The future of autonomous vehicles
29:23 Share

Hi, everyone. It's Russ Altman here from the Future of Everything. We're starting our new Q&A segment on the podcast. At the end of an episode, I'll be answering a few questions that come in from viewers and listeners like you.

If you have a question, send it our way either in writing or as a voice memo, and it may be featured in an upcoming episode. Please introduce yourself, tell us where you're from, and give us your question. You can send the questions to thefutureofeverythingatstanford.edu. The future of everything, all one word, no spaces, no caps, no nothing, at stanford.edu.

S-T-A-N-F-O-R-D dot E-D-U. Thanks very much.

In terms of applications, now we see that autonomous systems are now becoming a reality. So we have robot taxis providing service in some cities in the United States. Unmanned aerial vehicles are now used ubiquitously. And even in space exploration, autonomy, particularly in the context of orbital applications, is making key strides. So very exciting times.

This is Stanford Engineering's The Future of Everything, and I'm your host, Russ Altman. If you're enjoying the show, please consider rating and reviewing it. We like to get fives when we deserve them, and we would really appreciate it. Your input will help spread the news about the show and will help us improve it.

Today, Marco Pavoni from Stanford University will tell us how autonomous vehicles, self-driving cars, and self-driving space vehicles are improving. AI has helped a lot, and there's been great technical advances. It's the future of autonomous vehicles. Before we get started, another reminder to rate and review the podcast if you're enjoying it. That'll help all of us.

You know, seven years ago, I had a guest, Marco Pavoni, come on to The Future of Everything and tell us what the status was in 2017 about self-driving cars. Well, I have Marco back today, and he's going to give us an update on where we are with self-driving cars, and also he'll tell us about self-driving space vehicles, which he's also working on. It turns out that AI has not only revolutionized our ability to write memos and emails, it's also revolutionized our ability to train people

cars to drive more safely, and while they're driving, to respond to unexpected events. Marco will tell us about that, and he'll also tell us that some of these technologies are now very relevant in space. There's been a blossoming of new companies with new applications of space vehicles, some of them as small as a shoebox, to monitor the Earth and help with logistics, with clocks,

with positioning of course, and with many other applications that you might not expect. Marco Pavoni continues to be a professor of aeronautics and astronautics and of electrical engineering and computer science at Stanford University. He's running an exciting new center for space and car autonomy, and he's an expert on the technical challenges of getting these vehicles to work safely and effectively for all of us.

Marco, you were on The Future of Everything seven years ago. Time flies. What are the big changes in the area of autonomous vehicles, self-driving cars? And I know you're also looking into aeronautics. Well, first of all, thanks for having me. Well, a lot of things have changed, both in terms of the tools that we use to design autonomous systems and in how autonomous systems are now becoming a reality.

In terms of tools, of course, everybody knows about the incredible advancements in the field of AI. And obviously, autonomy engineers have been very eager to use those tools in order to design in a completely novel way how we architect autonomous systems.

And in terms of applications, now we see that autonomous systems are now becoming a reality. So we have robot taxis providing service in some cities in the United States. Unmanned aerial vehicles are now used ubiquitously. And even in space exploration,

autonomy, particularly in the context of orbital applications, is making a key stride. So very exciting times. This is great. This is great. And so let's go, you've just hit so many good topics. Let's go to the self-driving cars, because seven years ago, you know, there was a huge bubble of enthusiasm at that time. And I'm not going to say any bubble was popped, but the

Then reality kind of set in and people were very worried about safety and trust. And you and I talked about that back then. What is the status of the enthusiasm for fully automated vehicles versus some kind of hybrid? I think we discussed back then there's different levels of automatic autonomous vehicles and

Where are we? The robo taxis do seem to be absolutely autonomous. There was talk about trucks coming down the interstate. And then there was the idea of every individual having a car that could drive around. Have we changed our expectations or are we still shooting for the same goals?

