Good Disruption: Autonomous Vehicles
How close are we to fully automated vehicles and how will transportation itself change as a result?
00:00 / 41:50
Gen Xers will remember Knight Rider, a television show in the 1980s about a crime fighter who battles bad guys with the help of a technologically advanced Pontiac Firebird Trans Am that drives itself and talks, all powered by artificial intelligence. At that time, the idea of a self-driving car was futuristic; but now 40 years later, some elements of autonomy are available in passenger vehicles as a standard offering. Just how close are we to fully automated vehicles and how will transportation itself change as a result?
Mike and Yael talk with Will Shepherdson, Darden alum (MBA’21) Product Manager at Waymo, and Madhur Behl, an Associate Professor in Computer Science at UVA. Waymo operates autonomous ride-hailing services in Phoenix, San Francisco, Los Angeles, and Austin. Professor Behl conducts research at the confluence of Machine Learning, Predictive Control, and Artificial Intelligence with applications in Cyber-Physical Systems, Autonomous Systems, Robotics, and Smart Cities.
Transcript
The Disruption, a lively discussion between UVA Darden School of Business professors Yael Grushka Cocaine and Mike Lennox on cutting edge technologies and practices that are challenging the status quo.
Hey Mike.
Hey, how are you doing today?
I am, I am great actually.
This time of year is always a wonderful time of year and my, my birthday is coming up, so it always makes me happy.
Well, a few weeks away, but I, I start a long time in advance.
You're 35th.
Yes, exactly.
Let's just say that.
Yeah, let's go with that.
Let's go, um, you know, I love movies.
Have we talked about movies in this podcast before?
I think we have a little bit, a little bit, just with AI and entertainment and deep fakes.
That's come up a number of times now.
So one of my, uh, uh.
Um, not in my top 30 movies, but you know, up there is a Total Recall.
Do you remember Total Recall?
I do.
That's a Schwarzenegger vehicle.
That's a great, great movie.
Yeah.
So it always, um, every time we come to these podcasts, I think about that movie because it feels like, you know, the movie came out in the 90s, early 90s, and it feels like it already predicted half of what we talk about in this podcast.
So I thought you were gonna say I remind you of Arnold Schwarzenegger, but that's OK.
OK, if I can be Sharon Stone, you can be Arnold Schwarzenegger.
How about that?
Thank goodness for AI and deep fakes.
Well, today is a topic that I know is, um, uh, really of interest to you, um, and that is of, um, autonomous vehicles, vehicles that don't require us to drive around.
Um, I know that you're passionate and interested and fascinated by this.
Um, uh, is that true?
Do you feel still that still feel that way about the topic?
I do.
You know, it's funny when we started this podcast, um, like a year and a half ago now, we had talked about this as being one of our.
Topics and we did another related topic that also is a passion of mine which is electric vehicles and what I find fascinating is this is an industry that arguably has two major disruptions.
Uh, now as we talked about with electric vehicles, I think that one's a little more advanced in terms of what we're seeing in the marketplace and how it's transforming the market.
But be clear to our listeners, this autonomous vehicle future is also moving quickly and, and arguably kind of on the horizon here.
And, uh, and it's interesting to think about them, um, and I love that you brought up the idea of the EVs and, and the fact that we're now talking about autonomous vehicles today. um, is there a reason why they're so intertwined?
Is it the electric part of things that is making autonomous, uh. possible.
Is there any link between those advancements or not at all?
Yeah, I mean, 11 could imagine them being decoupled, right, that you could have an autonomous vehicle that's an internal combustion engine.
There's, there's no reason not.
However, there are some complementarities here that if you're going to design a navy, especially one from scratch, you, you probably would want to use an electric vehicle.
Um, most of the electric vehicles are heavily software-based, as we've talked.
About before lending itself obviously to AV and uh you can think about things like charging, you know, you, you could imagine having an autonomous, uh, you know, gas pump it's possible, of course, but it would be a lot easier to think about uh autonomous charging in an electric vehicle.
OK, fascinating.
So we're talking again about vehicles transportation, um, but the implications here of autonomous vehicles and autonomous driving.
Um, is, is just in some ways very different from the electric vehicles kind of industry.
There's just so many implications for society, so many implications for the economy, so many implications for the environment, uh, that are really worth this discussion, I think, uh, in, in detail.
Yeah, no, and I think that's as a business strategist.
What fascinates me about this is this can actually totally restructure the way we consume transportation services.
Um, this is not just a change in the kind of form factor or drive train of the traditional way we consume cars.
This pretends that it could, it could be quite, quite dramatic and, and.
