Ferguson Research

Transcription

DAVE FERGUSON: The other Google X moonshot that I'd like to speak about tonight is self-driving cars. So these are vehicles that can navigate without any human input whatsoever. And when I talk about this project to people, they sometimes respond, I can drive. I like driving. What's the point? Why aren't you guys working on something more useful? So indeed, what is the point? Well, for starters, time. In the US, we spend an average of 50 minutes a day commuting to and from work. In New Zealand, we drive an average of 30 kilometers every day. And this is an incredible amount of time that could better be spent working, sleeping, relaxing, watching cat videos on YouTube. Self-driving cars could give this time back to their occupants. Now, it's estimated in the US just this time due to commuting is a wasted 50 billion hours a year. Secondly, they could also be more efficient in how they use the roads that we already have. So they could drive closer together. They could also drive on narrower lanes. And a recent estimate has shown that just this optimization could save 10 billion liters of fuel a year in the US. Secondly, they could be a lot safer than the current vehicles that we have. Worldwide, we lose over a million people every year to traffic accidents. It's the number one cause of death for young people in the developed world. And over 90% of these accidents are due to human error. Imagine if you had vehicles that never made mistakes, always paid attention, never tried to tweet while eating a Big Mac and drive all at the same time. Now, these are two of the perhaps more obvious examples of where driverless cars would benefit society. But this is really only scratching the surface. Imagine not needing parking spots or parking lots at any places of interest. Imagine not needing to own a car because there's always one available on demand for a fraction of the cost. Imagine vehicles that didn't get in collisions, so they didn't need heavy collision reinforcement, which meant that they could be much, much lighter and thus, much, much greener than the cars we drive today. And then imagine driverless cars that could run errands with or without people on board. Every time we have come up with some new technology, society has found ways to make the most of it. Think of the internet or GPS as two very recent examples. And yet, in general, humankind has a glorious tradition of not imagining what things could be like, of being anchored to how things are now and thinking only incrementally. So we look at the accidents that we have from driving. We look at the time that's wasted commuting. And we realize that these are inefficiencies, but we accept it as the tradeoff that we make for the convenience of driving, our current relationship with the vehicle. But why should we accept this when we could do so, so much better potentially? And then finally, does anyone disagree that they're kind of cool? I mean, who wouldn't want KITT from Knight Rider driving them around? Now the idea of tackling driverless cars may seem a little bit crazy, a little bit far-fetched. But maybe it's just far-fetched enough to be worth doing. And certainly, not everyone shares this opinion. Even some heads of car companies believe that this is never going to happen, that it's an impossible technology. But to me, that's even crazier than the idea of driverless cars. It's impossible that this isn't going to happen at some point. And of course, there are hurdles. There are lots of hurdles. I find it hard programming my TV remote. Imagine how hard it is to get a car to drive itself. And beyond the technology issues are the policy issues, the liability issues, the regulatory issues. But that's OK. Rarely is anything worth doing ever that easy. So we decided to give it a shot at Google X. And we asked ourselves how can we try to solve this problem that has seemed for a long time to be intractable? And the first idea was to put all of the intelligence and the sensing on the vehicle itself. So a lot of current and previous work has focused on changes to the infrastructure, to the environment. So cars can be equipped to talk to each other and tell each other where they are and what they're doing. Cars can talk to traffic lights. And the traffic lights can tell them what the color of the light is and when they should go. Cars can track markers or magnets in the road to tell them where they should drive. Now, this is all fantastic work, and it simplifies the problem enormously. But it means that we have to wait for this infrastructure to be available before we can unleash this technology and the promise of driverless cars. And we've already been waiting a long time. There was a very successful research project where they put magnets in a road in a highway in California and had cars track it. That happened in the early '90s. So we've been waiting a long time. And if we can put all of this intelligence and sensing on the vehicle itself and not need to rely on any changes in the environment, then we don't need to wait anymore. The problem is that's really, really hard. So some of the sensing that we might put on would be radars that can detect where other vehicles are and what speed they're going, lasers that can generate a three-dimensional representation of everything around the vehicle, and cameras to detect lights and signs, traffic