Expectations have changed for sure. As you said, as every technology, the self-driving technologies had its ups and downs. In 2018, I would say it was probably on the low side. As you said, realities kicked in and this technology proved to be very difficult to

productize and to make money out of it. So in those years, a number of companies shut down.

But in the past couple of years, I would say there is a resurgence of interest and excitement about this technology, but in slightly different ways. So now many of the companies have been shifting their focus from full self-driving to autopilot systems, similar to, for example, what you might have on a Tesla vehicle. And those systems provide value.

and also are actually quite profitable. So that's why a number of companies have been redirecting their interest toward that use case. We still have a few Robotaxis companies, but now we have a lot of companies focusing on very advanced driver assistance systems. Again, the autopilot of Tesla is a good case in point. - Right. - And technologically, as I said before,

All these new AI techniques that have been introduced in the past couple of years have provided a boost in the reliability and performance of this technology. So overall, I will say that right now we are in an upward trend, but with a scope that is a bit different from the one that we discussed seven years ago.

Now, I know you think very deeply about the technical challenges, and I definitely want to get to that. But a social non-technical challenge to some degree is the public's perception of these autonomous vehicles, not just safety, but like the weirdness factor of being in a car that doesn't have somebody holding a steering wheel. Are people getting used to this?

No, I'm not a UX expert, so I cannot really give you a quantification. Of course, this is why I ask you. But my perception is that in general, people are in awe for the first 10 minutes, and then they feel completely natural and they don't bother anymore about the technology in the Robotaxis space.

And this is actually a bit of a problem in the autopilot case because we want to keep people engaged. So we want to make sure that people also understand a little bit the limitations of this technology, at least in the autopilot setting. And so this is where actually engineers should have a clear conversation with the public to keep them informed. It's a very powerful technology, but still has limitations.

Okay, so let's move to AI. You said the first thing you said actually was that, and we've all seen this obviously in the last couple of years, the proliferation of AI applications and really performance in some areas that is quite remarkable. So tell me the ways in which AI is useful in the design or implementation of these systems.

There are multiple ways. Of course, AI is a very large field. And for example, AI techniques for simulation are finally making simulation integral in the development process of an autonomous system. But probably one of the most exciting AI advancements is represented by so-called foundation models, large language models or chat GPT being an example of them.

And these are models that are trained at internet scale. And so you may ask, well, why does that help? Well, the reason why a human person can learn how to drive in a few hours is that the human brings to the task of driving a lifetime of experiences. So you have accrued during your previous 18, 20 years experience.

a lifetime experiences in terms of detecting objects and the reason about how the world works and so on and so forth. Yes, I was on a bicycle for 12 years before I touched the car. Right, right. So basically, this foundational models allows you to bring in all these generalist experience that can be learned from videos on the internet to the task of autonomous driving, dramatically reducing the amount of data

that is vehicle specific. So this provides a big boost in performance, particularly in terms of reasoning about very difficult events that if you just

try to experience while driving, maybe you never see. But if you think about internet scale, maybe there is something close enough, not maybe related to task of driving, but it still allows you to reason about those corner cases. This is exactly what I wanted to ask you about because I remember in our last discussion, I think we talked about the importance of getting the edge cases, that you were training these models, but you needed to see the weird events, you know,

the clown on a bicycle or the baby in the carriage that's being pushed by a dog, you know, crazy things. And I was wondering as I was preparing for our discussion, if the ability of these large language models where you can tell them like, you can, of course, you can tell them, write me a poem or write me a short story, but now you can maybe say, and this is a question, even though I'm saying it like it's true.

create for me unusual situations that a driver might face. And I was wondering if these foundation models are any good at creating a useful set of edge cases that make your training of the driving vehicles more robust.

They are. So probably seven years ago, when you asked me this question, I didn't have a pretty good answer. This time, I think I do. And so basically, what these models allow you to do is two things. As you alluded to, when you design these models, you can use them to produce in simulation those corner cases. Of course, you want to make sure that you don't just generate a

arbitrarily crazy scenarios. But there are different ways to actually to ground this generation process in something that is plausible. For example, something that we have been doing with language models is to mine all the police reports and all the crash reports that have been collected in the United States, feed them into a large language model and use it to create scenarios

that are initialized from those crash reports. So that's basically the link to reality. So we generate scenarios that are similar, maybe a little bit different, but actually close to what has happened.

