ai.bythebay.io: Christopher Heiser, Autonomy at Scale: The Need for Interop
Recording: ai.bythebay.io: Christopher Heiser, Autonomy at Scale: The Need for Interop
you so first of all thanks to Alexi and Mason and AI by the bay for inviting us to talk to you today my name is Christopher Heiser I'm the co-founder and CEO of Renault vu we are a company that's focused on automated mobility on-demand we're going to talk a little bit about what that means to us and how we it is affects the way we think about a vehicle design and systems design so I think what's important to think about right now is that automation is a relatively poorly understood field if it were understood people would have commercialized the product by now and there's a lot of different approaches there's a lot of different even agreement on what sensors are going to be used how they're going to be the deploy who's going to control them none of this is really understood but the goal is to actually get this to scale that once we operate these fleets in hundreds or thousands and tens of thousands of vehicles in cities it transforms a lot of important things we'll talk a little bit about that but I want to get into some detail as we go along a little bit about us as a company we're based here in the Silicon Valley we are a team of technologists some of us with backgrounds from big companies here in the Silicon Valley and we are thinking about how what we've been working on over our careers affects the vehicle we're also very deep into cars we understand safety critical systems at a low level we build all of our own prototypes in-house we're very self-sufficient and that's important because as we're going to talk about in a bit here automating the car is actually a collision of two very different worlds and and they don't speak the same language there's a sort of massive impedance mismatch between these two organizations automotive and and self-driving so that's one of the things that we work we focus on we also test a lot and we like to find tough environments that are very extreme so we'll show clips from one of those test programs as the partnership with Stanford Stanford is doing research to figure out how self-driving can improve safety by controlling cars at the limit of traction so as we can teach cars to handle vehicles outside of normal conditions we can prevent more accidents and so this is a program that we work on with Stanford this is Marty and it is a self-driving DeLorean and so this is Adam Savage's first ride in it from Mythbusters and it takes us about 90 seconds to boil a set of tires off of Marty so if the test is mostly us changing tires than anything else but the point is that in these environments small errors result in massive problems when you're driving at 10 miles an hour on a city street if you make an error there's lots of time to change that you can run your autonomous loops that slow is maybe 50 or 100 Hertz in this environment if you make a tiny error the car spins out of control so we like these environments because it pushes our technology and our systems to the limit and allows us to think about how it's going to affect systems like this Shanghai 24 million people so the types of things that you see on a skid pad in a city are very different but at some level of the system it begins to look the same and that's that interface between these two worlds is a big focus of our company so a tiny little bit about me I went to Carnegie Mellon that works on robotics when I was there I worked at I do product development down in Palo Alto building really cool products I worked on pioneering the camera phone so I loved disruptive businesses but unlike most people in a self-driving car world I actually liked driving I'd like to really enjoy it it's fun I do it to clear my head and I want to sound like it's the nostalgia thing more importantly driving or the freedom to move from place to place I think it's very fundamental and when you look at societies that are becoming developed nations it's one of the first things that people reach out to some of the greatest growth for automobiles is not in the US it's certainly not in Europe where GM is about to divest itself of its entire car business there to Pujo its in places like China in India so driving is something that that speaks to a freedom that we have and a lot of people at Renault go really identify with that and one of things are trying to figure out is how does that translate to a car that you may not drive or you may not even own that's a very interesting problem as well because I enjoy driving and racing I'm attracted to figuring out how these two things work together so I'm an accounting adviser to self racing cars created by Joshua Schachter so you proud native heard about this earlier in the talk today so I won't dwell on it too much except to say that the real beauty of this is it is a grassroots movement it's an autonomous racing competition that's designed to make things inexpensive and easy for people to do so if you are a researcher an AI or you're working on anything to do with automated systems come to the event there's an event early next month we'd love to have you out there you'll see lots of cool stuff you may even find applications for the things that you're working on but what I think's more interesting about this is that it's a sort of a counter to Robo race to some of the very expensive systems that only a few groups will get to work at and Joshua's dedicated to making it be something that's going to be accessible to everybody hopefully even down to high school level so that's super exciting so self racing cars it's on the web so I one of these setting things about addressing a technical group of people is that you all use logic and math and reason on a regular basis so what we're going to do is we're actually going to the reason why the a mod car is not only the apex of what we should be working on but kind of the only thing we should be thinking about long term so unfortunately that means audience participation so hopefully people aren't too uncomfortable