Self-Driving Cars Better Than You-Driving Cars
We are fast approaching an era where robot-driven cars will not only be practical but mandatory.
Rachel Swaby asks, “In a Race Between a Self-Driving Car and a Pro Race-Car Driver, Who Wins?”
A self-driving car and a seasoned race-car driver each speed around Northern California’s three-mile Thunderhill Raceway loop. Which car will get the fastest time?
Before you place your bets, a little setup: Send a pro out on a racetrack, and the driver will automatically find what is mathematically the quickest route around it. They navigate with such adept muscle memory that elite drivers can handle sudden changes in friction on the road without increasing cognitive workload.
The autonomous vehicle is a creation from the Center for Automotive Research at Stanford (CARS). “We tried to model [the self-driving car] after what we’ve learned from the best race-car drivers,” explained Chris Gerdes, the program’s director, yesterday at The Atlantic‘s Big Science Summit in San Jose, California.
So who would win in a battle between skilled human and a wheeled robot imitator? Humans, of course. But only by a few measly seconds.
“What the human drivers do is consistently feel out the limits of the car and push it just a little bit farther,” explained Gerdes. “When you look at what the car is capable of and what humans achieve, that gap is really actually small.”
Ninety percent of accidents occur because of human error, and even a really smart algorithm isn’t going to maneuver a car out of every dangerous situation. Gerdes spoke specifically about the problem with recognizing pedestrians. You can teach a car to recognize something with two arms and two legs as something to avoid, but “Go to [San Francisco’s] Castro for Halloween, and the pedestrian system needs to recognize strange cases.” Costumes or not, Gerdes believes that by teaching a car to operate at the level of our most skilled drivers, we’re better equipping them to take care of us.
Gerdes can imagine a scenario where cars would assist drivers in difficult situations. Hit a nasty ice patch? All that skidding around Thunderhill Raceway has taught the cars — and their programmers — how best to deal with the slippery situation.
Two things are noteworthy here.
First, the comparison is between the very best human drivers and a self-driving technology that could be easily replicated. That means the self-driving cars are already better than most of us. And they’re only going to get better.
Second, tough, this is in a very controlled environment. While racing at very high speeds is dangerous, there are fewer variables at work than would be encountered in an ordinary commute. I’m not sure I’d be ready to put my fate–much less those of my girls–in the hands of a robot car while dealing with the hazards of driving in the DC metro area.
Still, that self-driving cars are already this good is impressive, indeed. It shouldn’t take long for them to be safer and more efficient than human drivers even under the most complicated conditions. And, at some point, that will make it unreasonable to let human beings drive their own cars, given how hazardous that has proven to be.
If (when) we get there, that’s going to be a tremendous change in the nature of transportation in this country.
And insurance. I know there’s been some chatter at our company about it. Who do you insure? The manufacturer? How do you do the rating?
It’s not that these questions are especially difficult to answer. It’s more that they’re just weird, because things have been the way they are for a long time. It will be quite an adjustment.
I think there will be a fair amount of resistance to this, even if the technology is ready (and I rather doubt it’s really read yet).
I think the thing that might ultimately tip the scales might be the insurance side. If self-driving cars are proven to be safer (as in they’ve been used for years and the loss data clearly shows them being superior), there’s going to be a bunch of discounts offered for having such a car. Possibly massive discounts. That will be a big incentive, even if you enjoy driving.
@Doug Mataconis: Definitely. A lot of the plusses of taking public transit for commuting revolve around reclaiming time – the ability to nap, read a book, meditate, etc. On its face, the robot car would combine many of the plusses of transit with the autonomy of driving one’s own vehicle.
@Rob in CT: I’d think robo-cars would hasten the move to no-fault insurance. People would just insure their own vehicles against theft and damage–as well as damage to real property and injuries to those not in an automobile.