And then of course, during the operation of the system, again, for the reasons that I mentioned before about leveraging generalist experience, this instance may allow you to reason about those corner cases, maybe by making correlations with the weird things that have seen from, I don't know, the internet.

So, one of the things I also wanted to ask you, and I think it's related to all of this and it may be related to AI, is one of the things that I know you publish about and you study is where the cars are autonomous, but they're also communicating with the other autonomous cars on the road. And I find that fascinating because that requires cooperation, not just between the cars and the owners of those cars, but

it requires the manufacturers to agree on standards for communication. And I was wondering how that's going because I could imagine that being not only technically difficult, but also again, socially difficult where now the socialness is the companies that are both competing, but they have an interest to some degree in cooperating about standards for interaction of their products with one another. So how is that playing out?

I mean, you summarized very well the challenges related to vehicle-to-vehicle connectivity and vehicle-to-infrastructure connectivity. And this is the reason why, to a large extent, companies working both in the full self-driving car industry and in the semi-car

automated car industry are not relying too heavily on communication with outside entities being other vehicles or infrastructure. At some point there is of course some level of connectivity. For example, you want to get some data out of your fleet at the minimum to retrain your models in order to improve them. But in terms of leveraging the data for real time control,

Most of the companies are not relying on the data exactly for the challenges that you mentioned. And there are even more, like for example, cybersecurity challenges, cost of setting up that infrastructure and so on and so forth. And the environment is very competitive, so that's why

That's another challenge against introducing a T-shirt. - Yes, I mean, obviously. So the reason I asked this question is I don't know Elon Musk personally, but I have a sense of him from the press. And I know that there are these new companies, Lucid and others that are coming up, and it might not be his first instinct to say, let's take a meeting with them and get our engineers talking about how we can, and in fact,

The benefits would be great because you could imagine a much safer set of autonomous vehicles if they're doing basic checking with one another, like what's around the corner? Do I have a big turn coming up? All this stuff that Waze tells you in a kind of a crazy way. But this could be based on actual data. So, yeah, so I guess it will be a challenge. It's an interesting case where we're told we have this problem first. Yes. Yeah.

Exactly. Okay. You said it very well. Okay. I wanted to go, well, first of all, is there anything else that you wanted to say about what AI is doing? Because you've now, you've given us a good idea about it's useful in the training. It's useful maybe in the communications. We talked about this just a second ago and it's useful in real time. Are there other ways that AI might be surprisingly useful for autonomous vehicles? Yeah.

I think these are the two main ways. So how they accelerate the design of the systems and how they improve their decision-making capabilities. So basically their deployment. So I think these are the two most important points.

Now, I know that in your writings in the last couple of years, I see a lot of, I saw one very interesting paper where you use the phrase risk-averse planning. And, you know, forgive me if that's one of hundreds of papers you've been involved with. But I was, and also a lot of papers about safety assurances. What is the state of our understanding of how to maximize the safety of these vehicles?

So typically the way that safety is handled in safety critical systems is more as a constraint rather than an objective that you want to maximize. So basically you have a requirement in terms of how often you can fail and that requirement

basically stems from considerations regarding what happens if you fail. So what is the severity of your failure?

what is exposure of the failure, so how often that failure might happen. So you start from a requirement and then based on that requirement you try to come up with a design on the one hand and a validation strategy on the other hand that allow you to design a system that meets the requirement in a provable way. Here by provable I don't mean necessarily theorem but typically more statistical analysis.