with that so let's start with just briefly defining what we mean you pricing this chart earlier today if you haven't I apologize for my previous speakers this is the self driving automation scale for cars when you hear level 1 level 2 level 5 that's what this means we'll have that we'll put a PDF of the deck with Alexia you guys can get this but the bottom line is that when people talk about fully autonomous cars or self-driving cars what they're really saying is highly automated vehicles and in the SAE standard which is now pretty much being adopted by ever in the world that means level 4 and level 5 and the critical thing is that this means the human is not in the loop there's a little subtle feed of difference between level 4 and level 5 in level 4 the vehicle can ask the human to intervene but the human is not required to intervene whereas in level 5 is expected that the human has never asked to intervene so when we talk about it we try not to use the words autonomous we try to talk about automation because autonomy really has to do with choice and automation has to do with repeating a process accurately to a specification you don't want an autonomous coffee maker that chats you up and decides that you'd rather like your latte a little bit differently today then yesterday you want an automated coffee machine that makes it the way you want it on demand so this is really important and you'll see that in the technical information and the specifications people get very very specific about the difference between these two so we're focused on high level automation so just by a show of hands who believes that this is a bounded problem that we can solve we can have a car drive itself automated you guys should all be optimist because you're in AI so so please raise your hand so hopefully it is what's interesting about this is that even in the 1960s people understood what the hard problems weren't they said hey this thing is going to drive on this magnetic track and RCA is going to beam radio waves down to it but when you get to the intersection you're gonna have to take over and steer it I thought that was kind of interesting that even you know 50 years ago 60 years ago people understood some of the conceptual problems with driving and what would be hard what would be easy so it's a bounded problem and I think if you're George hops your argument is that bounded by the power of a cellphone if you're if you're Danny in a video you're hoping it's a trunkload of px2 s but either way it's somewhere in between these two worlds that that the computing power is there and the advancements in machine learning particularly with object classification vision processing have made things a lot better today or five or ten years ago so we kind of see a light at the end of the tunnel so we'll automated systems be better than humans so once we've built these systems and we train them who thinks in the aggregate these systems will have fewer and less serious crashes and incidents than humans okay good that's awesome so your right eye and the important thing to think about by the way that the fun bit about this I directed this Total Recall movie 1990 probably is is older than some of you here there's a fantastic scene where he's driving around and sort of conversing with this robot and in the end Arnold decides not to pay and the robot tries to kill him so this really is an autonomous system in the fact that decides to go after the people inside of it but anyway there's a fantastic article in the Atlantic from three or four days ago about how AI is really like what people think AI is in the movies and I think some people self driving is what a self driving is in the movies it's Minority Report at this but anyway so the good news is that most of the problems that happen on roads are due to people anyway so if you fix that problem you solve the vast majority of injury and deaths on the roads by the way 1 million people over a million every year die of road injury globally it's one of the top 10 killers of humanity amongst with cancer and other diseases but more importantly most of these crashes are not someone was coming it's not a trolley problem people know a trolley problem is right if that's not it there's not thousands and thousands of trolley problems there's not thousands 2000 that you can see intersection and a boat was crashing down and there was a blimp exploding it's not that you were texting you fell asleep you have too many drinks you were yelling at your kids that's what happens automated systems are not you ask them to are not going to yell at your kids they're not going to make these kinds of simple unforced errors that humans make all the time we're really good at driving at a kind of a minimum level across lots of platforms so we're actually kind of crappy at driving and and these automated systems are going to fix a lot of that so great so it's a bounded problem and the robots will be better than us so once we've achieved that level who thinks that having a car that also allows you to drive it in addition to you driving it will be more expensive than one that just drives itself okay so people raise your hands you're seeing a pattern here yes it's always the answer yes because it's really expensive to do this as soon as you put humans in the loop you have safety critical systems you have to design you have to validate that you have to build crashed structures around that you have to integrate into the cabin pedals steering wheels buttons knobs display that center console is an acyl rated device that means that it takes about two and a half years to design and build it and you cannot really change it that much once it gets into production there are some companies like Tesla that are pushing quite hard on this but if you ask people at the major OS a sake don't touch the cluster because it's so critical to the safety function of the driver this is a big problem it's a lot of cost and if you're just driving autonomously it is dead weight and more importantly its expense which means that adoption is slow if you look inside the way