henry.liu - av safety assessment program + panel talk

4:41PM Oct 25, 2024

Speakers:

Keywords:

autonomous vehicle safety

international standards

behavior competency

naturalistic driving

simulation testing

probabilistic safety

national framework

third-party testing

collaboration opportunities

regulatory framework

cost per mile

infrastructure support

probabilistic proven safety

liability minimization

industry progress

They are solving one of the most difficult problem in AI, and also have how the great potential to solve to help us to improve safety, as well, as you know, mobility issues. And Jing Huang asked, What are the things we can we should compete, and one of the things we should collaborate, and I think one of the things we should collaborate, is to have some consensus at how we can evaluate the safety performance of these autonomous vehicles. Are they safe enough to be deployed on the road, and how we can do that. And so I want to start, oh, the clicker. Okay. We do have clickers. Okay. So I want to start with this slide. This is self reported in terms of safety performance from some of the autonomous vehicle companies, cruise, Waymo, and also level two of Tesla and these, I did not really pick this company, simply because this company has reported safety performance. And as you can see, and every company reported, they have, you know better than human drivers, they also have, you know benchmark in terms of human drivers as well. And then I would say there's a lot of caveat within these numbers. And you have, when you calculate in terms of the accident rate, you need to look at in terms of what towards the accident, what miles being driven, and what the drivers drive these miles. And so there's a lot of things going into these numbers, and these are soft reported, these not verified by third party, and there's no consensus on how we evaluate and test autonomous vehicle, as I mentioned before. So, so what I like to say is, there's also international standard. There's international standard related how we're going to test autonomous vehicle. And this is, this is the standard about two years ago, in terms of using scenario based approach to really look at two things, two objectives. One is to look at what are the behavior, basic behavior, competency. Can these vehicles doing some of the basic things like car falling, Lane changing, navigate through roundabout and unprotected left turn and things like that. Second, what are the safety performance in comparison to human drivers, particularly in terms of accident per mile and things like that. Okay, so there's two objectives. However, this international standard doesn't really give you anything related with how we're going to do this. There's no procedure, there's no recommendation in terms of how we're going to do this, and so, so that's why we we've been working on this and cities Safety Assessment Program. We're trying to provide a framework for the industry to to adopt in terms of how we can help evaluate the safety performance. And we have two parts, one where I call driver license test. Well, that what I really means is that we want to make sure these vehicles has basic behavior competency before they can put on the road. We have seen from our own test track. By the way, I'm sitting it's a test track with test autonomous vehicle every day. And we have seen some of these vehicles on the road being deployed to our lace, and they failed, in some of our cases, basic behavior competencies. And so the second thing is that we like to see how these vehicle perform in comparison to human drivers in terms of accident per mile and and then things like that. And so there's two parts, in terms of our framework, how we evaluate the safety performance. And so both of these two parts, in terms of our framework, really build upon two things. One is we want to build these naturalistic driving environment models so that we can realistically evaluate the safety performance of these autonomous vehicles. And that's not enough, because then even in simulation, you will need a really long time in simulation model to run to get to that point, so we have to somehow to accelerate. And so that's why we call it naturalistic and auto zero driving environment, so that we can accelerate the testing process for autonomous vehicles. And so that's essentially our framework. We're trying to build the model for Netflix driving environment. We're trying to accelerate this evaluation as well. Okay, so I can't show you in terms of our commercial testing, but we do testing for open source system, and all the wear is open source autonomous driving system is openly available online. And so we tested our latest work on all the way our universe. And so we do two parts. And first of the number one is looking at their basic behavior competency in terms of driver license test, we test 14 of the scenarios. And then if you look at, in terms of, out of these 14 scenario, almost seven of these, they failed in some of the scenarios, okay? And so these are just basic behavior competency for these, for this particularly open source automated driving systems. And so that's the second column. In terms of pass and fail. You see number, your number of fails over there in terms of some basic behavior competency. Okay, we also so this is how we do the do the test. We test this in both the test track and also in simulation, as well, as you can see on your left hand side. Not sure whether I can speak. Let me see whether. Okay. The right side. It seems working. If you can, if you can click on the left hand side, this is how we do the real test in Michigan. You can see snow as well. And so autonomous vehicle crossing the intersection. At the time, there's a background vehicle doing the left turn. We want to see whether autonomous vehicle will be able to handle that. Obviously, you can't in the if you use a real background vehicle, you can't really do test these dangerous scenarios. So we that's why we on the right hand side, we test this in mixed reality. So the background vehicle is actually provided by through a simulation model. Okay, we also do the driver driving intelligence that we want to look at overall, in terms of accident rates and things like that. We look at in terms of how run the simulation model until they convert. And this, this is some of the situations. Let me see whether, whether you can click on any of these. Yeah, sure. Sure. Just this one, yeah. So you can see, in this type of situation, the red car is the autonomous vehicle. Blue car copy in and and so the autonomous vehicle actually slowed down. That lead to a rare end accident, okay? And so these are the things we are testing within the simulation to see in terms of accident rates. And then there's multiple type of situations, and I don't think I have time to go through. Okay, all right, so that's my sort of my last slide, why we need a national framework for autonomous vehicle testing. I want to say three things. One is for autonomous vehicle, because we are evaluating not just the vehicle systems. We are evaluating the vehicle plus AI based drivers. Okay, so we want to move beyond from the functional safety, which we do for all the automotive industry, we do functional safety testing, we need to move beyond that into more behavioral testing from third party, from consumer perspective, how these vehicles perform in a naturalistic driving environment. The second thing I want to mention, we need to move beyond from the so called deterministic safety to more probabilistic safety. What we really mean by that is, if you look at these autonomous vehicle companies, in fact, if their safety performance is being validated, they are doing better than human drivers. We should help them to commercialize to scale. We should not penalize them because one single failed case, so that we that's what we were arguing. We should move beyond this deterministic case based reactive investigation into more probabilistic, overall proactive type of licensing and helping them to scale and to scale, to commercialize their technology. I think I have one second. That's my last. Thank