Provable was, I've seen the word provable in your papers, and it's very exciting. And the degree to which you can use that word, you know, truly would, I would guess, be extremely reassuring to both to the public and to the technical community. So that line, we weren't, I don't think we were talking too much about proving things at all seven years ago. And so it seems like there's been an advance. And I also am also very intrigued by your comment that in the old days, there was the

The objective was like get the person from A to B and the constraint was and do it safely. And it sounds to me like there's been a little bit of a change where now it's I want you to drive safely and I also want you to get the person from A to B. Yes. So basically we talk about the availability of a future and the safety of a future. So we want to be safe.

up to a given requirement. And we want to maximize availability of the future. So basically for how often we drive autonomously. But definitely we don't want to compromise availability with, we don't want to compromise safety in the quest for additional availability. This is the Future of Everything with Russ Altman. More with Marco Pavoni next.

Welcome back to the Future of Everything. I'm Russ Altman, and I'm speaking with Marco Pavone from Stanford University. In the first segment, Marco gave us a great update on where self-driving cars are, how AI has really helped improve them, and what the current challenges are. In this segment, he's going to tell us about how some of the same technologies are being used to create autonomous vehicles in space. It's a very crowded field.

both literally and figuratively, because there's a lot of satellites up there and there's a lot of companies trying to create new applications for space satellites. Marco, I know you also work, and this is incredible, on autonomous vehicles in space or in the air. So tell me, what are you doing there and what are the challenges there? Well, first of all, with respect to seven years ago, a lot of things have changed.

including the increasingly large roles that private companies play in the space sector, SpaceX, of course, being one of the most notable examples. New missions, new things that we have learned, particularly regarding exoplanets, that is planetary systems outside of the solar system.

And in this context, autonomy, again, is playing a big role. For example, there is a lot of emphasis to go back to the human-- sorry, to the moon. And of course, a prerequisite is to make sure that there is enough infrastructural capabilities to build an outpost on the moon.

And since you can't really, you know, transfer workers easily between Earth and the moon, it stands to reason that most of that work will be done by robots. And for example, we have collaborations in that regard. And again, AI is going, is poised to play a big role

role in the design of autonomous systems for these new space applications. We understand that, of course, the environment is different. Spacecrafts are very limited from a computational standpoint in terms of amounts of data that can crunch with respect to what you can do on a self-driving car. And data is very limited.

like to a large extent, you can relatively easy gather data for self-driving cars by just driving around. But gathering data in terms of how to land on the moon or maybe on some satellite of Jupiter,

You don't have the data. And so in a context where data is becoming increasingly important, how do you devise techniques that are much more judicious in how you use that data? So there are a lot of promising studies. There are path forwards, but...

overall the strategy would be different. Are we mostly thinking about unmanned, unpersoned vehicles or because that takes away a whole bunch of safety? I'm sure you still don't want these things to blow up and crash because they're expensive, but it's different when there are humans involved. So is human part of the equation right now or is it mostly entirely robotic systems?

I mean, in some cases, there is a debate whether you want to have a human or not. In some other cases, there is no debate at all. So if you want to explore icy bodies in a solar system, that for me are among the most exciting targets for space exploration, such as, for example, Europa around Jupiter or Enceladus.

You can't send a human there. It just takes literally too long. And maybe the human will be able to get there, but definitely wouldn't have time to come back. So nice and easy. So there you want to have a robotic fully manned platform.

For other tasks, like for example, going back to the moon, you probably want to have a combination of astronauts and also robots that take care of the dull jobs, like for example, construction jobs.

And then for operations around the orbit, like for example, for in orbit assembly or maintenance of space assets, again, there is a strong push toward robotic systems.

from an economic standpoint to basically minimize cost. So the short answer is yes, there is a strong push toward increasingly more unmanned platforms. Are there similar issues as for self-driving cars in terms of coordination of multiple vehicles? Or is space so big that you can make the assumption that it's a freedom to operate and you don't really have to worry about other devices in this space?