no car there's nothing inside it to drive now not true exactly they can take the center console out and stick a steering wheel in there and they do from time to time if you feed a little slot down near the footwell you can plug in brake settles and gas pedals but lame-o built a vehicle that could be seen as a fully level 5 a mod car and it doesn't have any of that there are other people that have experimented with this idea this is rolls-royce and sort of the opposite end of the spectrum from the way mol Koala car I this is a you know I think five meter long luxury sedan but all it has is super comfy seats and they're really awesome TV in the front of it by the way this is the reason why a lot of other companies are interested in a mod because they get to put what's on that screen and today that's a big problem for them if you look at Apple's experience inside even companies they partner with like W and Ferrari it's kind of crappy but if there's a giant screen and there's nothing safety-critical going on it fire up Netflix do whatever the hell you want it is a completely open environment so for a lot of reasons a mod is also what content providers and communications companies want so so it's not going to have a steering wheel who can tell me what this is good it's a map the sweats right this is a visualization of the roads of the United States so every dark line is a road and every not dark line is not a road and this is relatively in the middle of the country so the amount of roads you need to drive on to traverse all of the United States is massive and when you buy a GM or Toyota or a BMW the implicit contract is that your car's going to work on all of these roads and in fact at least for now roads in Canada and Mexico so you can buy your BMW and drive it anywhere in the North American continent that you want that's a really hard problem to solve and it explains why when you hear someone from Daimler say we think self-driving is something that is solved in 15 to 20 years I agree if what you're trying to do is have your 5 Series or your your C class do this but having a vehicle do this is an order orders of magnitude smaller problem it's still really hard but once you solve this you get 80% of the population in the United States if you just cover the urban areas so this is called geo fencing the concept that these vehicles will never be able to leave these areas and once you agree on geo fencing the cost and the validation profile if your product comes way way down and that's really important so these cars will be geofence what about electric today if you buy a car I'm sad to say I'm a huge fan of v's EVs are more expensive than gas cars for the same functionality and that will continue to be the case for some time now it turns out that the reason for this is how many miles you drive it it is cheaper per mile once you amortize the cost of it but at ten or fifteen thousand miles we can't do that doesn't make sense to buy a Model S based on economics but if you drove your Model S 30 40 50 thousand miles it starts to and the fun thing about these a mod cars is they're going to be driving tens or hundreds of thousands of kilometers every single year so evey won't be something that's driven by regulation or subsidies or anything it will be driven by the people operating these networks buying them because they're cheaper they will make more money if they use Evie so Evie will be cheaper and on the flip side a mod is going to drive the adoption in cities i forget the fact that cities are probably going to ban any views anyway in the next five to ten years but without that you don't need the regulation you don't need to push from government to make easy work in this context and so finally who thinks the model on the right is a more scalable model for guang growing flowers in the model on the left yes absolutely the idea that you're going to buy a car and then let it go and do errands or make money for you while you're at work I think is enticing and it sounds cool it's totally not scalable or another way to put it is that the first person that does that at a fleet level will drive the price down to the point where it's no longer profitable for you to do it so the notion that owned cars cars you buy are part of a mobility on-demand solution for cities it's not sustainable in a large economic model these fleets can buy in bulk they can charge them all in one place they can do their own maintenance they can do predictive maintenance they have massive advantages they can profile how many times they crash and negotiate better rates or the insurance companies no individual has that benefit and even if a company like Tesla steps up and tries to offer some of those benefits a well-managed fleet of these vehicles will always find a lower cost point than an individually owned car just say nothing in fact of the financing around individually owned cars is really really as compared to what these guys do in fact these ones mostly people don't even own their own vehicles they're bought by banks and wreaths and at least to them so this this models already understood for Hertz and Avis and everyone else in the world so great so when we say a mod what we mean is a fleet managed geo-fenced electric car without a steering wheel that's summoned by now and our argument is that this is not just the effects it is the only solution that operates at scale inside of large city centers which again is where 75 to 80 percent of the developed world live so if you solve this problem you solve a massive transportation challenge and then stuff that we can talk for another talk at you know Eco by the bay all of the massive advantages to parking and congestion and sustainability and energy management and pollution and and and there's massive benefits to doing this I'm going to kind of accelerate through so I think I'm running out of time how fun how far have I gone ten minutes oh yeah okay so lots of people are thinking about a mod some people have bad days yesterday Volkswagen launched this and Jalopnik sorry verge called it an angry toaster and and