Right. I presently don’t take the bus. I actually have ridden it a few times and found it pleasant enough. I don’t take it b/c of a combo of flexibility concerns and also because my wife and I commute in together (we work in the same bldg), forming a bit of a family carpool. And our car gets solid gas mileage. And I have a sweet, sweet parking spot…
But if you told me I could have the upside without the downside? Well hell. Of course.
We’re obviously leaving cost out of this…
When I put on my programmer’s hat, I think that successes are more specialized than the general public understands. Most driverless cars rely on extreme mapping, with road paths defined down to the inch. It works for demos, when such a map can be prepared and validated, but it would be much different … unworkable IMO in a national system, with constant road changes.
I’m also skeptical of the economic benefits. Do we really need to eliminate millions of jobs for high school graduates? And to replace them with capital equipment?
It might be slightly dystopic.
@Rob in CT:
Good guy bus driver stopped and let me pass today on my bike. Here’s hoping good guy computer is as sharp.
Maybe bicycles will get transponders.
Thinking some more about this – some of the downside remains. Parking. You still have to park the thing. Or, if you want to avoid that by ordering it home, you’re wasting fuel. Either way, that’s one downside you still avoid with public transit.
When push button elevators first came out, people were wringing their hands about how they couldn’t possibly trust an elevator controlled by a computer instead of a professional human operator. Now we think nothing of it, and indeed, would consider a human operated elevator do be bizarre.
I suspect self-driving cars will be a similar process. Those of us who remember driving ourselves will never fully trust them. People born after they came out and who see them as normal will rightly consider us all paranoid nuts.
Which is why I don’t think the tech is really near ready. It’s apparently workable in controlled environments, which is nice proof of concept. But dealing with
self propelled obstructionspedestrians, construction work, tight/chaotic urban environments, etc… well, that’s gonna take more time.
I AM NOT A CRANK.
Gerdes spoke specifically about the problem with recognizing pedestrians. You can teach a car to recognize something with two arms and two legs as something to avoid, but “Go to [San Francisco’s] Castro for Halloween, and the pedestrian system needs to recognize strange cases.” Costumes or not, Gerdes believes that by teaching a car to operate at the level of our most skilled drivers, we’re better equipping them to take care of us.
This is an odd explanation of the problem. You don’t teach the car to avoid things with two arms and legs, you teach it to avoid hitting physical objects of any kind. Perhaps he is talking about programming priority, so that if the car has an option between hitting a person and hitting a wall, it hits the wall, but it doesn’t seem like it. Do you teach it to avoid moving objects? How does the car know the difference between a curb and a person collapsed on the side of the road?
This kind of technology has a long way to go. Their only comparison between human and robot car was speed on a closed track, which tells us nothing about how it would perform in a real environment when none of us are driving at top speed and there are obstacles everywhere, as you note.
This could work in an environment like the one Steven Spielberg imagined in Minority Report, where the roads were uniform and all of the cars were self-driven. In a way, they were more like personalize trackless trains than cars.
Teaching robots to drive on our existing road system, with all of its variety and chaos, is an enormous undertaking. I think we’ll see flying cars before we see that happen.
It won’t be long before we get here… Red Barchetta
Actually, the section of the 15 that’s was set up for testing driverless cars down here had small magnets (or something else that was easily tracked) embedded into the road to define the lanes. That combined with off the shelf GPS was the technology they were testing at that time. I’m not sure what this current implementation uses.
Of course, the robo-car will work great until it doesn’t. Either an internal failure due to real world, not prototype maintenance, real world dynamic road/driving conditions, hacking, etc. Not they won’t have their place. But often, the best drivers drive by foresight, e.g., kids off to the side a head kicking a ball, you slow in anticipation of the chase into the street, sound of a siren as you approach a light, someone texting approaching a crossing. That’s a wider view than the cars currently have. That doesn’t mean they won’t have their place, in interstate driving, especially out of town, for example.
The combo of the computer needing a human able to take over in extremis and the tendency of the passive driver/passenger to zone out, will create a problem in dynamic environments.
