What I’m interested in is whether they counted hypothetical accidents. As far as I know, there always has to be someone in the driver’s seat, ready to step in at any time. How often did that person have to step in? Now that would be an interesting metric, especially when plotted over time.

As is, it should be pretty unlikely for the autonomous cars to cause accidents, even if they were much less reliable than human drivers (since there is always a human ready to step in).

I think autonomous cars are awesome and very clearly the future, but I find it hard to believe that humans never had to step in, especially not in the beginning. I mean, that’s why they do those extensive test drives, right? To find bugs that they can squash.

Computer driven cars also have the option of just stopping when things aren't looking good (analogous to a panic). For all we know they could stop when a leaf blows across the road - unlikely to cause an accident because of the reaction time.

The "just stop" approach isn't available for planes or nuclear power plants for example.

The "just stop" approach is also problematic on a busy highway.

I don't know where you live, but where I live the highways have lots of stopping during busy times. And of course drivers always have to be aware of stopped cars as they do happen. I suspect highways are where the automated cars have the easiest time, and it is far more likely side roads and suburban areas that are more confusing. Abrupt stops are more likely there (eg kids/pets running across the road). In the US I do believe that running into the back of another car under almost all circumstances makes you at fault.

Where I live a panic stop in heavy traffic on anything called a "highway" could get you killed. I imagine a lot of the stops we see on the highway that do not result in accidents are made safer by the network effects of lots of cars doing the same thing over time and communicating visually. It's different in nature than somebody (or some machine) getting confused and stopping suddenly in the middle of the road.

Don't get me wrong: things aren't so grim for the "just stop" approach of dealing with a problem. There's no reason an automated vehicle won't be able to communicate with other cars, warn everyone of an emergency stop (at a minimum), even find the best path off the road, depending on how serious the problem is.

Not if all of the cars can do it just as easily.

I love the mental image of a whole highway full of cars slamming on their brakes to let a leaf blow by.

How would a full switch to automated cars work anyway? Would they require different roads?

If their functionality included stopping on a dime at a duck on the road, and human reaction speed isn't equivalent, it isn't safe to human drivers. But cost and manufacturing has to factor in before we can just replace every car on the road. People still drive 20 year old cars.

I expect it will be a mixture of carrot and stick like any good phased replacement. Dedicated lanes (since they can drive 2ft from the car in front) and higher speeds because they are able to cut the reaction time down so much lower.

Eventually making it mandatory for some classes of roads (motorways / highways).

At some point in the very distant future your classic car will only be good for showing off on track days or classic car meets.

Actually, it kind of is available for planes (at least the autopilot systems). See the AF447 accident and how the autopilot disconnected itself when it started getting conflicting data from the pitot.

Agree. Did humans step-in when they noted conditions which they believed the car unable to handle and perhaps prevent the majority of likely accidents?

My suspicion is that the figures quoted around these Google cars and their safety record, if you take them at face value, lead to a ludicrous level of optimism.

I think that Google have simply done a better job of creating a self-driving car than you are willing to let yourself believe.

They have done this by attacking the problem from the 'Big Data' mindset which Google are intimately familiar with. Instead of trying to crack hard AI problems (computer vision being the most obvious) they are simply counting on being able to record and use enough information both before and during the actual driving that they simply avoid these problems.

From what I have seen this approach seems to be working. My real concern is it will not be economically viable to integrate the kind of sensor arrays they are depending on into a real car (2 radars and the laser scanner being the killers).

Indeed. The whole self-driven car isn't a new idea - various individuals and university disciplines have been trying the idea for the last decade over in Europe - it's legal to test these ideas on the road in some countries over there.

If you make the assumption that all cars will be AI, then traffic-handling becomes very, very simple. It's not obstacles or accidents that cause serious problems for traffic, it's the fact that you can't know what all other drivers are thinking.

If all vehicles behaved to a certain rule set then all vehicles actions could be predicted, and as a result designing logic to accommodate that is made much easier.

That might work if the road system were entirely isolated from the rest of the world, but it's not. There are walkers, cyclists, animals, children and other random obstacles that might exist, none of which will ever be controlled by computers.

