pedestrian
link |
nudged out of the way, that's an experience problem,
link |
not an algorithm problem.
link |
Who's the main competitor to Aurora today?
link |
And how do you outcompete them in the long run?
link |
So we really focus a lot on what we're doing here.
link |
I think that, I've said this a few times,
link |
that this is a huge difficult problem
link |
and it's great that a bunch of companies are tackling it
link |
because I think it's so important for society
link |
that somebody gets there.
link |
So we don't spend a whole lot of time
link |
thinking tactically about who's out there
link |
and how do we beat that person individually.
link |
What are we trying to do to go faster ultimately?
link |
Well part of it is the leadership team we have
link |
has got pretty tremendous experience.
link |
And so we kind of understand the landscape
link |
and understand where the cul de sacs are to some degree
link |
and we try and avoid those.
link |
I think there's a part of it,
link |
just this great team we've built.
link |
People, this is a technology and a company
link |
that people believe in the mission of
link |
and so it allows us to attract
link |
just awesome people to go work.
link |
We've got a culture I think that people appreciate
link |
that allows them to focus,
link |
allows them to really spend time solving problems.
link |
And I think that keeps them energized.
link |
And then we've invested hard,
link |
invested heavily in the infrastructure
link |
and architectures that we think will ultimately accelerate us.
link |
So because of the folks we're able to bring in early on,
link |
because of the great investors we have,
link |
we don't spend all of our time doing demos
link |
and kind of leaping from one demo to the next.
link |
We've been given the freedom to invest in
link |
infrastructure to do machine learning,
link |
infrastructure to pull data from our on road testing,
link |
infrastructure to use that to accelerate engineering.
link |
And I think that early investment
link |
and continuing investment in those kind of tools
link |
will ultimately allow us to accelerate
link |
and do something pretty incredible.
link |
Chris, beautifully put.
link |
It's a good place to end.
link |
Thank you so much for talking today.
link |
Thank you very much. Really enjoyed it.