Again, the question isn’t whether the robo cars are perfect, but whether they’re better than people. Human drivers get into many accidents as a result of unexpected mechanical failures as well (e.g. rolling the car because they try to slam on the brakes after a tire blow out at highway speeds).
More important than having a car drive itself is to have a 360-degree vaporizing device, for when those clowns on their bikes decide to go all “critical mass” and to shut down traffic. Also advisable would be an automatic Gatling gun that activates itself whenever meter maids approach. But the Holy Grail, of course, would be a self-driving car that serves alcohol and hor d’oeuvres and doubles on the interior as a hot tub and spa.
Modern airliners can essentially take off, fly, and then land using fully automated instruments on board the airplane. So why don’t we start using them fulltime and eliminate the cockpit crew? Are you willing to bet your life on a pilotless airplane?
The buzz I have been hearing is that the auto-pilot technology is moving forward very rapidly, and that solving the legal and regulatory issues is probably a bigger roadblock to implementation than the technical challenges.
Why did my comment go into moderation?
We’ll probably get there if it’s economically sensible, but the current status is far from close to real drivers. The article overstates how well the robo-racer did. It was around 30 seconds per lap slower than the best human racers in similar classes over a 2:30 lap. That’s street pace, not race pace. Getting lapped every five times around isn’t a “few seconds off”, it’s a lifetime. Still, it’s an impressive effort as it is, and doesn’t need the hype.
Well, I can’t wait till the Top Gear guys get a chance to do a real race with the robocar up against The Stig.
Again, there’s a limitation to what the marketplace will accept, regardless of whether it’s rational or not. Most flyers are convinced they need a human pilot to be safe, whether or not that’s really true. So the airlines provide them the illusion of human control even if they’re really not doing much anymore.
Obviously the self-filtering comment system can’t be trusted, and OTB needs to hire a human moderator to read all comments and make decisions about what is or isn’t spam.
1. There are instances where you as the driver will need to disobey traffic rules, for instance someone is hurt and need to get to the hospital as soon as possible. How do you get the self-driving car to disobey traffic rules?
2. Moral decisions. If the self-driving car ends up in a situation where it has to either hit a child or an adult, will it be able to chose? Or a situation where it either has to hit a number of children or end up totaling the car, killing the driver?
@Rob in CT: Interesting. I would think personal injury liability would flow to the manufacturer and thus over-time, some form of products liability system. The only residual liability to the driver I can think of would be a failure to maintain the system or the vehicle. Perhaps over time states would drop the requirement to have proof of insurance and instead require quarterly to annual vehicle inspections. Compliance with scheduled maintenance requirements, particularly if set by government, could further reduce individual liability exposure. If the risks covered by an individual auto policy become too small, they might end up being wrapped into homeowner/ renter’ insurance coverage.
That would leave manufacturers with the liability exposure, and with the constant threat of mass torts and dependence on things like GPS technology that they don’t entirely control. That may be enough to prevent this technology from going forward even if it is better than human drivers. Supporters would probably need to go to the government to seek a compromise that caps manufacturer exposure in the return for greater ease in getting paid.
I’m not on the underwriting side, so I’m just passing on some idle chatter (via my wife, who is IT, not underwriting). So I’m basically playing telephone here…
@Rob in CT: Right. I am a hybrid commuter at the moment. I drive to a Park and Ride and then take an express commuter bus that goes from the Park and Ride directly to downtown Austin, where it makes drop-off only stops around UT Austin, the State Capitol Complex, and the lower half of downtown. In the afternoon, it does the same route in reverse, doing pick-up only through downtown. The Park and Ride is in the middle of an area with a lot of retail, big box stores, apartment buildings, and restaurants.
It is a great hybrid if you’re a day student at UT or work in an office building somewhere downtown. But that’s still a limited population – my guess is that this particular express route takes about 350 – 450 passengers in the morning and the roughly the same number in the afternoon. Obviously there are a number of other Park and Ride facilities throughout the service region running the same type of service from the suburbs into town.