you. Let me invite shouty out of hotel on this stage as well.

Great. Thank

you, all of you for the presentation there. Maybe I'll start with Henry. You're very humble. You said this is the MCT license test, right?

Yeah, we don't call it license test, because when I say license test, they're always questionable. We call this safety assessment service,

yeah, but I title your talk as national AV test, right? So what would it take to move from MCT test to a genuine national test?

I think one of the things I like to see is to have consensus, not only from academia like us, but also from industry, from the autonomous vehicle companies and also third party testing agencies, and not help us to put together this framework so that we can all agree some safety assessment. And I know safety testing is part of your job. You you have to do testing in your own company. But to consumers, we also want to know, are these products you build safe enough and so third party some sort of behavior based testing is really needed? I see, oh, thank you. So consensus will be, will be required, right? Yeah,

thank you. Yeah. So Fauci, I introduce you as the founder, CEO of bot, dot auto, of course, you're also the founder CEO of two simple right. Over the years, I like to share with us what you learn from two simple experience and what make you to really have this new framework, a new perspective looking at our own striking

industry today. Yeah, well, in the early days, I thought the technology can solve every problem, and in fact, my experience told me that, you know, you need more than that. That's why I think, right now I'm trying to be the architect to bring the commercial side of the world with the technology side of the world

I see, oh, thank you,

Raquel, here, while we had a fantastic year last year here, right? Yeah, but tell us, what do you see in the next two years, right? What will make you happy see two years from now, what can achieve?

What will make me happy, or what? Yeah, sorry, jokes apart, yeah. So what you're gonna see from us, I guess the question is about our roadmap, right? Yeah. So next year we're gonna launch. So today, we have commercial operations in Texas where you can see also driving trucks with a safety driver. Next year, the safety driver will not be on board anymore, right? So it's a very, very exciting time. So, you know, talking about safety. So our safety case with, you know, kind of to the next level, where we call safety 2.0 bring in a full assessment that is statistically of that. Can you really, you really prove that it's not human before deployment? And so you what you gonna see, and then it's a scaling of product, and that's what you're gonna see. You know, over the next couple of years, the bottleneck for the industry is actually OEM production of the redundant platform. Today, our technology will know, the bottleneck in terms of adoption is really, you know, think of an electric rice not it took time from the first redundant, from the first electric vehicle to the time where you can actually create hundreds, 1000s, right? Two years, three years. So the same is what you're going to see with the driving trucks, and that's going to be what you know limits adoption, right? Great. Thank

you. Yeah, as I mentioned before the session, during this panel, I really want to say, how do we collectively grow and cultivated the autonomous vehicle as the industry as a whole, right? One specific question asked, Where should we compete and where should we collaborate? Henry, you answered your version of the question during the presentation. I'd like to hear from Xiao Diane request, where do you see us? Yeah,

