nuTonomy

Startup Exchange Video | Duration: 17:22
October 18, 2016
Video Clips
  • Interactive transcript
    Share

    EMILIO FRAZZOLI: One of our core main advantages is exactly that of how to teach a computer how to obey all the rules of the law. Actually, we do something that goes beyond that. As we all know, there is a body of rules that control what we can do and what we cannot do under them.

    However, very often, we violate them. And it is accepted it is OK to violate some rules under certain circumstances, as long as you pay attention. For example, you may be stuck behind a garbage truck. You can either stay in your lane and wait half an hour until you get to the next intersection and can get away.

    Or you can just look. If nobody's coming from the other lane, you can get in the opposite lane, pass the garbage truck, get back into your lane. And as long as you do it safely, and you check that nobody is coming from the other direction, nobody will really fault you for that.

    What we are teaching our cars is a principal way of using some form of judgment to violate some of the rules when it is necessary to do so but is also safe to do so. So essentially mimicking the human reasoning in this kind of situation.

    That said, I think that there is a danger, in everything I read about autonomous cars, to humanize the car. Like this anthropomorphic view of the driverless car, where we assign the driverless car the same kind of ethical criteria or ethical capabilities that we would expect in a human.

    I think that is somewhat dangerous, because the self-driving car will not be human. It's not meant to replicate the human. Some people talk about having a Turing test for autonomous cars. That is, would I be able to distinguish, from the outside, a car driven by human or a car driven by computer?

    And the idea is, do I want to make them essentially indistinguishable? I'm not sure that's a good idea. That is, if you think of it, it's tantamount to saying that humans are the perfect paradigm for driving. And as we all know, they're not.

    This could be a little bit controversial. But one way that I think that we could think of autonomous cars is really as appliances that happen to drive from place to place. An appliance is something that does a certain task, it's a very well-understood task, reliably and safely.

    If we take a refrigerator or a washing machine, we don't evaluate what they do according to ethical standards. But clearly, thinking of autonomous car as an appliance will not help selling the autonomous car as a product. However, if I think of autonomous car as something that is just offering me the service of transporting me from A to B, then I think of it as an elevator, a horizontal elevator.

    Or a more sophisticated elevator that, instead of just going back and forth along a fixed route, can actually travel over a complicated road network. Essentially, take me anywhere I want to go in a city.

    I think that is a somewhat critical change of perspective. So we don't think of elevators as something that needs to behave like a human would behave. An elevator is something that, when you push a button, you expect it to come and open the door at your floor.

    And you get in. You push another button. It takes you to the third floor, or whatever you want to go. And that's it. So this is the service that it provides. And as long as it provides this service in a very well-understood, reliable, repeatable way, then it's fine.

  • Interactive transcript
    Share

    EMILIO FRAZZOLI: One of our core main advantages is exactly that of how to teach a computer how to obey all the rules of the law. Actually, we do something that goes beyond that. As we all know, there is a body of rules that control what we can do and what we cannot do under them.

    However, very often, we violate them. And it is accepted it is OK to violate some rules under certain circumstances, as long as you pay attention. For example, you may be stuck behind a garbage truck. You can either stay in your lane and wait half an hour until you get to the next intersection and can get away.

    Or you can just look. If nobody's coming from the other lane, you can get in the opposite lane, pass the garbage truck, get back into your lane. And as long as you do it safely, and you check that nobody is coming from the other direction, nobody will really fault you for that.

    What we are teaching our cars is a principal way of using some form of judgment to violate some of the rules when it is necessary to do so but is also safe to do so. So essentially mimicking the human reasoning in this kind of situation.

    That said, I think that there is a danger, in everything I read about autonomous cars, to humanize the car. Like this anthropomorphic view of the driverless car, where we assign the driverless car the same kind of ethical criteria or ethical capabilities that we would expect in a human.

    I think that is somewhat dangerous, because the self-driving car will not be human. It's not meant to replicate the human. Some people talk about having a Turing test for autonomous cars. That is, would I be able to distinguish, from the outside, a car driven by human or a car driven by computer?