Uh, really, uh, turn over the industry in, in kind of interesting and unique ways, but it's not only how we, uh, move around and, and our, our mode of transportation, but this, I've heard visionaries, and this was a while back, which I'm gonna come back to the fact that the visions have been out, yes, exactly, total recall, right?
So a while back I heard the, uh, uh, folks describing not only the, uh, impact on how we get from place A to place B, but this affects who takes the.
The kids to school who, you know, how we spend our time, this affects where we park our cars.
This affects how we build our buildings, the parking lots, the insurance companies that we need to pay.
Like there's so many derivative and auxiliary effects that, um, that are just mind boggling that we're talking about completely different way to live our lives.
Is that, is that the case?
That is.
And you know, but let's take a step back first.
Let's, let's talk a little bit about what, where we are and where the technology, uh, hopefully may, may go here.
Um, so, uh, there, there's levels of autonomy that we talk about and arguably we, we already have quite a bit of autonomy in many of our vehicles.
You think about driver assist, various forms of cruise control, vehicles that will automatically parallel park.
These are all forms of autonomy.
But the, the gold standard, if you will, where, where, you know, we, we'd like to see it go is what we call level 5 autonomy and. 5 autonomous vehicle.
I, I like to joke and the simplest way to think about it is, it's the one I can fall asleep in the car behind the wheel or maybe no wheel at all, and, and feel safe and be, be taken care of.
And the fact is, again, we're getting closer and closer to that, but we're not quite there yet.
We're not quite to that point where we would trust the vehicle in that sense, despite some YouTube videos of people doing this like in their Teslas down the highway and the like.
Um, so what are the main challenges with getting us to that level 5?
Well, as you know well, I mean, at the heart of this, this is, this is AI again, you know, once again we're talking about artificial intelligence and the way these vehicles are trained is on these massive data sets, and the way to think about those is they're trying to train on all the different contingencies and problems that you might face while you're driving.
The challenge is, uh, kind of the classic 80/20 rule, though I always like to joke with autonomous vehicles, it might be more like 95/5 at this point in the sense that for 80% or even 95% of your driving situations, we can actually do really well with autonomy.
But that last percent is incredibly difficult, uh, different weather conditions when it rains or it snows, um, just those kind of random events where, you know, someone darts in front of the street or a traffic pattern is, uh, not what you'd normally expect.
That's the hard part, like that's the hard part to get this to that full level 5 autonomy where we would just trust these vehicles to drive us around.
So that sounds to me like you're describing more of a software challenge, a data training challenge that we've talked in, as you mentioned in AI, but what about the rest?
I mean, autonomous vehicles, I've heard of many different technologies that are hardware, you know, the, the sensors, the cameras, uh, the different components that sit above, beyond and back of the car that help us kind of really, uh, maybe translate some of the, the world around us, pick up on what's going on around the vehicle.
Is the hardware solved?
Is that part of the equation already done?
No, in fact, there's some, you know, critical questions still to be resolved, though, you know, I imagine some might think it already, they know the answer.
But, um, obviously it's a mix of various types of sensor, sensors, radar, just cameras on the car.
One of the critical questions is the LDAR, uh, which is this, uh, advanced technology for basically seeing the environment around the car.
Uh, most, I think, of those developing level 5 autonomy or moving towards level 5 have see LiDAR as a critical component of the vehicle, um, and that's when you see cars, uh, maybe you've seen some of these around the road that have like the bubble on the top of the car, that's usually the LDAR system.
Uh, Tesla in particular has, and Elon Musk in particular has been adamant that you don't need LiDAR.
You can, you can solve this with enough data using just simple sensors, simple cameras, and the like.
Uh, and there are debates, let's say, and not everyone agrees with Musk that you can do this with just, um, without a LiDAR system.
Well, it's interesting that you bring up Tesla, um, and, and I know that we're gonna, you know, the name Waymo will come up in our conversation for sure today, um, as a player in the field, so we're seeing a lot of interest from, um, maybe pure data companies like Google, the tech companies, the tech giants, um, uh, and maybe Tesla is kind of representing those, those, uh, you know, vehicle, maybe it's more of a it's not even a traditional, uh, uh.
A car manufacturing company.
But what about the more traditional car manufacturers, the Fords of the world, the, you know, the BMWs of the world, are they playing in this space?
They're definitely playing in the electric vehicle space, but are they interested in the autonomous vehicle development?
Yeah, it's been really interesting.
I think all of them are investing at some level again for these kind of advanced driver assist features, so especially the luxury car manufacturers, but others as well.
How they're approaching kind of true level 5 autonomy, uh, varies.