Jalopnik called it a space cat accordion for a caterpillar from outer space anyway sometimes it doesn't go well but people are trying this is my colleagues and I do thinking about mobile workplaces on demand so there's lots of ideas people are playing with them some of them are building concepts some of them are building prototypes you're going to see a lot of this but today in the marketplace those things aren't reading with exception of way mo no one has really built one of these things and driven it around they're all using production cars and this is one of the big challenges that this car is not a platform for development it is a highly lockdown embedded proprietary wasteland of development and anyone that's tried to build on these has a lot of challenges and there's lots of companies popping up to help people do this some people are taking different approaches this is oops this should make John Eggert really happy there's eight Vala dimes yep so this is good for the sensor guys they love this some people ilan says no lidar that makes John unhappy but but it I think there's there's a lot of people doing tests out in the field in different ways some of them are partnering lightly with Alize like uber and Volvo and some of them have really really deep connections like acquisition in the case of cruising n GM but the funny thing about this is is that if they're going to build an a mod cars that's the pulling of all this work they're actually doing something that was is going to reduce their sales by massive amounts that this is then this is why the car companies are having a bit of a I think a challenge in fully embracing this that if you have lots of vehicles in these cities that are shared the number of vehicles people need to put into their driveways drops precipitously and the total number of cars drops might not be five to ten but even if it was a 20% drop in their sales that would be earth-shattering for any one of these ohi's so this is a big challenge and they're trying to figure it out so why is this a big deal can anyone identify this awesome city yes thank you Austin City about eight hundred ninety thousand people in the city center two million if you account everybody in the suburb so study was done last year and I said hey would you want to use these a mod services 41% of people said yeah they would use at least once a week thirteen percent said that they would use it all the time at a dollar a mile now it cost about 80 cents a mile to operate a vehicle to own and operate per mile of vehicle yourself in a big city and that's parking and everything else factored in so dollar mile is really really close to that and there's a really decent take rate what's interesting about it is that when you do the math on that if you achieve that take rate your net revenue of that service would be bigger than Ebers total net revenue globally in 2015 so this is Austin Austin has one tenth as many people as Shanghai if you look in the united states the total opportunity three services over a hundred billion by 2025 some of it has challenges and that it snows sorry Boston and New York but it snows and that's hard but if you just look at Texas alone over ten billion dollar opportunity over the next seven years so this is really really interesting and it pushes us to develop technology that can get there and eventually scale up to the size of a city like Shanghai so I am awed and AI this is where the big challenge is I think for all of us it's not a technical talk if there's not an SK CD in your talk so just to nail that I'll give you guys a second to read this and appreciate it so so when when we talk about self-driving cars we're sort of missing the point that's not what these things are it's not a self-driving vehicle this is replacing the human component in a vehicle that's very different that there are people that are looking at AI on object detection classification tracking Asian intentionality all these great and we work with companies to do all of this but those are pieces very important but the pieces of it that in order to get this to work together you have to solve a bigger problem you have to replace the taxi driver the uber driver and all the things that person does drive the vehicle talk to the passengers talk to other people outside make sure you haven't passed out in the back right there's a lot that the human does and we do it pretty well that we have this amazing set of sensors they're kind of crappy by comparing it to lidar lighters really great your eyes are not that great but they're deeply embedded into our system and there's a lot of distributed processing that happens in a human and so the notion that you're going to build a brain for a car I think is really dangerous I think it under serves it doesn't describe what you're doing and it makes it sound like there's going to be an intelligence that just runs the whole car and everything else is going to be sort of dumb and that's not the case that's not the way your body works it's not the way that that any sort of complex system works today that you wind up distributing because you have to solve lots of different challenges and we're really good it's amazing that human can navigate these types of diverse situations but when you look at it this is what we're replacing and so when you talk about a light or our camera you're talking about an embedded sensor that is connected to systems in your body that are processing at multiple levels some of them completely autonomous right this is this is what's happening the brain is half of the thing at the top of this everything else the autonomic systems the sensor systems motor control when you pull your hand off a hot stove your brains didn't have anything to do with that when you there are things that happen slow level that are important for us to think about because we're talking about taking this whole staff and re representing it in a vehicle and today these vehicle systems are deterministic hard-coded embedded computers hundreds of them some in some cases and kind of cars so those systems have very simple