Future Current

The DARPA Grand Challenge race required robots to drive without human assistance along a 131-mile long course through the Mojave desert. The competition was won by the Stanford team led by Sebastian Thrun, having developed significant new AI technology for robot perception and decision making. At the 2006 Singularity Summit at Stanford he discussed how autonomous car capabilities compare with human-level cognition.

The following transcript of Sebastian Thrun’s Singularity Summit at Stanford presentation entitled “Toward Human-Level Intelligence in Autonomous Cars” has not been approved by the author. Video and audio are available on the Singularity Institute website.

Toward Human-Level Intelligence in Autonomous Cars

I’m delighted to be here. I think the issues we are discussing are of fundamental importance. As Doug said, it is important that we take a serious look at the future of this. I am a bit of an odd-ball as a speaker because I’m not here to scare you. I’m here to tell you something about a technology that we are inventing at Stanford and many other places. Ray has this wonderful picture of what makes a human human, and to drive a car is still on the wall. From our definition of playing chess, something that involves real intelligence, to something you can do drunk, we are still left with certain areas that make humans humans. I am just about to take this away from you. At some point in the future, you will be better off trusting your computer than driving yourself.

This is in large part a talk about the DARPA Grand Challenge as a technology that is coming out of artificial intelligence that in many ways was thought to be impossible a decade ago. DARPA, our main sponsor in computer science, has put almost half a billion dollars into the development of unmanned vehicles. The vision there is that if you put self-driving cars into Iraq, for example, then you lose fewer soldiers and you can do different things. After all this funding they decided that the capabilities are not quite there yet, so they came up with a new model of saying let’s double the speed of the technology and make it more robust, and cut all the funding. Instead, give people prize money.

The race first took place in 2004. It was meant to be a robot race without a person inside a car and without any remote control. So basically a car programmed to go from Barstow to Primm. Originally they wanted to go from Los Angeles to Vegas, but there are just to many kids in Los Angeles. You can’t do that. So they went from Barstow to Primm. You push a button on one end, and something like ten hours later the car arrives on the other end. 2004 was the first time it took place. You might have followed it. At the time about 135 teams registered and fifteen raced. Half the teams built their own vehicles and the other half used off the shelf SUVs. Everyone was concerned with programming these machines to be smart, and what came out of this was just a disaster. The furthest any team went was 5% of the course. Some of them had no orientation. Some of them had GPS orientation from their home state, and drove backwards to their home state. Some had very severe control problems.

The competition got stiffer in 2005. Some people said that all of a sudden we invented autonomous driving. That is of course entirely wrong. We are building on 30 years of history. Let me tell you some of the basic problems that came up in building a winning entry here. First thing is you have to get a person out of the car and replace them with computers and sensors. Very basic AI questions, like what defines rotor, are actually fairly hard questions.

We had to come up with ways of doing very simple and mundane tasks which people can do very well. The technological solutions aren’t exactly humanlike, although there is some resemblance. This is part of the history of finding new ways like in chess, where we can build really fast computers to beat humans with completely different methods than humans play chess. They involved laser terrain acquisition and some sort of a probabilistic assessment of where obstacles are. A laser scans the road and can see vertical elevations. If you think of driving as an intelligent act, this is really dumb compared with what humans do.

From there it can actually build these wonderful road maps that allow it to assess where to drive. Red means there is danger, there could be a cliff, while white means it is drivable. The other lines basically are what the organizers gave us as a description of the race course, using GPS reference points. Then, to see really far, we had to move away from lasers, which don’t see really far, and build a vision system to find the road. It took us months to write a simple piece of software to find the road. What we ended up doing was instead of solving the whole problem of what is a road, we said our laser is perfectly capable of finding the road at near range, that gives us data of how the road looks like. Then we can try to find the same stuff further out.