So one of the big moves taken in the industry was a few years back when General Motors acquired Cruze, who is a startup in this space.
So Cruze is in essence the autonomous division for, uh, General Motors.
Uh, obviously you have Tesla, as you mentioned, who really from day one has been talking about autonomy, and they continue to talk about it on their earnings calls and the like.
And then of course you mentioned, you know, the big tech companies, so, uh, Waymo spun out of, you know, Google's famous, uh, Xlab, um, you have Amazon getting into the space with Nook which was an acquisition that they made, uh, Apple is, you know, rumored.
Continuously to be, you know, getting into this space that more recently they said they're pulling back from their, uh, autonomous, uh, efforts there, but who knows if Apple will jump in there, um, so it is attracting a lot of attention from a, a kind of a wide variety of different players, um, so here at Darden we have a lot of efforts that are, um, focused on, um, uh, talking about responsible AI or leadership, ethical leadership in the, in the, in the world of AI.
Let's talk about ethical leadership in the world of um autonomous vehicles and in this disruption I guess one concern that people may have which could cause them to label it bad disruption would be this whole notion of, you know, giving over the total control to to the machines whether phys hardware or software and then catastrophic outcome.
Like accidents happening.
So how do we think about that?
How, how do we get our heads around, uh, acquiring, uh, a whole fleet of these, uh, cars for our business?
Yeah, I, I mean, I'm of a couple minds with this, right?
Uh, we've talked before, whenever a new technology comes by, there are often risks associated with it that, um, you know, cause consumers and others to be resistant to the new technology.
Um, that is not surprising in this case, obviously, given the nature of the, you know, of what these cars are doing.
Um, at the same time, you know, progress is often made, opinions change, and then you, you get more comfortable with the new technology, and the enthusiasts for AVs would argue that AVs will be actually safer than any, you know, real world scenario where you and I are driving our vehicles on the road there.
So that this again, the future will be an accident-free future once we're all using autonomous vehicles there.
Um, I do think there are some critical issues.
Um, are you aware of the trolley problem?
I am aware of the trolley problem.
Why don't you explain the trolley problem for our, for our listeners?
The trolley problem, trolley problem, um, way back when I studied programming, uh, uh, was always presented as, as a, a, a real challenge there when you have to hard code, uh, rules and, and make, you know, description of situations for, for software to basically make a decision and to decide.
To take Route A or Route B, you almost have to program in certain difficult pathways and different, uh, forks in the road, and here the trolley problem is quite literally which route do you take, uh, when you're faced with two very unwanted outcomes, um, you know, yeah, exactly.
I mean this is a classic moral dilemma question.
You're going down a, a, a hill in a trolley, the brakes give out, your choices are to continue forward and plow into let's.
A group of people or you could divert, uh, and maybe kill everyone on the trolley.
It was very morbid, morbid, an old lady crossing the road, yeah, a baby in a stroke, right?
And but the critical thing here is that the AV is going to have to make these decisions and right now it falls to you and I as individual drivers in the moment to make these decisions, and we have a whole, you know, legal apparatus around liability and the like, uh, that assesses whether we made maybe the right or wrong decision.
Um, that's gonna have to be hard coded in and, uh, maybe it will have to be regulatory that tells the, you know, the AV makers what they should do in certain circumstances.
Maybe it will fall to the AV manufacturers.
Lots of questions about liability, um, you mentioned insurance before, you know, who gets insured in an, in an AV?
Uh, is it the, the rider now as opposed to the driver?
Um, and so these are just some critical issues that are gonna need to be resolved, uh, before we, we truly find that AV future, you know, it's interesting about trust and how do we build trust, uh, um, and the, and, and some of the challenges for this industry to evolve.
Uh, my son is 17 and he just, uh, got his, uh, uh, license, uh, just a few weeks ago.
God bless you.
Yes, thanks, exactly.
This is exactly my point.
I have.
Not slept for about a month now since he got his license, um, and the thought around do, um, and by the way he's a great driver and I'm very proud of him, um, and he's been doing a great job, uh, but it is nerve-wracking and to trust a 17 year old to take his thirteen-year-old sister, you know, my my daughter to, to school, um, that's a big jump and do I prefer that a machine does it?
I don't know, it's a good dilemma, right?
Um, yeah.
Um, I, I remember when I was teaching my daughter to drive a few years back that like saying to her like she might be the last generation that learns how to drive.
Like, will she be teaching her kids how to drive, or again will we all be buzzing around in our, uh, on our AVs?
There's another industry that will be disrupted eventually, maybe the whole, uh, uh, driving lesson.
