This is fantastic! The adoption could be very fast, e.g. if insurance companies give benefits if you drive a self-driving car.
One current roadblock is the price of the system: the LIDAR (the thing on the top of the car in the picture) retails for ~$75K currently. There should be significant volume to drive the price down. But a lot of people would buy them for prestige, too (e.g. many early Prius adopters), so if the cost of the system can be reduced to perhaps $4K-$5K levels people will seriously think about this.
Not entirely. There's three main parts: liability, comprehensive, and collision.
Liability is the part that's legally required in the US. If the numbers showed that self-driving cars were much less likely to cause an accident, it would be considerably cheaper for them (a much more dramatic version of how anti-lock brakes can make your liability coverage cheaper).
Comprehensive (mainly stuff that happens while you're not in the vehicle) would remain unchanged. But it's also the cheapest part.
Collision is what covers repairs to your own vehicle if you're at fault or if someone at fault isn't insured themselves (you can also skip the first part and just get uninsured motorist coverage). The uninsured at-fault driver case doesn't change until there are enough self-driving cars on the road that accidents as a whole go way down, but the first case (damage to your own vehicle that was your own fault) should be much less likely and thus cheaper to insure.
Insurance covers much more than drivers running into other things
- theft of vehicle, parts of vehicle, or damage to vehicle and theft of contents
- fire + natural disasters (flood etc)
- damage to the vehicle by other drivers, both moving and parked
- personal liability for damage while using the car
- damage to the car from accidents
Note that in multiple vehicle accidents sometimes 'blame' is shared out between insurance companies as a way of sharing costs.
As to the fiction that accidents will never happen with autonomous cars, this is simply false. Of course there will still be accidents. The accident rate should decrease but this will take a long, long time to happen.
Take this example : workers have a self-driving pickup, but don't secure a ladder on the roof properly. While driving along the freeway, the ladder comes off and smashes into self-driving car behind. Neither self-driving car can prevent this. Same goes for self-driving cars failing to detect black ice on the road and skidding into railings/trees/other traffic.
Liability for accidents with self-drive vehicles will fall on the owner of the vehicle, possibly with insurers pooling 'fault' when one or more self-driving vehicles are involved.
"PriusBot has detected you are not authorised to use this vehicle. In accordance with law XYZ-00A, you are now being driven to the nearest police station as a precaution."
PriusBot would be unable to do much about being loaded on a tilt-tray, taken to a chop-shop and relieved of the various sensors, air bags, catalytic converters and body panels.
"PriusBot has detected you have been classified as a public menace. In accordance with statute XYZ-00N, you are now being driven off the nearest bridge."
This could actually be a serious concern. By riding in a self-driving car, we're putting our life directly in to a machine's hands. Of course, this isn't very new: pacemakers, for example, are controlled by computers. They can be reprogrammed, but only when you're physically near the target [1].
But a self-driving car would almost inevitably be connected to a network: an central computer could algorithmically control where all the cars are, where they're moving, how fast they're doing it, and basically make accidents nonexistent; such a system would be enormously beneficial to society at large. Yet, the downside is that all the cars are connected to a network, which an attacker could use to, as you say, send the target "off the nearest bridge".
Our investigation shows that an implantable cardioverter
defibrillator (1) is potentially susceptible to malicious attacks that violate the privacy of patient information and medical telemetry, and (2) may experience malicious alteration to the integrity of information or state, including patient data and therapy settings for when and how shocks are administered. Moreover, standard approaches for security and access control may not always be suitable for IMDs due to tensions between security (e.g., access for pre-authorized parties only) and safety (e.g., access for previously unauthorized parties in emergency circumstances)
Nsa will have e911 equivalent and warrantable backdoor to anything DMV/govt approved, though one would think.
[1] Some physical design limits were cpu and power, plus other operational concerns.
Since nothing about that is totalitarian at all, I don't think the term Orwellian is correct. Having computers recognize us and authorize or authenticate is solely at the whim of the consumer. It becomes Orwellian when the government demands that government controlled sensors be included in every car.
The US has alterted a few folks in the middle east they they were [Unauthorized] to [drive their vehicle] any longer, because they were on bad terms with the NSA. Try adding [Predator Drone] and [Targeted by Hellfire Missle] into the same sentence.
You can figure it out when you look at the cost of your insurance.
It depends on what you pick for coverage and what car you are driving, but for me roughly 50% of my bill goes to liability and 50% goes to comprehensive coverage.
In a world with self driving cars, your liability wouldn't be 0 but it would be lower. The 50% that goes to comprehensive coverage would remain. That covers theft, falling trees, flooding, fires, etc.
That 50% will not go completely away. The total cost of liability insurance will only drop insofar as driving gets safer. You may not be responsible for hitting that other car, but someone will have to pay.
Because of that, you likely wil have to pay a premium on top of the 'real' costs of such a car to the seller of the electronics, either as a lump sum, or as a per year and/or kilometer contribution.
Yes, it should be. However do you think that insurance companies are going to allow that? They'll form the insurance-equivalent of the RIAA before that happens.
I highly doubt that. Insurance companies don't represent a "precious resource" (aka celebrities), so it would be easy for somebody to come in underneath them.