Here is an example of this technique at work. You are driving through the desert. The red stuff is clearly road, and the blue stuff is not road. You can see at near range with a laser, and then use it to train the car. So the car adapts all the time, just like to some extent people adapt all the time. The adaptations are really powerful. When you go from one surface to another, you can see the color jumping quite a bit. We can use this model to extrapolate further out, up to 80 meters in distance.

After about a year of work we had a car that could drive hundreds of miles. We have a lot of video documentation. We weren’t the only ones. In fact, Berkeley was doing good work, and since this is the Bay Area I can’t resist putting up one more Cal video. The way these cars were then tested and selected from 195 to eventually 23 and then eventually one was through what is called a qualification event, where cars were tested in a relatively complex environment using barriers and parked vehicles. They had to make decisions of a simple and mundane form, not anywhere as complex as you do when you drive home drunk - like, find the road, and if there is no obstacle, ether go left or right. That was pretty much it. There were some high-speed sections and a tunnel that shielded the GPS perception. One of the key driving technologies here is something people don’t use at all for driving, which is GPS. That really helps the vehicles to understand where they are, so these tunnels take that away. This is taken from an hour-long NOVA documentary I recommend.

The race itself was one of these moments I will always remember in my life. It started at 4:00 in the morning. We were sent a top secret data file and gave us about two hours to put the data into our computers. It took us twenty seconds, and then we sat around and talked to the media four one hour and fifty-nine minutes. The race was a staggered race. Carnegie Mellon was the fastest at the time and they got the first position. We got the second, and so on. It was a total of 23 vehicles going out in the desert. Whenever I look at this, I have this feeling, I assume some of you are parents and have sent your kids to college. You work on this technology for so much time, like a parent, and then the thing goes off by itself. It goes out just like when a kid goes off to college, and you hope that it comes back without a dent, without falling off a cliff, maybe not pregnant, whatever.

You just sit there and drink beer, hang out and talk to the media. Larry Page comes by, and lots of other people. The cars drive, and they are pre-programmed to race. In this specific race, the race was very tight. Five teams finished and all of those were equally deserved to be called the victor, although we were happy to have the cash because we were eleven minutes faster. The race started about 80 miles in the flats. At mile 80 or so, Carnegie Mellon developed an engine problem and at mile 103 we were given the opportunity to pass the front running vehicle. It recognized it as a piece of metal sitting in the desert and successfully avoided the collision. From that point on we started leading the race.

It merged into a treacherous mountain pass. This was kind of frightening for us. We can see the cliff on the left side, and we fear that falling down could be a final catastrophic event of the type just discussed. Stanley descended this mountain pass and then emerged from the desert. Coming out of the desert, we were all sitting outside and waiting for our own creation to come back. It was a big moment for us, of course. First you see a helicopter, and then you see a little bit of dust, and you think there might be something moving. Then you start realizing, wow, the car could actually do it. Then you see a blue dot, and finally the thing comes closer and closer. Then you are being buried in water and champagne, and you have no clue what’s going on. You say all kinds of strange things to the media, then you get a $2 million check.

Critiques of this race, just like critiques of chess and other innovations, said “This is not the real thing.” “You’re not in traffic, you’re just in an empty desert.” And they’re right. DARPA created with Stanford’s participation the next Grand Challenge. It’s urban driving, so if you’re living in Palo Alto for the next 18 months, be careful; we do experiments.

It’s hard. There are really hard AI problems, in that we are able to perceive scenes like those effortlessly, whereas robots have a really hard time. Now, in the few minutes that I have left, I want to give my personal reflections on some questions. I don’t claim to be an authority, I’m just the director of the Stanford AI lab. I’m not a visionary. Does it take away from human identity? If you identify yourself with a fast sports car, then it does, and rightfully so. If you really care about that, when 97% of the human population used to be in agriculture, worry that you not working in agriculture takes away from your identity. I don’t think there’s an issue, because all we are doing is building technology that ennables people to be more successful, not to replace people. The core of AI today is about making people more effective, like the core of the industrial revolution was not about replacing people, but making people more effective.