Yeah, well, hey, let's, let's talk a little bit about, um, the business side of this because.
Because what's interesting about AVs is that at the end of the day the whole business model could change, uh, and in fact companies like Waymo, uh, are moving in as ride sharing companies, so they're really displacing, if you will, the Ubers and the Lyfts of the world, um, but again could change just the way we even consume cars.
Do we even own vehicles in a, uh, in a fully level 5 AV future?
Yeah, I think this is a time for us to, uh, bring in people with a little bit more knowledge than us.
What do you think?
I think it's always good, uh, to bring in some others with some knowledge here.
So I have a real treat for us today, A.
For the first time in Good Disruption podcast history.
I love we innovate ourselves.
We do.
We actually have two experts.
This topic is too big for just one expert.
We have two experts with us.
Uh, so I'm very pleased that we have, uh, Will Shepherdson here, a former student of both of ours, a Darden MBA, uh, and is a product manager at Waymo, who we had mentioned, one of the leaders here, uh, in developing autonomous vehicles, and then we also have Maher Bijel, who is an associate professor of computer science at UVA, uh, who does research, uh, on machine learning, AI, but also on.
Autonomy, uh, and does work on the robotics and small cars that, you know, we're talking about here so we're very excited to have, uh, both of them with us here.
It's fantastic.
Both academia and, uh, and the business world are represented.
So I, I'm, I'm excited.
I have so many questions.
Yes, well, uh, Ma, I'd like to start with you, uh, given your expertise and your work here.
Can you give us any sense, level 5?
Like how close are we here?
I, I, I always joke like Elon Musk said, you know, we, we'd have it like 2 years ago and he keeps saying every year we're we're, we're there.
What's your, what's your sense?
How, how close are we to this level 5 dream?
Yeah, I think you, you, you are right in assessing that most of the predictions are, are very hyperbole, in my opinion.
This is a very, very difficult problem, and I heard you talking about this earlier, that this is really a problem in, in AI, right?
Because this is a very smart robot, essentially, which has to understand how the world operates and be able to deal with these uh all possible situations that could arise.
And as flawed as humans might be, we, we are surprisingly good at generalizing the skill that we have to, to different situations.
That's sort of the, you know, the remarkable ability that humans, humans possess.
Like, so for example, your, your son who just got his license.
Uh, he will learn to drive in maybe, you know, about 3 or 4 days or maybe a week at most, and, uh, he'll improve throughout, throughout his lifetime as he drives more.
Uh, but, but he has the ability to generalize that, that experience, right?
So I was actually surprised when I, when I came to the US, uh, many years ago, I got a license, and then, uh, I learned that I could drive a U-Haul truck with the same license, right?
So because we trust that driving.
And analyzing how the world works, getting a sense for what different situations arise, how to navigate in the presence of other vehicles, pedestrians, cyclists.
Uh, we trust the human intellect and this generalized intelligence to, to behave correctly.
And this is really where the, you know, level 5 is, is very hard.
And just to clarify, level 5 and level 4, they are nearest neighbors.
Level 4 is when the human does not have to intervene at all.
So it could be completely hands-free, and you could still sort of sleep, and you're not required to intervene.
Uh, you know, if you can sleep, uh, without being anxious, that is.
But, uh, but the difference between level 4 and level 5 is that level 5 really means that This car, which is already human free, uh, and we've had many of these operate in a few selected cities in the US.
Uh, level 5 means that they could drive anywhere under any conditions not restricted to a couple of neighborhoods or just in, you know, certain, uh, cities, and not just in the sunny parts of the country, but, uh, under fog, under rain, any kind of, uh, uh, conditions.
That's the level 5.
So, so this is precisely why level 5 is way, way out there in the future, like getting to level 5.
Uh, autonomous driving, in my opinion, would really mean that we have had a breakthrough in generalized artificial intelligence, because I would say we, we don't even have generalized driving intelligence right now.
We can all agree that driving is a very intelligent task, but, uh, you know, these cars, the software on these cars, it cannot work on a different hardware, just like a human could drive a sedan or SUV.
Um, so that's something fundamental is missing in that sense.
And it's a very difficult problem because at the end of the day, it just boils down to, uh, uh, this, this sort of uh word, which one word which could describe why this is such a technical problem is it's a problem of coverage, right?
So the, the In the real world, you have uncountable number of objects in all shades, colors, and varieties that could appear in the scene in front of you.
You have different environments, um, which we've sort of, you know, it's more beyond just, you know, lighting environments and, and, and weather.
Um, and then you also have the interplay of both of these, right?