Besides, insurance companies make money on the difference between rates and claims. If claims fall significantly, rates can, too, yet the insurance companies can still take home the same sized paycheck.
>insurance companies make money on the difference between
rates and claims.
This isn't true, but is a common misconception.
Insurance companies make money by earning an investment income on the retained premium (the premium earnt between when a policy is paid for, and a claim is paid out). You can easily see this by reading analyst reports on publicly-listed insurance companies, which will analyse in detail the return on the retained premium.
An insurance company collecting more premium than claim costs is overpricing itself. An insurance company paying out more claims than premium is underpricing itself.
The idea is that the risk is managed through actuarial study, so the premium/claim payout is managed, giving an optimal time lag between collected premium and paid out claims.
The ability of insurance companies to create a very large pool of investment funds is why Warren Buffet buys them. He is an expert at investing, and the premium pool gives him the size he needs.
I agree that it will market forces will drastically lower the price, but if the margins stay the same the insurance companies will still make less money as the overall premiums will be smaller.
With self driving cars, why would you own one? Why bother making a capital investment and paying to park a rarely used vehicle when Übercab could provide rides everywhere at much lower cost with much greater availability?
If you had a human driven car, would a malfunction in the car causing a crash be your fault? [1] I can see that if you hadn't updated the firmware of the car then maybe it could be claimed that you had not performed adequate maintenance?
Vehicle integrity is the responsibility of the driver for human driven cars, I assume it would be the same for autonomous vehicles. The problem comes when you are driving a 15 year old car because you can afford new. For non autonomous vehicles mechanical safety is dependent on powertrain, steering, and brakes, all of which are under manual control. Maintaining the integrity of autonomous system--its computer, its sensors, its radar/gps, the accuracy of its maps, will be even more important than the maintenance of those manual system.
I would object. If you want to use the HOV lane, have multiple people in your car. I understand the rationale for extending it to things that are sufficiently fuel efficient. I don't see it here.
HOV lanes promote multiple people to reduce congestion, and self-driving cars are touted as a way to reduce congestion. If the latter claim is true, then HOV lanes could reduce congestion by allowing self driving cars to drive in them.
I actually think they will do the reverse. A self-driving car will likely do more miles. What's more, it will do empty-passenger miles. The travelling patterns of people are unlikely to change. But there remains the distinct possibility that people will get their cars to drive to work, then home (or at least to remote parking) again if 'work' does not have parking.
Thus the potential for cars without people driving around. This is inherently worse for congestion than 1-person cars.
Ah, right - I guess I'm just so skeptical of the claims that didn't think of it. I expect self-driving cars to be driving like a cautious human driver for at least 5 years. If we get to the point where they're in fact driving closer together, then there is indeed an argument for letting them in the HOV lanes.
I think the LIDAR approach will have to go. Laser systems are inherently susceptible to jamming with a cheap laser diode. We cannot afford to have mass adoption of cars where somebody can create weeks of citywide gridlock with a few thousand dollars of lasers.
Thrun said in a talk I saw posted on HN earlier today that this wasn't being considered a problem. I'm not certain if targeted LIDAR jamming was on his mind, his comment was specifically in reference to operating with many other autonomous cars on the road.
He was concerned about GPS jamming though, but thankfully these cars have been designed from the ground up to not rely on GPS (as a signal from a satellite is not reliable). I believe as long as the area has been 3D mapped by Google, the car is capable of driving and navigating without GPS assistance.
You assume that the car is occupied by a sober adult who can drive. Not something you'll be able to assume any more when mom can send the car to do the school run.
You cannot drive your car in 1 meter spaced convoys. In many cities jamming the navigators would cause the cars to not fit on the roads, resulting in gridlock for however long the jammers could be kept going.
And dumping 500 caltrops off an overpass would likely have the exact same effect. Worrying about someone actively attacking this kind of system using theoretical jammers that could seize up multiple lanes of traffic (without easy line of sight!) and operate for hundreds of feet at this point seems premature.
You can't do the caltrop trick at 50 locations for weeks on end.
The thing about lasers is they use a very narrow slice of wavelengths. Environmental light like sunlight has very little power within that slice, so its easy for the LIDAR to see its own laser. But that makes it really easy to jam. You could probably build a jammer today for a few hundred dollars, which means that in the future every disaffected monkeywrencher will do it. Ergo the approach is not suitable for the mass market.
The eye is broadband and the desired light is bright, so it takes a very high power density to jam it. LIDAR receivers are very narrowband and the desired light is fairly dim, so jamming it will be quite easy. Modulation does not help when the light detector is saturated.
To keep the detector saturated for any period of time you would need to have aim that's a little too perfect. Besides, my central point was that there are lots of vulnerabilities in existing cars and they're still driven.
In case you decide to test your theory about lasers and human eyes, please don't use someone else as a test subject.
I think it can be done without aiming, assuming the light detector uses a wide angle lens. You just have to illuminate the scene observed by the LIDAR with a few watts per square meter of laser light. That's doable today, and easy in the future.
One current roadblock is the price of the system: the LIDAR (the thing on the top of the car in the picture) retails for ~$75K currently. There should be significant volume to drive the price down. But a lot of people would buy them for prestige, too (e.g. many early Prius adopters), so if the cost of the system can be reduced to perhaps $4K-$5K levels people will seriously think about this.