The idea of all cars following the same set of rules even if they are all AI driven is actually way too complicated. Bugs and misunderstandings of the standard would cause deviations that the other cars would need to react to anyway.

> If all vehicles behaved to a certain rule set then all vehicles actions could be predicted.

... or, potentially, hacked.

They are already putting radars into higher end luxury cars as part of their cruise control systems. Granted, google cars involve a good bit more than those systems, but I am fairly confident that the usual trend of features trickling down into normal consumer cars will continue.

I don't think that Google would say otherwise. This is just a status report in the beginning of the program.

really? 300,000 miles without an accident is not particularly amazing for a human. I mean, it's good, but not amazing.

I work for an insurance company that specializes on auto insurance. It'd be filthy rich if people only had an accident every 300,000 miles.

The company is pretty well off, but pays 95% of the premiums in claims payment.

Doing a very back-of-the-envelope calculation, 22.5% of our insured cars had a claims payment this year (august to august), so an average insured person has a traffic accident every 4 years or so, and does way less than 300,000 miles every 4 years (more like 100,000 being generous).

I was saying not that it wasn't better than average but that it wasn't amazingly good. I'm pretty sure my mom has gone 300,000 miles between accidents; (growing up, we went on a lot of road trips.)

all I was trying to say is that I think the best human drivers can pull off three hundred thousand miles; it's not inhuman. and that's what we are going to need for self-driving cars to become widely accepted; something that would seem impossibly good to a human. And I bet they will do it, eventually.

I was responding to

>My suspicion is that the figures quoted around these Google cars and their safety record, if you take them at face value, lead to a ludicrous level of optimism.

and I don't think 300,000 miles without an accident is a ludicrous number; it's something a human can achieve, (even if it's something most humans don't.) It's a milestone, sure, but like I said, for self-driving cars to be accepted, they can't just be better than the average driver; they need to be better than the best drivers.

Ah, ok :) , I hadn't understood your statement correctly.

I'm looking forward to self driving cars because of their convenience ! Imagine being able to sleep during road trips, using them as a taxicab for places where there's no parking (and ask it to park itself :) ), sending them to pick up somebody... the possibilities are endless !!

What type of accidents though? Do people get in a number of fender benders over their lives? or head on at 65mph?

What sort of accidents will GoogleCar be prone to? Fender benders, or a head on at 65mph every 500K?

I'm very hopeful about this. Most people don't want to drive most of the time. As long as I can still drive my V12 Aston martin manually when I want to, this is awesome.

I imagine, once automated cars are safe and popular enough, they will mandate that all new cars must have it and set a date after which only automated cars are allowed on the road. It will become an offence to take over and drive manually unless the circumstance demands it. You'll be fine driving your Aston Martin, but your kids, or their kids might not be.

I expect that as the automated cars get sophisticated enough that they no longer act individually, but rather each car is in constant communication with the cars in the area so that they can coordinate maneuvers and engage in global planning, a semi-manual option might be added to the system so people can drive for fun.

They way that would work is that when you car is in semi-manual mode, it tells the others cars that a human is controlling steering and speed. The other cars will then give the human extra room (e.g., no drafting or close formation driving on the freeway, no crossing at right angles in an intersection at full speed just inches apart, or that kind of stuff that the cars will do when computers are driving all of the cars in an area).

The automated system will still be monitoring the human and can take over if the human does something bad, so the other drivers don't have to worry that a human driven car will suddenly cross the line into their lane and kill them.

This could even be turned into a game, where the automated system can score the human driver based on how well the human drove and how often the system had to intervene to keep the human from crashing.

Note that racing could also be allowed in such a system. Want to take that Aston Martin out for a race against your neighbor's Ferrari? Let the system know, and it can give you a course through the city, clear a bubble around you two, and let you go at it.

Most of these things will not be possible merely because they give you an unfair advantage. What you call "racing" I call "being at my destination much sooner", and, yes please, I would like every other car to step aside so I can get home faster. Which, of course, only means that everybody will try to abuse this.

Make "race mode" cost money then. Cheap enough that an occasional race is not ridiculous, but expensive enough that it is not practical for a commute. It probably wouldn't be an incredible advantage anyway. Your car could drive much faster if it knew exactly what every car around it was planning on doing.