    And the idea is, do I want to make them essentially indistinguishable? I'm not sure that's a good idea. That is, if you think of it, it's tantamount to saying that humans are the perfect paradigm for driving. And as we all know, they're not.

    This could be a little bit controversial. But one way that I think that we could think of autonomous cars is really as appliances that happen to drive from place to place. An appliance is something that does a certain task, it's a very well-understood task, reliably and safely.

    If we take a refrigerator or a washing machine, we don't evaluate what they do according to ethical standards. But clearly, thinking of autonomous car as an appliance will not help selling the autonomous car as a product. However, if I think of autonomous car as something that is just offering me the service of transporting me from A to B, then I think of it as an elevator, a horizontal elevator.

    Or a more sophisticated elevator that, instead of just going back and forth along a fixed route, can actually travel over a complicated road network. Essentially, take me anywhere I want to go in a city.

    I think that is a somewhat critical change of perspective. So we don't think of elevators as something that needs to behave like a human would behave. An elevator is something that, when you push a button, you expect it to come and open the door at your floor.

    And you get in. You push another button. It takes you to the third floor, or whatever you want to go. And that's it. So this is the service that it provides. And as long as it provides this service in a very well-understood, reliable, repeatable way, then it's fine.

    Download Transcript
  • Interactive transcript
    Share

    EMILIO FRAZZOLI: The technology that we are developing at nuTonomy, it really has its foundation in the work that Carl and I and other colleagues at MIT and elsewhere have been carrying out over the past-- I would say at least 10 years-- partly at MIT, but also at the SMART, the Singapore MIT Alliance for Research and Technology.

    We started doing this work at a time of the DARPA Urban Challenge, which by now is about eight years old. So we keep talking about it. But we had to realize it is also eight years in the past.

    So the [INAUDIBLE] has moved a lot. And I think that our main advantages are really twofold. On one hand what we have is a technology that allows our cars to behave in a way that automatically or systematically satisfies all the rules of the road. This is something that we did in a very painful way at the time of the DARPA Urban Challenge. And at the time essentially what we had to do is essentially hand code all the different options, all different things that could happen and what would be the right behavior in all possible situations.

    This was very painful to-- it was very painful to write the code. It was even more painful to debug the code. Later we actually figured out how to do that in a systematic way without any need for hand coding. So essentially what we do is we provide the software with a list of rules. And then the software that is running in the card automatically satisfies all the rules. I think that, you know, this is something that is very powerful, a very promising approach, and I think that this is something that is unique to us.

    Something else that is unique to us is not only the vehicle technology, but I also think the business model. So how you will actually operate these vehicles, what kind of service you will provide to the customers, how would you operate the fleet. So you've written some of the seminal papers, you've done some of the seminal research on how would you manage a fleet of autonomous vehicles in such a way that you can provide the quality of service that your customers would expect and the auto service that your customers need to be convinced to use your service.

    For example, what we did is we did like a case study based on data from-- like a real data from Singapore. And what we saw is that a fleet of about 300,000 vehicles in Singapore could provide everybody in Singapore with the same level of mobility that they enjoy nowadays.

    How many is 300,000? At the time where we did the study-- or at the time where they didn't study the first two, there were about 800,000 vehicles-- passenger vehicles in Singapore. So this is actually a reduction of about 60% in the vehicle count.

    And if you think of it, this is half a million cars that you can just-- that you will no longer need in Singapore. So you can just sell them to some other country. And half a million car means typically one million and a half parking spots. And this is all space they can give back to people.

    And the thing that this is something that is clearly very important in a place like Singapore, which is very geographically constrained. But if you think of a lot of the downtown areas in big cities worldwide, get rid of all the rubber and metal and all the parking spots would actually allow people to get back their cities and enjoy them in a very different way. So I think that there are huge potential also in terms of urban planning, city development, and how people really live in the city.

    Download Transcript
  • Interactive transcript
    Share

    EMILIO FRAZZOLI: There are these two main mental models for me. Most of people think of autonomous cars as a product to sell. I like to think of autonomous cars as a service that I can provide for the users.

    I think that the service model will actually allow us to make much faster progress. If I buy an autonomous car, or something that is sold to me as being an autonomous car, then I want this to be autonomous at all times, day, night, all weather conditions, any location where I want to use it. And by the way, I don't want to do that much maintenance.