I really don't see any competition at all. To be honest, the market is so huge, like even in my most optimistic projection, I don't think I'll see our call in 2033 so this is how fast this industry is. So there's no competition, only collaboration. And we want to bring orders, bring consensus to the industry. That's what we really need. For example, I think years ago, the MPI was, was a terrible counter example, where every company is, in the end, posting about their own MPI with making the MPI, I think, like making the one who wants to report the real MPI, a lot of pressure, which is unnecessary and and in the end, nobody cares about that metric. But I think right now, for this time, as I mentioned in my presentation, the cost per mile really has a chance. Because, you know, this is the auditing term. Every if you, every auditor that you hire should give you the same number for cost per mile. And I think this is a very good metric for for indicating the future commercialization for autonomous driving. For example, a call has said next year they're going to do driver out similarly for us, we're going to also have the same roadmap for next year. But maybe, I think in the initial run, the driver out will be maybe one or two demos, five demos is not going to be like in one night, all of the truck, all of the truck with the safety driver are going to be shifting to 100% driver out, in my view, is going to be a gradual process. But throughout this gradual process, in reality is that the number, the cost per mile, reduces over time, but how quickly it reduces, we need a verifiable way of indicating the progress.

Yeah, in our prior conversation that your business model, because you are signed contract with actually delivering the service, therefore you as a company absorb the risk to what degree this transition to the drivers do? Yeah, that's actually interesting design. Tell us more about that. Yeah,

because I think in the end, we wanted to provide capacities like this is a super complicated systems like like a mainframe computer. In 1970s you sell the mainframe computer. IBM sells a mainframe computer to bank. And you know what? With a computer, there are two PhDs associated with a computer and maintaining it every day, and that is the level of completeness for autonomous driving. I'm not talking about the software. I'm sure that with with Gen AI, your company, my company, we're gonna all enjoy the Gen AI benefit of it. But I think there is always hardware problem. Like, you know, the server is over, hit, okay? The sensor is not clean, or the sensor is not clean enough, right? All of these issues are going to bug the day to day fluency of the operation, and that is all about cost, like one minute in the delay of the presentation probably mean, like $20 in some real operation. And we really wanted to make sure that we calculate every number in here and what is the best way to encompass everything we transform our own business.

Thank you. Yeah, roquel, what's your thought? Yeah. So

coming back to the collaboration opportunity, right? So I think there is a tremendous opportunity to collaborate, and it's in everybody's interest that we are actually working together in the safe deployment of this technology. If somebody, one of us is not doing the right thing. It's going to hurt all of us, right? So that is only us about, you know, safety. I often say safety shouldn't be proprietary, but it's also about, you know, we we work instead driving trucks. It's a fairly complex ecosystem, right where, on your side, on one side, you have regulators, you have, you know, your customers with private fleets, with carriers. You have the OEMs. You have the suppliers, or some of the sensors compute. You have insurance, some of them that are present in the in the audience here, you have the free brokers, and you know, this has to name a few, right? You have terminals, you know, etc. So we need to collaborate in terms of building that ecosystem so that we can enable that deployment at the scale. And I think that's lots of opportunity. And there is some of the associations that are represented in the audience are a way where we actually collaborate. And I usually believe that MIT has the opportunity to become a key player of really going from, you know, the deployment, or a few sub driving trucks, you know, in product, as I was saying, for 2025 but making that a scalable solution for, you know, this trillion dollar market.

Thank you. We actually aim to do that. Yeah, go ahead. I

just want to add what just said. This is where, in terms of where we should collaborate. I think this is where technology is not enough, and there's a lot of support needed for this technology to scale. For example, the business model is itself in terms of how infrastructure can help right low weight infrastructure, can we make up their life a little easier by doing some landmarking better than markets even, and how we how we can ensure these vehicles, right? So, MCD itself, we are a part of private partnership. So State Farm is part of our partners. They also struggling in terms of how they can ensure Tesla, even these, these ADA systems, right? So, so, how do we ensure these vehicles as well, and so, so there's a number of things. Is pre competitive, is across everybody, and then we can work together on those to help them to scale their business and build out we can help us to improve safety and efficiency. Thank you.