So how these different entities in the world around you interact with each other is again, some infinite long tale of possible situations, and no engineer or group of engineers could adequately describe what these situations are a priori.
Um, fascinating.
Um, it, it's so that has helped so much to kind of, uh, make it clear to folks exactly where the challenges lie.
Um, will, on the business side, how do companies kind of, how do you as a product manager on the business side cope with some of this challenge and some of these, uh, um, uh, real hurdles that, that we face still in the advancement?
Yeah, um, and just for sort of context, uh, you know, Waymo's mission is to be the most trusted driver.
And, um, you know, I think sort of safety is at the, um, the center of everything that we do.
You know, our goal is to reduce sort of traffic injuries and fatalities.
In the cities in which we are operating and now we have sort of quite a bit of experience and history working towards that goal.
We actually operate, you know, a fully autonomous ride service in 4 different cities right now San Francisco, Phoenix, Los Angeles, and Austin.
Um, and, and I think, you know, to, to the point about sort of L4 versus L5, um, you know, I think the definition of L5 is somewhat unconstrained, like it could be, uh, you know, a dirt road, an unpaved dirt road in the middle of nowhere that someone drives, you know, every once in a while, and I, and I think that from like a if our goal is like reducing traffic injuries and fatalities, a lot of those tend to be concentrated.
In cities and like we think there is both like an opportunity to meet this goal of improving safety on the roads and also sort of build a business and offer a product that our riders want to use so you know we have.
You know, uh, like a lot of enthusiasm.
We've been offering a commercial service in San Francisco since the mid of August in 2023, um, same in Phoenix and recently expanded in Los Angeles and Austin.
I think that like with an L4 product, um.
You know, our goal is to build an L4 autonomous driving system and operate a service that, you know, brings people from point A to point B safely.
I think the Waymo driver, you know, doesn't drive distracted or drunk.
So, um, you know, we, we, you know, think that there are sort of like real safety benefits.
We like believe, obviously like believe that there's real safety benefits to.
You know, operating the Waymo driver, and I think that some of the data actually indicates this.
So there's a couple of research partnerships I can tell about if that's interesting, um, but you know, one statistic is that we actually experienced like 85% fewer crashes that led to injuries than comparable human driving benchmarks over about 7 million miles of fully autonomous driving.
So this train, yeah fascinating.
Well, it also is interesting to think through like what is the correct logic for when these cars, and they already are on the roads, but come on the roads.
Is it 100%, you know, uh, no error rate, or is there some exception because again, let's point it out, there are millions of car accidents on the road now.
So, you know, maybe the baseline should be improvement over where we currently are, um, with these vehicles, but I liked putting together the two comments from, from both of you.
Uh, level 5 is gonna be really hard and it's gonna require a big improvement and developments in, in how we do things.
It's worthy of working towards, but at the same time, um, what maybe is, is exciting to hear about is how companies strategize on doing more local and.
And limb and and more narrow kind of investments to work within the L4 kind of opportunities is that is that that that fair?
Yeah, no, and I'm actually, will, a question for you, um, in the offhand, let's say, uh, a snowstorm hits San Francisco.
I don't know if that ever actually happens, but let's just say for the sake of argument that it does.
Do you, I assume you have like protocols then that maybe you shut the fleet down or how, how do you handle maybe, um, some of those outlying cases?
Yeah, so I guess one of the like common misconceptions is I think Wayma for a long time operated in Phoenix and like the suburbs like Phoenix greater metropolitan area, and I think people are like, wow, there's never any weather in Phoenix.
Like people have sort of gone there because it's easy and, and I, and I, you know, I think that that is maybe a little bit of a simplification.
Like Phoenix actually does get like severe storms in the summer and like it gets weather exposure.
It's sort of not sunny and beautiful all the time.
So I think that like, you know, even cities that have relatively mild climates experience weather conditions.
Waymo does a huge amount of sort of testing and evaluation and validation to operate in these more rare, uh, you know, conditions, and I, and I can actually talk specifically about weather because that's something that, you know, I work on at Waymo, which is that we have done a lot of intentional work around validating our performance in more rare weather conditions that we experience in the cities that we're operating in.
We also do sort of excursions and field trips to other geographies that have more extreme weather.
So in the past we have driven in, we've experienced like winter weather from Some of these trips that we've taken to New York City, Bellevue, Washington this past winter, we are operating a small fleet of vehicles in Buffalo, New York, and we actually have an engineering center in Michigan.
So we're able to like take the Waymo driver to these new geos and improve our learning about how does the Waymo driver perform in those conditions.
We can take like the relatively small amount of driving that we do on.