Do you have any evidence that human drivers actually process these decisions correctly either?
@Rob in CT: I’m just playing futurist.
@PJ: Good questions. I’m reminded that I was required to learn how to drive stick when I got my first driver’s permit, because you never know when the zombies will attack. But I only ever learned to fake driving a stick while the teacher yelled at me, and have never been in a position to drive one since. So flipping the tech off in emergencies for those types of scenarios is not entirely satisfying if the driver has lost the experience to execute.
* * *
We should think of requiring people of a certain age to use this technology.
@mantis: “You don’t teach the car to avoid things with two arms and legs, you teach it to avoid hitting physical objects of any kind.”
Your robot car is driving down the road when a deer jumps in front of it. Swerving to avoid the deer might actually be the more dangerous option to the car and its passenger. So, you plow into the deer rather than swerve off the road and risking a worse accident.
Now, what if it’s a kid and not a deer? How does Robot Car know that not killing the child is worth the greater risk?
I think what we’ll see first is self driving trucks operating on highways in lightly populated regions with humans taking over in more crowded areas or for delivery to destinations off the highway.
@Stormy Dragon: “Do you have any evidence that human drivers actually process these decisions correctly either?”
Life is full of people making those kind of split-second decisions, correctly and not, so we know that human beings CAN do such things. The question you’re avoiding is HOW do you give computers the same capacity. Just telling them to avoid all objects isn’t enough.
I’m betting if you actually studied the ability of general people to make such split section decisions you’d find out that in all but a few cases they either 1) fail to notice the dilemma at all (e.g. they chose to run into the adult rather than the child because they only saw the child and didn’t realize they were swerving into someone else) or 2) lock up and fail to make any decision at all, with the ultimate outcome being determined by inertia.
People with extensive training will likely fair better, but most drivers aren’t all that well trained.
It would be more fair (and more fun) to put it up against one of the celebrity drivers. It might actually beat Michael Gambon!
Note that railway locomotives, a technology much better suited for automation than cars and trucks on the open road, are still driven by human beings. The scenario that you’re envisioning, James, may take much, much longer to materialize than you think.
It should also be pointed out that a lot of the resistance to driverless cars is likely going to end up being a result of the Dunning–Kruger effect. Much like PJs portrayal of human drivers as philsopher kings, skillfully weaving their vehicles through accidents while weighing the offsetting utilities of each possible victim, most people assume the human response to accidents is far more rational than it actually is. In 99% cases, the response to a crisis behind the wheel is going to be “apply brakes while keeping the vehicle moving in a generally straight line”.
Increasingly the human driver is only there as a backup and isn’t actually doing anything. And some systems have begun eliminating drivers entirely:
I think you guys are missing the real advantage here. I drive one kid to one city and another to a different city for school, and their schools like to play shuffle-the-schedules, so sometimes the pick-ups are separate driving events. I would be able to send my car to pick up my kids. I would also be able to “drive” drunk. In fact, I’d be able to send my car out to pick up beer. Come on, that’s worth a few pedestrians.
Washington’s Metro had automated robodriven trains until a faulty sensor in the rails failed and killed a bunch of people in a crash in 2009. It’s been human driven ever since.
I’m not sure anyone who hasn’t driven or ridden on a racetrack understands just how big a difference there is between the street and the track. I’ve done several track days on a motorcycle, and the track is far safer than riding on the road. The track surface is uniform and does not change from lap to lap, while road pavement is rough and there could be a pothole or gravel around any corner. There is no oncoming traffic and you don’t ever look behind you, so you’re only concerned with going through the next corner. Going around a track is also repeating the same set of corners, not driving to a completely new destination through road construction.
I’m not saying it can’t be done, but it’ll still take some work.
I saw what you did there…
That’s why I wrote “railway locomotive” Most of the vehicles in the article you’ve sited are trams. Trams are as much better suited to pilotless operation than locomotives as locomotives are than highway-borne vehicles.