Bu, long check the time. I mean, I have a whole page. This can go on for two hours, right? But let me know what time format. Okay, great. Thank you. Yeah, so then we'll talk about this. There's a lot of common good in these industry, right? As we often mentioned that in aviation industry, we don't, we don't compete on safety. People know that one plane drop hurts everybody, right? It seems, it seems, for us to apply here, right? So given this question, what do you see are still missing in today's ecosystem here, right? So either from the company point of view, so if, say, federal government do something, my life would be better, or the state or city do something, or the Supply Chain system do something, what do you see as a missing still in this system?

Yeah, go ahead. Yeah. Okay, I'll start so I think the one of the biggest issue that I see is that if I see the upper bound of some current problem cannot be solved by autonomy, then that is something that bugs me. In the middle of the night, I think construction zone is actually one of them. I give you one concrete example, like I as a human being has run into the because of the very weird configuration of the cone. I ran into the wrong side multiple times when I encountered construction zone in my life. And I think, kind of think about this problem that, like, there's no way for a human being to even recognize will be the right way to get into for a construction zone. How can a computer do that? This is basically the you define the nature of the problem, not really because we don't have smart enough algorithm. Basically, whatever algorithm you have, there is always a configuration that can misguide you, right? And the social tolerance for autonomy is, is, you know, another factor that compounded on it. So I think really, I would really wanted to see more IoT technology to be embedded in here, to be having a more coordinated construction zone by the actual construction zone workers for full communication. Because in any cases, in the world, in the end of the world, there's always a way to configure the codes to misguide the system,

right? I look at which is here from for real point of view, how this infrastructure side and the AV side can really coordinate this, yeah, and also even how the cities or state government facilitates this conversation. Ocean

APIs, we are posting APIs that will be consumed into x using under Communication connected area. Give you an amplified awareness something that is happening with the right context and with the right petition in

the world, we need to talk. Pun intended,

yeah, that's one thing we want to contribute. John is facilitative conversation here right now?

Yes. So, you know, if I have a magic wand, I will say that having a global regulatory framework that, you know, across the world, versus states, you know, i We are headquartered in Canada, provincial, you name it, right? So that's and also, a lot of the regulations are made for humans, right? So there is some changes that need to happen for machines that actually will make the deployment one triangles is one example, right? And you know, if you get into an accident a little bit of the road, you have to put that one triangle. But obviously machines are not going to do that, right? So there is talks now the industry into how we can do something different, etc, but be open to the world is changing, and the world is changing, we have the opportunity to change it for a much better world, right? So let's enable that. And let's have, you know, as as as common as we can, you know, framework so that we don't need to go, you know, at every time that we change our province or between the state, we need to change the way that we do things,

indeed. Yeah, all right, if you look at the progress by SpaceX and also Tesla in terms of their safety performance, and one is in the space industry, one is in the automotive industry, I would argue autonomy, autonomous vehicle obviously, is harder when you see in terms of the technology development. And so what I want to say in terms of the difference between the industry as well, is that I really think we should, from safety perspective. We really think the safety performance as a probabilistic safety. So add to your point, in terms of proven safety, I think it should be a probabilistic proven safety, 100% 100%

I will add to that that it should be probabilistic, and you need to be for deployment. So it cannot be I drove billions, trillions of miles in the real world. So you need to rely on simulation. But you need to prove and verify that simulator. You need to prove the realism. You need to prove that that simulator cover all the cases that you might encounter, etc. And that's where, you know, the industry needs to go, you know, go to where it is today. There is the infancy to the next level. And that should be, you know, a requirement for any safety case, in my opinion.

So you don't answer, I think, yeah, safety is definitely important. But I think for autonomous driving, the key is actually not to guarantee safety, because it's not something that we can guarantee trucks can dodge bullets. I always remind myself about that. So in the end, I think it's more about the liability, and I totally agree with Henry on on one hand, we need to be 100% liable, non liable, for all of the road behaviors. But on the other hand, when it comes to dodging bullet time, we need to have a more probabilistic view on how can we minimize the safety incidents instead of like eliminating it, no way.

So we're running against time. Now, I just comment that compared to the session last year,

I really impressed how much progress we're making here, right hopefully a year from now, we're gonna see another level of progress and really brings the technology to benefit the general society. Please join me. Thank the panel for the presentation. Thank you.

Okay, you