Public roads that's complemented with lab and sort of chamber testing that's done at like a sensor level and at a vehicle level.
We also do closed course testing in extreme weather conditions and then closed course.
Oh, and then and then I need to mention sort of simulation.
So like simulation plays a large role in our development process and validation of the Waymo driver.
So we're able to re-simulate our experience.
And sort of continue to learn from it without having to drive more real world miles.
We drive far more miles in simulation than in the real world and use that to sort of gain confidence about the performance of the limo driver in these kind of rare conditions.
So I'd say like we actually know quite a bit about these extreme weather events both because of like accumulated experience over a number of years, but a lot of targeted testing and validation work that we've done.
My mom likes to remind me that I was 3 years old living in Buffalo during the blizzard of '77, so that was quite an extreme event that I don't know quite has happened since, but um, I wanna actually kind of pivot to, to Mader and here, um, we heard from Will how, how Waymo learn and, and train and get more exposure.
Can you tell us on the academic side, you've, you've shared with me a little bit that you guys have a, a, a racing team.
Like, can you tell me how the academic world pushes the needle on advancements here?
Yeah, for sure.
Uh, uh, so yeah, I do, uh, lead the university's autonomous racing team.
So, um, you know, on a, on a good day, any self-driving car would go, you know, the speed limit, uh, or, uh, but here we have a full-scale indie racing car, uh, with no driver, there's no steering wheel.
So it's, it's not level 4 quite because it's not comparable to running in the streets, but Um, we've been, we've been able to run this car at 150, 160 MPH, completely autonomously and overtaking other, other competitors.
Um, the idea being that, you know, uh, in academia, we try to explore the limits of any technology.
And so, my crazy hypothesis was that, um, just like motorsport racing has always served as some sort of proving grounds for regular automotive technology, um, you know, Any, any kind of, uh, any kind of improvement you have in your car, you can trace its origin to some innovation in motorsports.
In fact, that's why motorsport was invented, uh, because, um, back in the day when we were transitioning from horse-driven carriages to horseless carriages, there was the same skepticism.
That this new tech technology is not going to be safe.
And so motorsport became like a way to, to demonstrate the performance and safety.
So, now we are going in the beginning of the same transition.
We are going from driver-driven carriages to driverless carriages.
So why not use the same principles and pit these cars in extreme conditions intentionally against each other.
So, put them to the limits of what is, what breaks them, so that we can collectively learn in advance.
And that's why.
Uh, you know, racing is an extreme version of it, but, uh, broadly speaking, uh, if you go to any top machine learning, uh, robotics conference in the last 10 years, there are dedicated sessions on How do you predict better trajectories?
How do you handle pedestrians' behaviors better?
Because this is where these cars seem to be struggling.
They either fail to just realize what the world looks like, because there's so many things to detect and it can get confused, uh, or it fails to be clairvoyant and predict the next few seconds in order for it to plan its correct course of action.
So, academia has, has worked in this problem space for, for many, many years.
The entire self-driving car industry sort of started from an academic experiment with the DARPA Grand Challenges, and that collaborative spirit still continues, although the industry has taken the lead in, uh, you know, deploying them in the real world and learning lessons.
Based on the mistakes that they, that they encounter, you know, getting to the motor sport thing, uh, you know, I'm, I'm one of those watchers that like kind of maybe enjoys the accidents every once in a while.
Of course you don't want them to be like fatal or anything like that, but I, I don't know if we're gonna eliminate that if we don't have accidents with these, uh, matter.
I was curious, you know, Will had mentioned the simulation piece, which I find fascinating.
What, what are your thoughts as a computer scientist on this idea?
Of training our AIs using a simulated environment.
My understanding, and Waymo is, you know, millions of miles on the road, but billions of miles in the simulation.
So, so matter, what, how do you think about this idea of training AI on simulations?
I think it's a good, good idea and one which is, uh, you know, widely acknowledged that should be part of any, uh, good development cycle.
Uh, essentially, you know, if you, if you just wait for the car to rack up enough miles to generate meaningful data for it to be comparable to human-level error, then peer reviewed studies indicate that you'll have to wait for, you know, some 250 million miles per car or something crazy number, which will take decades for, um, for us to achieve.
But you could, um, you could drive them, as you pointed out, in simulation, at scale, and you could also vary, um, different sort of conditions that, that you may encounter.
So this is definitely something which is a good overall thing, uh, and it can help you improve how the car behaves in the real world.
The challenge with the, with simulation, which is also well, well studied and well known, is that there's always going to be a gap.
Uh, which is called the sim to real gap between no matter how sophisticated your, your simulator is, it either will have a gap in the physics or maybe it doesn't simulate the sensor level physics that will.