At least in urban areas, you're eventually going to end up having vehicle performance arbitrated via radio by a computer network outside the vehicles. (I'm pretty sure it's much easier than trying to do everything "peer to peer," if you will) This gets you intersections where nobody has to fully stop and vehicles merely coordinate and modulate their speed, which is a pretty powerful efficiency advantage in itself.

With a system like that you could easily design a market where vehicles can choose to pay a toll to get a favorable path through traffic. You might see a few options and incentives given for choosing a more fuel-efficient path, etc.

Honestly, I'd expect semi-fun mode to be implemented to prevent human drivers from passing, thus killing the "fun" and reinforcing the "driving is tedious" meme in order to serve a greater safety goal.

I don't buy it. In that scenario, how does anybody learn to drive in the first place? Someone who hasn't spent years of their life driving manually won't have developed the skills to intervene properly in an emergency. At best, driving would, become a specialty career/hobby, like piloting aircraft.

By the time all personal transportation is automatic, it won't just be the status quo with "robot chauffeurs". The whole transportation ecosystem will have to change substantially. Self-driving cars are just a first step to get people used to (or even enamored with) the idea of automatic transportation on a large scale.

I love driving. I find it incredibly enjoyable. I don't think automatically driven cars will change that because I'd happily give up 80 - 90% of my real world driving to a computer. Driving on a congested motorway (highway) or in a busy town just isn't fun, no matter how nice the car you are driving.

I think horse riding might be the best example of the way things will go if and when automatically driven cars become a reality. Once a life skill required by everyone, now a hobby in all but the most extreme of communities.

We're already heading down that route [1]. Although I can't see it ever being an offense to manually drive your car, the car will just automatically take over the driving duties if it doesn't like the way you are driving.

[1] http://www.autoweek.com/article/20120803/carnews/120809942

You're not factoring in the Media, which will raise a big fuss about taking away their rights and robots not being safe... you get the idea. I can't imagine how people will react once there's an actual possibility a car could be driving on the road with no human controlling it.

I believe that 10.8M accident figure is an estimate counting all kinds of crashes, even the ones not reported to police.

6M police-reported crashes in 2007. 1.71 million police-reported crashes with an injury. 37,000 traffic fatalities in 2007, 30,862 in 2009.

Traffic fatality data is here with only some high-level info about non-fatality crashes: http://www-fars.nhtsa.dot.gov/Main/index.aspx

A lot of those accidents happen in rain and snow, which the Google folks admit are not as well tested. Under clear conditions on open roads, humans probably do better than 300K miles too.

That doesn't seem right to me. A person driving 10K a year will hit 300,000 miles over the course of 30 years. Does the average person really have less than 1 accident every 30 years? Especially when they are younger?

Per mile is a pretty poor measure. My dad drives at least 40K/year (long commute) and has averaged one accident per 20 years - but obviously for a regular driver following a familiar route the chances of an accident are pretty low. I would expect someone who drives less would actually have more accidents/year.

Mentally reviewing the people I'm close to, the figure seems to be off by at least an order of magnitude.

I think it includes truckers who are far safer than normal drivers on average and do a lot of driving.

I put 40k on my first car and hit stuff enough to warrant replacing something - often just a fender etc - at least 5 times. Only one fender bender when I rear ended someone at ~15 mph talking on my phone like an asshole.

I know I'm a worse driver than most. But I also wouldn't say that my similarly young friends average anywhere close to 1 such error per 250k miles. Definitely closer 25k.

I don't know much about this google project. Can autonomous cars (i) drive on the highway (ii) drive at night (iii) drive in snow or rain?

Autonomous cars have driven on the highway in real-traffic conditions [1], [2], although they had a driver at the ready in the event of any problem. I don't know about night or weather conditions for these instances.

[1] - http://en.wikipedia.org/wiki/VaMP (their citations pages aren't responding at the moment)

[2] - http://www.argo.ce.unipr.it/argo/english/index.html

They talk about weather issues in this video at 3:40.

http://www.youtube.com/watch?v=QkYFIhJ3rnY

In summary: They can handle darkness and rain, but not snow.