    So I think that thinking of the autonomous vehicle technology as a product to sell is something that will proceed very slowly. Because a lot of the economic financial drivers are, in a sense, the wrong ones. What I favor is more of the service model.

    Now, if I am a fleet operator, and am offering a service that is based on autonomous vehicles, then I can choose when the service is available, where it is available, under what conditions. I have my staff that takes care of the maintenance. And I'm not selling them the sensor, these fancy sensors, or this fancy technology that is on board the car.

    What I'm selling time is a service. So now I had to think of-- what I have to do is compare the cost of the service that I'm offering to them, including the capital cost to me of acquiring the car and acquiring the sensor and [INAUDIBLE] technology, with the standard way of providing the service, that is with a carbon-based life form sitting behind the wheel.

    Now, if you think of how much I need to pay a driver, if I want to offer a mobile on-demand service, like taxis, essentially, how much would a taxi driver or a driver cost? Let's say the minimum, I would imagine-- $50,000, just to keep the numbers round. So let's say $50,000 a year.

    And that's for one person. If I want to provide 24/7 service, then I need to hire at least three people for covering three shifts. Now the cost of providing this service to me, just the person sitting behind the wheel, is about $150,000 a year.

    Now, having a sensor that cost $10,000 doesn't sound that bad anymore. Because really, what we are doing is we are kind of amortizing this expense over the life of the car, of the service that we are offering. Really, what we have to compare to is what would be the cost of providing the services using a different approach, and maybe a more traditional approach.

    Download Transcript
  • Interactive transcript
    Share

    EMILIO FRAZZOLI: My name is Emilio Frazzoli. I am a professor of aeronautics and astronautics at MIT. I recently started a company called nuTonomy with my colleague Karl Iagnemma, who is also a research scientist at MIT.

    The company really started two years ago as an outgrowth of our consulting business. And we have been working with a number of automotive OEMs and tier one suppliers to essentially transition our work and our expertise on autonomous vehicles to the market.

    For most of my career, I worked on unmanned aerial vehicles. Nowadays, most people call them drones. And it's a very popular research area. I've been doing that for about 10 years.

    Then about 10 years ago, there was the time of the DARPA Urban Challenge. That was a robotic competition for self-driving cars, for autonomous cars. And I started taking an interest in that field and applying a lot of the techniques that I had learned and developed over the years and applied to the aircraft case.

    I found the automotive domain actually very interesting, especially because after a few years that I was working on that, primarily on the technology development, I took a step back and started thinking about, what is the potential impact of this technology? How will this technology change the world? And if you think carefully about it, the answer is actually quite surprising in the sense that I really think that this technology has a big potential for a very profound impact on our lives.

    Essentially, there are a number of reasons why autonomous cars are being developed by many players in the industry now. This is a very common topic. And you read about autonomous cars almost every day in the news.

    Typically, the arguments that people make are about increasing safety. So as we all know, most traffic accidents are actually due to human errors. You remove the humans, you remove the errors, at least this is the idea. Then is the increasing convenience, so that if the car is driving, I can text to my heart's content legally. Things like increasing efficiency in a city of traffic, in a city reducing congestion, reducing environmental impacts.

    These are all great things. However, all of these features will not change the way that we think of cars and the way that we approach personal mobility. I think that the killer app for autonomous vehicles is enabling car sharing. Car sharing at the level of convenience that is essentially the same as owning your private car, but to the sustainability of public transportation.

    Most of us who own a car, these cars cost a big chunk of our disposable income, and yet they sit somewhere 95% of the time. Usually, this somewhere is a pretty expensive piece of real estate. So we actually pay for the privilege of not using the car that costs us so much money.

    So clearly, this is not a sustainable model, especially if you think of a situation where you have a car that is actually able to drive itself. If the car is able to drive itself, why leave it parked in the garage? Let it go pick up somebody else, drive somebody else, about their business, or go back home, take your children to school, or take your wife to her job, or anything that you can imagine. But also, you can think of a car that you're actually sharing as a transportation infrastructure that everybody can use.

    Download Transcript