Um, versus, you know, what, what, what the real world is, which is very, very analog, and you cannot really, I said earlier, have the, the foresight to, to seed the simulation in a manner that will give you the confidence that no matter what happens in the real world, uh, it will make the safe decision.
Now, I think it's reasonable to say that, you know, there shouldn't be, uh, there shouldn't be some expectation that these vehicles are never supposed to crash.
The trouble is that it's not just the number of crashes, which are indicative of the safety of these vehicles.
It's sometimes also the kind of crashes that the AI makes that no reasonable human would make.
Now, I know we have a mix of bad drivers, distracted.
Drivers, but we also have good drivers.
And so these cars have to level up and be as good, if not better than most good drivers who are not distracted, who are not drunk when they drive.
And some of the mistakes, which we don't have much transparency into, unfortunately, um, they seem to suggest that, you know, these are some common sense errors, which, which humans wouldn't make.
Like if you, if you hit a pedestrian, there's no way you're going to drag the pedestrian another 20 ft.
Uh, so, so, so that, that again. necessitates that some uh real world testing is required, but, uh, you know, it opens, uh, throws open this, this debate of uh whether, you know, what, what is that fine line?
Uh, because the concern is that the pace of deployment might exceed.
Our ability to test these vehicles for safety.
Yeah, and so I want to pivot to Will.
Uh, we've talked a lot about safety, and I can totally see the need to, uh, and any improvement in safety, quite frankly, will all benefit as a society given that accidents, uh, happen too, too frequently, and, and I, and I hear you maybe not a reduction to zero accidents, uh, my dear, but, but some reduction, but.
Will, you're a Darden MBA.
You have all your, your multifaceted kind of perspectives.
What about the sustainability, environmental benefits?
Like, are there other benefits that we're, we haven't mentioned, or on the flip side, other concerns, like, you know, some kind of jobs being lost or, or something that we should be thinking about holistically in the conversation around autonomous vehicles?
Yeah, I guess, I mean, you mentioned earlier that like one of your first, I don't know if it was the first or one of the first episodes was on sort of electric vehicles, and we have, uh, you know, a fully electric fleet which are Jaguar I paces, and, uh, you know, I guess 11 comment that I wanted to mention about that is, uh, you know, I think the, the like charging infrastructure.
You know, for fleets, both Waymo and other fleets, you know, has somewhat like lagged behind the adoption of electric vehicles.
So I think that that's something that, you know, it's sort of like both a benefit and a challenge, you know, this is something that like Waymo's been planning with our partners for a long time, but that's, you know, sort of one challenge for adoption of our fully electric fleet.
Um, I think in terms of, uh, you know, to your question about like, um.
You know, jobs, I think that there are like a lot of different jobs that are created by, you know, our service, like through operating our service, and, um, you know, I, I, I think that, um, I think that that is like sort of, uh, you know, inevitable that like jobs change over time, um, you know, that's sort of like my, my personal opinion, I guess not waymos, but I think it's sort of like inevitable that like, you know, jobs somewhat shift over time and I think that sometimes, uh, you know, when, when people, uh.
Yeah, I, I guess that that's probably what I'd say.
It is true when there is innovation we've talked a lot in this podcast with any innovation we see a, a whole industry being created at the same time, maybe a, a shift from one type of job to the next.
No, and this has come up repeatedly, especially as we've talked quite a bit.
About, uh, AI, I, I, I wish we could keep going.
I think we could go for like another hour, but, uh, our Gary, our producers in our ear telling us no, no, no, we need to, uh, keep on our normal limits here.
I just wanna thank Will and her for your time, uh, and for sharing some of your thoughts there, um, and, uh, again appreciate your insights.
Yeah, thank you both for coming.
I learned a lot and I can't wait to, to see UVA racing team beat, uh, beat the competitors.
Yeah, we, we, we race, uh, many times a year and we try to be safe, but we've had some wrecks, but every wreck is a very, very expensive learning experience for us.
Yeah.
Well, good luck.
And, I'm waiting for Waymo to come to Charlottesville, so that's, we need that here.
Next Next time in next time in San Francisco, I'm going to take a robo taxi.
Please do, yeah, encourage if you're in.
San Francisco, Phoenix, LA County, or Austin, uh, you know, it's a great opportunity to experience the product.
So I'm in two minds.
I'm so excited about a, a, a future where I don't have to drive, um, places, although I do like to drive my little mini around, but I can imagine that it's so much safer, especially on long drives from DC late at night.
Um, but it seems like there are some real hurdles before this becomes common in every household.