If you're interested, Sebastian Thrun has an intro to the topic as this course (you can preview all videos and problems)

http://www.udacity.com/overview/Course/cs373/CourseRev/apr20...

Driving at night and the rain will probably be where autonomous cars really outshine humans. Lidar systems don't care if you are wearing all black during a night at new moon.

Snow is much harder for the time being. A snow drift falling off a car can make a google car think a cooler has appeared in the road.

Could it handle Lombard Street?

How would it react to a roadway that's covered in several feet of water due to flooding (and poor road design)? Or sinkholes? One-way streets? Freeways which change directions at different times of day? Dirt roads? (Can we make it autonomously offroad?) How will it react during an earthquake or tornado? How will it react to a bicyclist running a red stoplight and into the car's path, when the car has a green light? When the car comes across kids playing football/street hockey/baseball in the street, will it stop for them? How far out of the way will the kids have to move before the car thinks it's safe to go past?

Will these cars have systems to react when they take damage?

Google's algorithm (which Strum seems to be pollinating everywhere -- and good, because the old algorithms were failing DARPA tests) is to pre-map the road. The car knows exactly where it is and what is supposed to be there, which it also compared with what it sees now.

How will it react to a bicyclist running a red stoplight and into the car's path, when the car has a green light?

It will stop. In fact, the "failure mode" of self-driving cars will be how pedestrians can bully them by stepping in front of them so they stop.

When the car comes across kids playing football/street hockey/baseball in the street, will it stop for them?

There is already video of this. The radar/lidar isn't good enough to see a baseball, but it's definitely good enough to see a kid.

Will these cars have systems to react when they take damage?

It will probably pull over and stop if something is wrong.

Interesting case here for carjackers. You're Google car probably won't floor it in the middle of the night through a red light when an armed gang approaches your car. Then again I guess that's what "manual" is for.

Carjacking a car loaded with recording equipment that probably phones home with GPS and other environmental data probably isn't the best idea.

carjackers aren't known for their brilliance

Yes, a Google autonomous car drove down Lombard Street: http://googleblog.blogspot.com/2010/10/what-were-driving-at....

Here you can see video of self-driving cars handling Lombard Street, as well as dirt roads, off-road driving, pedestrians and bicyclists and other obstacles in the road, plus tons of information about how they work:

https://www.ai-class.com/course/video/videolecture/48

https://www.ai-class.com/course/video/videolecture/209

https://www.ai-class.com/course/video/videolecture/210 (Lombard Street at about 4:40)

One thing that strikes me is how much more these robots can potentially see than a human driver. Imagine how much better you could drive if you could see in all directions at once with multiple range-finding sensors and cameras, track dozens of moving objects simultaneously, never get sleepy or look away from the road at the wrong time, react to events in milliseconds, and learn from hundreds of thousands of miles of training data as your driving "experience".

Everything you've just said would have come up in numerous meetings before they even started planning how the system would work.

I guess another good question is: "how well do humans handle these things?"

I think the answer is: "not very well."

The original DARPA Grand Challenge was set in the desert on dirt roads. It's not that difficult to identify the road given some assumptions about how roads work (e.g., it's likely to be a slightly different color or texture)

I thought they were limited to driving on the highway in the empty parts of nevada?

I drove behind one yesterday as it returned to Google HQ in Mountain View. It was pretty exciting. It was also rush hour, and people didn't like that it tried so hard to give a safe amount of space in front and behind itself - it ended up being a little bubble of lane-changing space that moved slower than the rest of traffic.

Not that I hold any of this against the car - it was clearly doing the intelligent thing. But people... we are animals.

Weird. Your post didn't have a reply link. I had to click the permalink for it.

Anyway, I wanted to say that I read an article about a guy in Seattle that always kept a couple car lengths ahead of him. I seem to remember the article saying he was able to get home quicker that way than by behaving like the rest of the cars on the road.

Meta: there's a delay between when a comment is posted, and when the reply link becomes active. That delay is larger for deeper comment threads. I've never heard an official explanation, but one can imagine a few reasonable ones, e.g. to encourage people to stop and think in heated back-and-forth exchanges.