Now, is this finally the advent of strong AI? That is actually a point where I have a profound disagreement with people like Hans Moravec. In all these exponential growth curves that I usually see, I see them for computing, for memory, but I never see them for core AI. Some of the key questions that we have, like how does perception work? I would claim that AI is still in its infancy. For us to even tell a tumbleweed from a bush from a rock is still a hard problem, as it was 30 years ago. If it is exponential, we are still in the flat curve in my opinion. The progress is not just faster computers or understanding the human brain, the progress is really also better algorithms. That has been much slower in my opinion. I don’t have a curve here. It probably is exponential, but every exponential looks linear for awhile.

Will this advance humankind? You bet, absolutely. The car is a young thing. In fact, right now the car is an invention that kills about 42,000 people a year. 42,000 people is just about as many as we lost in the Vietnam war. It is about fifteen times as many people as we lost in September 11. We never really think about it because we always think it is the drunk drivers that die, but it’s not true. For every accident there is a drunk driver who dies and somebody else. The vast majority of accidents is caused by human error. 1,500 people drive into trees and die. More than 10,000 die because of drunk driving or speeding. It is a technology that we are really bad at handling.

Aside from safety, you can look at productivity. The average American worker spends more time stuck in traffic, doing the same thing over again. Even with a Porsche, I can tell you it’s no fun being stuck in traffic. So, what if we take this time and make people more productive? They can sleep, carry out a conversation, read a book, or do email. Email I guess we do during driving anyway, but doing it in a safe way.

My personal story here is that my dad has ALzheimer’s and they are taking his car away right now because he is unsafe to drive. What that means is that he loses his independence. The classical way from there is that people deteriorate very quickly. You can look at young people. How about your kid can drive himself to soccer practice, or what if you could make blind people independent in terms of transportation?

Highway throughput is something that people don’t pay much attention to, but it turns out the national highway system is basically over capacity in areas like the Beltway, Los Angeles, the Bay Area, Chicago, you find these enormous traffic jams. If you take a highway at peak capacity, which is when it has the highest throughput, people drive about 55 miles per hour. If you take a photograph form the air and ask yourself how many of these pixels in my photograph are taken by cars and how many are still free, you find that only 8% are occupied. That is, 92% of a fully used highway is still empty. That’s because we are such lousy drivers. You need the side distance and the front distance.

You can talk about real estate. Half of Stanford’s campus is parking lots as opposed to office buildings because we don’t like to park far away. If we can have cars park themselves, then they could drop us off at our building and we could re-utilize the space. If you look at this, this is not a claim about artificial intelligence, but there is something coming up that is going to be disruptive. It’s going to be major. It’s going to look natural from the other side, even though it looks uncertain from this side.

People have asked me to put a timeline on this, and I’m not very good on timelines, but I think in the next two years in 2008 we will probably solve the next Grand Challenge. This means in only a five year period we made a significant advance in a field that we thought to be generically human and uncrackable by humans, which is reliable driving. It is going to take time to make these things reliable. 2010 might be a little bit optimistic. Maybe 2015, but no more than that. Then we have to worry about legal issues and societal issues for this new technology. At some time we will drive more miles autonomously than we drive manual. Maybe it’s 2030. That is actually not all that unrealistic if you think about it. Twenty years for a technology to mature is actually today a lot.

I hope with this presentation I could give you some insight in what is happening on the artificial intelligence side and robotics side. I think there is still a very clear line to be drawn not to confuse any of this with human-level intelligence. It’s all smoke and mirrors. There is no attempt and no methodology to take this and turn it into something that threatens us. I think what it really does is empower us to be more effective people. There will be identity beyond human driving for all of us, but I think it is a good example of artificial intelligence moving to the next generation of technology. Thank you.

This entry was posted on Wednesday, December 26th, 2007 at 11:55 am and is filed under Artificial Intelligence. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.