It's not entirely irrational - One driver suddenly goes crazy -> One car crash. Broken software update for thousands of self driving cars -> hundreds of crashes.

That can already happen. Modern cars have fewer mechanical linkages than you might think.

There are solutions to your proposed problem. You could require the car be taken into a shop for updates. This would make the update gradual (so the eggs are never all in one basket) and it would let mechanics ensure the update was applied correctly. You could also design the software to have a core that is never updated. It would monitor basic things like "Am I going too fast?" "Can I brake in time for any obstacles?" etc. Google uses this strategy IIRC.

>Modern cars have fewer mechanical linkages than you might think.

Even better, the more software and chips there are in your car the "better" the bugs get. Friends of mine had fun problems with a new Audi A5: about 10 times, it randomly turned off while driving taking about 20 seconds to come back, and handling a car with power-steering with the engine off is very hard! AFAIK, that problem hasn't been solved yet.

I've googled around a bit and here's someone with an 2000 Audi TT that just randomly turns off because the engine speed sensor broke: http://www.audiforums.com/forum/audi-tt-7/help-2000-tt-rando...

At least in Germany though, steering and braking have to be mechanically connected which maybe limits innovation in that area a bit (or reduces incidents where you cannot control your car anymore).

People still take the plane, and most commercial planes are almost entirely flown by auto-pilots.

Making self-driving cars safe isn't an unsolvable problem: we do this already for planes.

Yes, but auto-pilots are easier problems to solve, due to predictability. It boils down to maintaining speed and bearing, while considering some other factors such as weather, terrain and collision prevention with other planes (transponder based?). In case of a truly unpredictable event, the pilot takes over.

Cars on the other hand need to do much more than that. The process of automatically determining where the road actually is, is not trivial by itself. Cars also need to handle other unpredictable factors to a much higher degree; other manually driven cars, bicycles, pedestrians to name a few. Not to mention road obstacles, and animals (in my country moose collisions are common, and deadly).

> Not to mention road obstacles, and animals (in my country moose collisions are common, and deadly).

It always fascinates me that this argument comes up from the negative side of self-driving cars pretty much every time, but it's actually the poster-child reason self-driving cars are better.

A self-driving car is at least as likely to recognize a potential collision with a moose as a human is (given current sensor tech) and probably much more so since it can't be distracted.

The advantage of a full omnidirectional sensor array is that it can pay attention to the kid on the bike, the moose crossing the street, the 5 cars around you, the alien spaceship in the sky, and the kid using your bumper to accelerate his skateboard all at once (or at least in timeslices small enough to count as such). You can only pay attention to a couple of those things.

No, don't get me wrong. I am all for better sensors, drivers assist systems, or self-driving cars. I'm just pointing out that self-driving cars are probably more complex than air plane auto-pilots.

Although, I have no real knowledge of either.

A pilot begs to differ: http://freakonomics.com/2011/12/12/cockpit-confidential-the-...

Note that most people don't know that the plane is flown by the auto-pilot.

They will, however, notice that their car doesn't have a steering wheel any more.

"A virus has been detected in your driving software. Please upgrade now to remove this threat" ... and the consequences are potentially a lot more damaging than PCs.

When the target gets big enough, it will happen. And how to protect against this, when we don't really understand the weaknesses yet - a rogue radar/lidar imitator sending false echoes, strong EM fields messing with electronic components - it could even be an intermittent connection fault in a communications bus. Plenty of unknown unknowns in this field.

given that many cars manufactured today are drive-by-wire, isn't the exact same argument applicable today?

What makes you think that updating your driving software will be like installing iOS 7?

Nitpicking: I should point out that far fewer than 1.2 million -- more like 35,000 these days [1] -- die in traffic accidents, although there are many injuries.

[1] http://en.wikipedia.org/wiki/List_of_motor_vehicle_deaths_in...

The 1.2 million deaths per year figure was for the entire world.

I think he's talking world-wide, not just in the U.S.

I wouldn't say that it is irrational. It's about wanting to be in control of your own fate as much as you can be.

I understand where you are coming from but when you get on the road you put your fate in not only your own hands but the hands of the drivers around you as well. You have to remember that while you might be an amazing driver with a perfect record there are plenty of other drivers out there that aren't and when it comes down to it they are a bigger threat to you than you are to yourself (provided you are not under the influences or otherwise incapacitated).

Whatever accidents happen on a car driven by human, it is usually the fault of the driver. For self-driving cars, then it becomes that the manufacturers are responsible for all accidents. This is unmanageable liability.

You know what will be an unmanageable liability? When the insurance adjusters start taking into account the fact that you could be reducing your risk of taking their money by switching to a self-driving car, not to mention the fact that if you get into an accident with a self-driving car and you were in control the liability will be all on you (and them).

Once someone starts selling these things, insurance for city drivers who insist on driving themselves will skyrocket.

Make insurance part of the purchase price. Individual death is covered by insurance all the time.

But yeah, testing is going to need to be very rigorous to make sure there are no regressions.

Where's your data? According to Google, their cars have gone 300,000 miles without an accident under autonomous control. At this rate, we wouldn't need auto insurance... so probably this liability cost will simply be part of the car's purchase price or a new type of insurance product offered to those who buy these cars.

On well-maintained roads that the software has been specifically trained on, with very expensive sensors and equipment, with a human driver that can take over instantly. Not that the work isn't impressive, but this 300k-no-accidents claim is just marketing spin and doesn't tell us anything about how close the system is to being ready for use.

It is more like the manufacturer is suddenly in the role of a public transport company. Self driving cars are more like small buses than personal cars.