Which demonstrates how irrational we are about such things. The Metro has had numerous accidents involving trains with human drivers:
Yet a single accident involving a driverless train is considered enough to scrap the system.
The June 2009 crash killed 9 people, the highest toll in any Metro accident. It was the post accident investigation that turned up all kinds of problems with equipment and maintenance.
Yes, but the deaths were mostly due to the fact one of the cars telescoped during the crash, a problem that had been previously been identified during similar accidents in 1996 and 1982; both times the system refused to heed NTSB recommendations that the cars should be reenforced or replaced to prevent this.
That is, the high fatality rate had nothing to do with the driverless system, which was just used as a scapegoat.
I assure you that a train traveling at full speed of about 50-60 mph crashing into a stationary train would still kill many passengers. The issue there wasn’t so much the telescoping but the faulty equipment in the automated system.
Well, the NTSB disagrees with you:
The failure of the automated train system caused the action, but it was made far worse by the Washington Metro’s continued reliance on outdated rolling stock.
From elsehwere in the report:
In other words, no one died ANYWHERE on either train except for the one car that got telescoped.
OK, there was just a long discussion of this topic in the Technical Forum at autosport.com. And I have what I consider to be a little bit of expertise on the matter. I can’t and won’t comment on the legal issues, but to cover humans vs. robots real quickly:
Long story short, robots are going to get rid of a large percentage of accidents which are caused by simple inattention, more than enough to make up for their current deficiencies. Some of you may not know this, but some cars on the road today *already* have a simulation of the car running on an onboard computer, for the purposes of stability control and steering response through the power steering system. They are simple for now, but it won’t be long before they can almost instantaneously test various possible avoidance maneuvers in order to reduce or eliminate an impending collision. They may make decisions such as it’s better to rear-end another car than to swerve out and hit a pedestrian.
What are humans better at? Mostly unusual situations that require complex and/or intuitive knowledge of how other the world works and particularly how other humans behave. Say you’re going around a mountain bend and there are some bystanders at the side of the road staring at something you cannot see yet. A human will detect this is a situation to slow down for. (But even so, there is plenty of research going on into reading body language; someday the robots will either be able to do that or have such complex sensors that they know what’s around the bend already, even if it’s a dead cow on a patch of ice in the middle of the road.)
The car that telescoped was the first car of the train that was out of control, which crashed into the stationary train ahead at full speed. I should think that if that car didn’t telescope, it would have folded into an accordion and the resulting deaths would have been the same.
Why? The last car on the stationary train (which was 5000 series car) didn’t.
Don’t tell the Libertarians. They couldn’t handle calorie counts or new light bulbs. Just wait till the nanny state says they can’t drive a car….
I have a tendency to do the driving in my family. It got to the point where my wife and I had to make a concious effort to make sure she drove at least once or twice a week, or she got noticeably rusty.
And man, my 86-yr old father’s driving has deteriorated to the point where we may need to have The Conversation and arrange for others to drive for him. Which is not going to go well.
@john personna: “When I put on my programmer’s hat, I think that successes are more specialized than the general public understands. Most driverless cars rely on extreme mapping, with road paths defined down to the inch. ”
IIRC, the DARPA challenge involved cross-country driving.
And I imagine that any Pentagon-funded research is both (a) huge and (b) does not rely on extreme mapping.
@Gromitt Gunn: ” But that’s still a limited population – my guess is that this particular express route takes about 350 – 450 passengers in the morning and the roughly the same number in the afternoon. ”
Please note that that’s a lot of parking spaces freed up in downtown.
@Stormy Dragon: “Do you have any evidence that human drivers actually process these decisions correctly either? ”
I’d expect most situations like that for a human driver to be no more ‘intelligent’ than a coin flip.
@Stormy Dragon: ” In 99% cases, the response to a crisis behind the wheel is going to be “apply brakes while keeping the vehicle moving in a generally straight line”. ”
I would bet on half of that actually being ‘hit person, and only then realize that they were there’.