API is but the idea that you know there are sensor companies today that are putting object tracking inside of their their sensor so how do you understand what their object means to your higher-level system and which one do you trust and how do you negotiate that discussion the idea you can send everything to a central processing unit and do it all there that's one approach but it's not the best for latency it's not the best reliability and it's almost certainly not the best for cost the edge processing and processing along the network and having AI along the way in between the edge the motor control systems and the sensors and the centralization is going to be keyed in making these systems scale so it's not this this is this is a bad metaphor that will of course be repeated millions and millions of times in in tech press for the next ten years but what's really needed is for people to figure out ways to take components in these systems low-level systems networking systems high level processing cloud processing and figure out how to work together that it's unlikely there's going to be one central thing that just does every thing it's going to be more distributed than that it's going to be more brokered than that and this is when you start thinking about scaling up from one car to a thousand cars to fleet of tens of thousands of cars in a place like Shanghai there's a lot of value in being able to move this information around and share not just from the vehicle but from vehicle to vehicle but there's no framework for this today there's the car companies are completely locked down there they're walled gardens they're Apple there they're embedded systems are completely you know unavailable from the outside and there's also not an understanding of how other companies will work together to solve this problem and one of our big views on this is that having a bunch of vertically integrated companies plow capital into solving this all by themselves we've seen that before in Silicon Valley we see it like once every 10 years and then it then it falls apart and so I'm hoping that falls apart in this space a lot faster because I would really like automated with the building on demand I think it'd be really good for everybody so so we do a lot as a human and trying to figure out how we represent that in a system is harder than just thinking about it's a computer program running on a PC in a car and one of the things that we work with our systems that help this so our platform shuttle around massive amounts of data just to give you an idea if every car company in the department if every car sold us are running our platform we'd process about eight times as much data it's Facebook does on an annual basis so this is a you know exabyte scale kind of problem and there are lots of different players that are trying to get at shared data inside of the car and that's a big problem too not just from bandwidth and latency but also security reliability and things like that so these are tough problems these are the kinds of things are working on and so we work with companies that are developing AIS and sensors and networks and things like that and we work with our own AI systems as well and so we entreat you that if you share our vision that a mod is the future and maybe the best future for transportation we'd love to talk to you find out what you're working on and see if there's ways that we can collaborate and I think I have one minute left so if there are any questions I'd love to take them thank you in the back sir oh okay one little thing maybe it's a misunderstanding you have this example where a self-driving car actually was offered for the cost of $1 per mile while a self operated car was cheaper so my question is when one of your points was that autonomous car automatic cars are significantly significantly cheaper why would you offer it for $1 per mile and why would people use it if it's actually more expensive than owning their own car what you don't understand the people who say I would use distance that's the cheaper version of it yep so I don't have the full that's a great question so I don't have the slide here but I'll put it into the PDF the target right now is actually lower than 80 cents per mile the study that was done in Austin asked the dollar per mile question and they got a take rate I don't think most people know what it cost to operate their own car per mile here's a fun fact if you live if you live 10 miles away from work how much money did you take out of your retirement by living 10 miles away from work as opposed to working at home how much money did you lose if you had taken the money for commuting and invested it at kind of standard low return rates for a 30-year career answers one million dollars so people don't make rational decisions in the space I think the illustration for us was that even at a dollar a mile people were very interested in the service when you look at the cost curve the target really is for us to get under 50 cents a mile we think that 50 cents is what people in the suburbs pay for a Toyota Camry if you can get below that 50 Cent's then its sky's the limit everyone will almost everyone will jettison their cars for a mobility solution but but people make really poor channel decisions uber drivers another great example if you do the long term analysis of how much your asset is depreciating versus what you're making versus your time driving for uber doesn't make sense we're kind of attacks on people that cannot do net present value calculations on their assets and and so so I think there are lots of irrational decisions I think we're trying to say is there is a really good cost curve to this technology and the human is 50% of the cost of existing drive a ride hailing services so getting that out helps then managing as a fleet optimizing the car around not being driven each one of these things gets you lower and lower to the point where there's no way and own the vehicle of any configuration will compete on a per mile basis than one of these cars thanks again appreciate it [Applause] [Music] you