lights, tail lights, markers on the road. But perhaps, even though this seems like a very, very hard way of solving the problem, maybe there are other things that we can do to make it slightly easier. And the first idea of making it easier is through mapping. So what if we were to map the entire world and then use that map to tell us what we should be doing at every point in the world? Now, at first, this also might seem a little crazy. And to be honest, this was one idea that I thought was a little nutty to begin with. But what if we were to give it a go, and then see what that allowed us to do that we couldn't do otherwise? And maybe we can figure out how to do this mapping itself later efficiently. After all, we do have a couple examples of companies that are able to keep world-size maps up to date pretty effectively, wonderful companies with wonderful maps. The second idea that we had was to focus and simplify the problem, and then simplify it some more. Now, you might think that in order to drive in an environment, you need to understand everything that's going on in that environment to make the right decisions. But this is an incredibly hard robotics problem. It basically means that we need to solve what's known as the artificial intelligence problem, where the robot has to have full common knowledge and common sense to reason about anything that could happen. But in fact, the act of driving requires coming up with exactly two values-- what angle to have your steering wheel, and how hard to push the gas or brake. That's it-- two numbers. So perhaps we can take all of this complexity that's in the environment, and we can filter it down to just the key components that really make a difference in us coming up with those two small numbers. And so that's sort of what we've tried to do. So we take everything that's in the environment, all the vehicles, all the pedestrians, all the static objects, and we try to filter it down to only the things that matter. And then we only consider those for our task. And basically, once we have this problem transformed to be as simple as possible, we get a bunch of smart people together, and we try to tackle the bits that remain. And ideally, people that don't know any better, that don't realize just how hard this problem is. And that's what we've done. We have a fleet of vehicles that have driven over 700,000 kilometers, mostly in the northern Bay Area. This is fully autonomous driving. And it's been in all sorts of different road situations. So we drive in suburban streets. We drive in hilly areas. We drive at night. We drive during the day. We stop for baby carriages. We stop for red lights. We hopefully don't stop for too many green lights. We drive through toll booths, highways, bridges, congested traffic areas, heavily congested pedestrian areas, and even down Lombard Street in San Francisco, which is the last clip here. Now, in a little more detail, the entire system can be broken down into just a few key steps. The first step is to get an idea of where the vehicle is in the world. So to do this, we first get a rough estimate using GPS and some inertial sensors that we have on the vehicle, such as the speed of each wheel. Now, this tells us which road we're on and maybe what the closest intersection is. But it doesn't tell us exactly where we are in our lane, for instance. So we need to improve this estimate. And we do this by using Maps. So the idea here is that we take the laser that's on the vehicle, and we look at what it can see around it, and we compare that to what we have on the map. And by doing that comparison, we can figure out where in that map we are, based on what we can see. And this tells us very accurately where the vehicle is. So we know where it is relative to its lane, where it is relative to crosswalks. And once we have this accurate position, we can then overlay some of this key information about what the world is like and what the vehicle should do. So we can add in where the lines are, where the boundaries are, where crosswalks are, where intersections are, where the nearest Taco Bell is, all of the really pertinent information that it might need. And we can then augment this with dynamic information about what the vehicle is seeing using its onboard sensors. So this is all of the other vehicles that are in the vicinity, pedestrians, traffic lights, stop signs, and so on. Given all of that information, we then filter it down, as I mentioned before, into just the key components that are important for making the decisions of where to drive the vehicle. And we then come up with a trajectory for where we would like it to go. And this consists of a path through the world, along with a desired speed at each point along that path. And this takes into account things like slowing down for a vehicle in front of us, stopping at a stop sign, going through a green or red light, and so on. And once we have this trajectory, the speed profile with the position estimate, we then feed this to the vehicle and have it executed. And we then repeat this entire process several times a second. And the resulting system looks something like this. So on the left here, you have a video taken from an onboard camera looking out into the world. And here you can see this run was done at night. And on the right, you have an internal representation of what the vehicle is seeing in the world and what