I too live in the Seattle area, and I've heard of this technique. Sadly, when implementing it, all you see is the angry drivers behind you who rush to pass and gobble up the traffic bubble you've been saving.

I thought that this was covered by the Traffic Waves guy who said his method deals with that problem quite well?

Yep: it doesn't happen, FAQ: http://trafficwaves.org/tfaq.html#five

- if truckers can maintain a bubble, how can it be impossible?

- yes, aggressively weaving drivers sometimes jump in ...then they jump out again.

- at merge zones, the entire point is to let people into your lane.

- In theory, everyone else wants your precious bubble. In practice, not.

...and in a burning building, we all know the best way to escape, eh? EVERYONE AGGRESSIVELY PUSH FORWARD, DON'T LET ANYONE ELSE GET PAST, COOPERATION AND TEAMWORK ARE FOR LOSERS! Yep, that certainly won't trigger a huge clogged mess where everyone dies.

That's the one. Thanks!

Seems like the safe distance in front of a robot car would be significantly less than for that of a car driven by a human. Of course, most humans I know think that the safe distance is 2 feet, which is probably less still.

They probably have to assume the human co-pilot may still need to assume control and avert disaster, and give them enough extra time to do that (beyond the time they'd need after starting out in control). With no need for failover, the robot should reliably react faster, though it's probably only reacting to immediate events rather than predicting drivers' (mis)behavior.

Most people stay way too close to the car in front, though. You're doing well as long as nothing unexpected happens. But if someone for some reason has to hit the brake really fast, you'll get a completely unnecessary pileup.

The Google cars probably also have additional margins of safety to allow for the driver to react to unexpected computer behavior _and_ actually step in and perform the maneuver himself.

  > Most people stay way too close to the car in front, though.

Yes, and people who don't care about this risk are very annoying drivers for the people who do.

I usually lightly tap the brake pedal, to let the lights come on, when they're too close behind me.

I've seen the Google self-driving car navigating within morning traffic in San Jose. There were two men with laptops in the driver and passenger seats, presumably monitoring the car and available to step in if something went wrong. It was also kind of surprising to see the roof-mounted device it used to map its surrounding. It was fairly large and rapidly rotating.

Did it look like this: http://velodynelidar.com/lidar/hdlproducts/hdl64e.aspx

Very impressive spec - as you'd expect for a $75,000 product.

The device on top is a laser scanner creating a cloud point map for spatial awareness.

that is for a car with no human in it, afaik. the google cars drive on normal roads with other cars present

To date, all Google autonomous cars have had two occupants - a driver (who is strictly hands-off - she's only there in case something goes wrong) and a tech who monitors the data to make sure things are proceeding as expected.

Only problem is sample size really. We need 2.9 trillion miles on Google cars with x accidents then.

If the Google Car had gotten into 100 accidents over 300 miles, we easily could claim it much worse than most drivers. However, as with most rare events...

I think the true rate of human accidents is higher than 1 per 250K miles - many are simply never reported to the police or DMV, especially because many states have a cost threshold ($750 in California, higher in other states) or require an injury to be reported

I'm guessing that more accidents happen in intersections than on the highway. My understanding is that Google's self-driving cars are currently limited to highways and known-course driving.

I don't think it's going to do a very good job driving in San Francisco city streets (but then, who does?)

From what I know it's always known-course. If you were to get a google car today for your commute (you can't), you would first drive to and from work yourself, probably more than once, so it can map everything.

Knowing where there should be a stop sign and traffic lights significantly reduces risks. The computer still needs to be able to learn when new ones appear, of course.

It would also help with some roads that are incorrectly marked on the GPS. There is a highway in ND that has a 4 mile stretch that has a miss marked road. Multiple GPS maps show the road about 100 yards west of where it actually is. Damn annoying driving 4 miles with the GPS constantly saying "Get back on the road". Thought it was just a garbage GPS map, but it happened on a couple of models.

Guess I really don't have to worry until they figure out how to get the car to drive on snow covered roads.

That's 10.8M crashes. The vast majority aren't really "accidents", since they are caused by human actions.

Census needs to fix their reports to be consistent with Department of Transportation terminology...