Well, believe it or not, collision avoidance is also a problem in space. Okay, okay. All the junk that you have around the Earth orbit. Yeah, I think I recently read that Starlink either has or will have 6,000 satellites and is working towards 30,000. Does that sound... Yes, yes, that's right. And especially in the past, people were not very...

careful in decommissioning space assets. So that's all the junk that we still have. And that junk does not communicate. Does it even put out a signal to say I'm here? I mean, you can see it through sensors, but it doesn't really communicate or definitely doesn't cooperate in order to communicate.

That said, in terms of communication, actually that plays a big role in space missions in the sense that the idea of sending groups of satellites or space assets that jointly achieve a task that will be impossible or very expensive with a monolithic mission is actually becoming almost a norm in space.

Is this because the thrust required to get them up into space is much cheaper if you break it up? Part of it. And the other is that, of course, by distributing your space assets, for example, you might have a much better coverage than what you could achieve with a monolithic architecture. Or if you want to live in a monolithic architecture, you might need to require much more expensive sensors.

I could also imagine you were saying before that it's hard to get training data for landing on the moon or landing on Mars. And I'm making this up and this is not what I do, but I could imagine that if you have a bunch of vehicles headed towards the moon, there's a certain triangulation they could do just with each other and with various stars and other milestones to perhaps do more accurate estimation of the flight. Is that a thing?

Yeah, absolutely. That's one of the ways that distributed space systems might actually work well by exploiting collaboration. And if I may, advertise a new research center we have set up at Stanford. Together with a colleague of mine, Professor Simone D'Amico, we have started a new center called the CISAR. We're both Italians, so there is a legacy in the name. And this center is all about

investigating ways to use AI techniques judiciously in the design of distributed autonomous systems for space exploration.

Okay, so that's very exciting. A lot more to tell you in seven years from now. Yes. And actually, I want to go back because you led off this discussion of space by saying it's very different because of private industry. And so tell me a little bit about that because this is kind of a new thing. In the old days, aerospace meant NASA, right? That was like I grew up at a time where they were synonymous. What...

What changes from a technical point of view are introduced when you have multiple players? We already talked about on the road the difficulty of communicating with a competitor's car. Is it any better when it comes to the space race? I think there are a number of reasons why...

private entities have become involved in the space sector. One of them is technological innovations. For example, in the past 20 years, we've seen how you actually can do excellent science or

fulfill useful tasks with miniaturized space assets. So you don't need a satellite that is as big as a container, that might be very expensive, but even a shoebox type of satellite can already provide value. So that of course has opened up opportunities for a number of stakeholders that of course cannot invest billions of dollars in space missions.

So there is one aspect of it. The other aspect of it is that a number of opportunities have opened up for business, for example, using space assets for communication, for surveillance, for logistics. And the people have been very quick in identifying those opportunities and together with the technology innovations that now make those opportunities profitable.

That's why you have several private stakeholders out there now.

This is very exciting because just the list of applications you mentioned went way beyond what I was aware of. And I can see that really almost every industry can now think about, would a little shoebox flying up above the earth be useful to my logistics or to my business? And the answer might be yes, a lot of the time. This is also obviously an international challenge. How is the level of collaboration across all the different international players?

Yeah, that is a little bit outside of my domain of expertise. Of course, space is a topic that unites mankind, but at the same time is a very sensitive topic from a military standpoint. So with some countries, of course, there is a close collaboration with others, as you can imagine. Right now, there is a much more tense conversation.

Thanks to Marco Pavone. That was the future of autonomous vehicles. Thanks for tuning into this episode. You know, we have more than 250 episodes in the archives, so you can find pretty good conversations about the future of almost anything. Please also remember to hit follow in whatever app you're listening to, to make sure you never miss an episode of the future of everything. You can connect with me on X or Twitter at RB Altman and with Stanford engineering at Stanford E N G.

If you'd like to ask a question about this episode or a previous episode, please email us a written question or a voice memo question. We might feature it in a future episode. You can send it to thefutureofeverything at stanford.edu. All one word, the future of everything. No spaces, no underscores, no dashes. The future of everything at stanford.edu. Thanks again for tuning in. We hope you're enjoying the podcast.