it's reasoning about. So here you can see it has static obstacles. It has its map. It's detecting other vehicles. Here the boxes represent other dynamic vehicles that are on the road. It's also reasoning about traffic lights and intersections. Here, we started out in a suburban area, and we whipped onto a highway. We then zip along the highway for a while and pop back off. So this all looks pretty good, right? We have a vehicle that has driven hundreds of thousands of kilometers in all sorts of different conditions. Seems like we've sort of got stuff figured out. Maybe we're done. Maybe it's time to unleash it on the world? Well, not quite. Unfortunately, when you're operating in real road scenarios there are a huge number of special cases or weird anomalies that you might need to deal with. One such group of anomalies are interesting vehicle shapes. And when you're first putting together a vehicle detection system, you might not anticipate that you're going to have to encounter cars that are shaped like hot dogs. But if you drive enough miles in enough places, particularly in the US, you will see all sorts of interesting things. You also need to deal with very erratic behavior that people in vehicles or pedestrians on the street may exhibit from time to time, perhaps after jumping off one of these vehicles on the bottom middle here. As well as the complex vehicles that you might need to deal with, there are also road scenarios that we might need to deal with. Now, we love using Maps to improve the performance of our system. But the road can change based on construction or re-painting. And we need to make sure that we're able to detect these situations and respond to them safely. We also have to deal with dynamic road situations, such as accidents or emergency vehicles that may come in our path. And then finally, weather can present quite a number of challenges for us. If the road is entirely covered in snow, if our sensors feel the view is occluded by rain or heavy fog, we have a number of additional things that we need to solve as part of this overall problem. And then finally, we have extra-special situations, like this wonderful fellow here. Basically developing a system-- [LAUGHTER] DAVE FERGUSON: --developing a system that's robust to anything that the world can throw at it is really, really hard. But again, this is the good stuff. It needs to continue to be hard to make sure that people are still interested in working on it. And we are still very much interested. So what's next? What's next for the project? What's next for driverless cars in the world? Well, let's take a look. [VIDEO PLAYBACK] -Hands-free driving. Cars that park themselves. An unmanned car driven by a search engine company. We've seen that movie. It ends with robots harvesting our bodies for energy. This is the all-new 2011 Dodge Charger, leader of the human resistance. [END VIDEO PLAYBACK] DAVE FERGUSON: So I love that commercial. And you know, Google does have a lot of power-hungry data centers that could use the energy. But actually, we're more interested in a future that looks something like this. [VIDEO PLAYBACK] -OK. -Well, Steve. -OK. -Off we go. [CAR BEEPS] -Auto-driving. -Here we go. -Away we go. -Look, ma, no hands! -No hands anywhere. -No hands, no feet. -No hands, no feet, no nothing. -I love it! So we're here at the stop sign. -Yep. -Car's using the radars and laser to check and make sure there's nothing coming either way. -I find myself looking. -Old habits die hard, man. -Hey, they don't die. Hey, anybody up for a taco? -Yeah, yeah. What do you want to do today, Steve? -I'm all for Taco Bell, myself. -All right. Well, let's go get a taco at the drive-through. -Now we're turning into the parking lot. -Yeah. -How neat! -There we go. Now we kind of creep along here. Does anybody have any money? -I've got money. -No, I've got my wallet right here. If you roll down your window and order a burrito. Yeah, push that. -How are you doing today? -I'm doing very well. How are you today? -Good. Thank you. -This is some of the best driving I've ever done. 95% of my vision is gone. I'm well past legally blind. You lose your timing in life. Everything takes you much longer. There are some places that you cannot go. There are some things that you really cannot do. Where this would change my life is to give me the independence and the flexibility to go the places I both want to go and need to go when I need to do those things. You guys get out. I've got places I have to go. -Yeah. Bye now. -Hey, it's been nice, you know. It's been nice. [END VIDEO PLAYBACK] DAVE FERGUSON: Most of you here today live in the developed world. You and I have already won life's lottery and been given the opportunities that we have. What are we each going to do with our winning ticket? And you might think that driverless cars or internet from balloons are kind of silly ideas. And that's fine. But what is it that impassions you? What might you push forward in the world? Can New Zealand be the first country to use 100% renewable power? Can we solve the diabetes or obesity epidemic? Google X and Solve for X is all about trying to take on these big, massive challenges and try to make an impact. There are so many huge problems in the world and so many that could benefit from some Number 8 wire ingenuity. So you tell me, what is next? Thank you very much for your time. [APPLAUSE]