Do you agree?
I do.
I do, but I think it, you know, we see this with technology all the time that when it first comes to market, it's often niche entry.
So for example, we didn't talk about this yet, but one place you're seeing quite a bit of autonomy is in tractors, um, closed environment, don't have some of the.
Uh, so John Deere pedestrians, you know, hopefully no one's running across the field.
Uh, John Deere has a number of products out there that are basically, you know, farmerless, uh, tractors that, uh, drive across the fields there.
Uh, we see it in delivery vehicles.
So again, it's not transporting people.
The risks are lower on a lot of college campuses.
You see these, uh, little delivery vehicles.
For, you know, pizzas and the like to the dorms again, you know, the risks are much lower.
Um, I think if it hits you, maybe it breaks an ankle at worst, you know, uh, so, so the risks are lower, and I think that's what we're gonna see, we're gonna see, you know, continued usage.
I, I think one of the interesting use cases that I, at least my understanding is we're getting much closer on long haul trucking.
So what you might see our trucks on the highway be fully autonomous, uh, and then when they pull off that last mile problem of getting to wherever they're going might need a human driver behind the wheel, but, but we might be pretty close to seeing that fully autonomous, uh, truck going down the highway.
A few years ago I took, um, not wheel's cohort, but the year after I took a group of students to Sweden and we heard from a company investing in, in trucking, autonomous trucking, but there they were, um, they were actually.
Um, manipulated and basically they were people in an office with joysticks basically driving them from far away, which is a very different kind of new, new innovation around that same concept still labor intensive, but maybe not as dangerous for them and uh and and just something kind of peculiar to think about.
Yeah, yeah, a little bit different.
Well, I think it's, it's time for us to answer the big question good disruption, bad disruption, no disruption.
Oh, it's a tough one.
I kind of know your answer already.
Um, I feel that, what do you think my answer is gonna be?
My, I think your answer is gonna be good disruption.
I think you're in favor of it.
Um, I think you think it's a disruption for sure and that it's good.
Um, I think that it's, it's funny for me.
I haven't seen it disrupt us yet that much, right?
I think that, uh, in order for it to be huge disruption, I would expect.
Industries changing or or key players fundamentally flipping from certain manufacturers to another, um, I think arguably the EVs have done maybe a little bit more of that with Tesla's prominence and how it's affected other, um, other car manufacturers, but.
Um, maybe it will be good disruption.
I just think that it's a little bit more in the future than, than people would like to think.
Yeah, I, I actually don't disagree with that.
I think, um, I, I definitely think as it comes to being, it will be disruptive, a large part because of also the business model implications again in a world of ride sharing and we maybe aren't purchasing vehicles ourselves, uh, that's, that's a radical shift in the way the market, uh, currently operates, obviously how long it will take.
For us to fully get there is is obviously the open question, and I think, you know, to Will's earlier point that, you know, cities, right, you know, will we see a city make a choice to just make only autonomous vehicles to remove some of the, you know, so houses don't need parking and you can build different housing structures and neighborhoods, yeah, and I'm just thinking about some of the risk and safety here, so there wouldn't be the human drivers kind of interacting with these, uh, vehicles.
Maybe we'll see that.
Um, um, but it probably will be this kind of longer transition.
I don't think it's no disruption though.
I think there's enough proof of concept of the technology.
I don't think this is like a bubble where, you know, we laugh about the fact that we were trying this, you know, back in the 2020.
Uh, I, I, I think there's enough effort taking place, enough progress, enough market actual adoption, uh, that we'll, we'll see this continue and, and continue to disrupt the industry.
That's great and it's interesting because it's, um, it might maybe we need to broaden.
Our definition of disrupting which industry it's gonna affect maybe different types of industries as you pointed out with the with the tractors, um, and maybe even the opportunity is yet to be realized.
Oh yeah, and I think again the spillover we talked about insurance, uh, parking garages, is it, you know, we have parking garages in the future, you know, there's a whole bunch of industries that are reliant on us, especially in the US, driving our vehicles that we own from point A to point B.
Um, and that, that might look radically different in the future.
OK, wonderful.
Well, there's hope for all parents there from the, you know, for kids ages 10 to 17 when all your day is spent being a taxi, taxiing them around.
There is hope for somebody else to do that one day.
That is a good disruption for sure.
If there ever was, that's a good disruption.
Well, thanks, Mike.
I learned a lot.
And let me thank, as always, uh, Gary, our producer, and Becky, our, uh, researcher extraordinaire.
Uh, this is good disruption, and thank you for being with us.
Good Disruption is a podcast from the University of Virginia Darden School of Business.