Strategies like Autopilot are the worst of both worlds: They don't free up the driver and they are still susceptable to faults. Driving aids which engage to prevent an accident are a useful way to improve safety, driving aids which mean the driver is not in control and is just monitoring the system in case something goes wrong increases the likelihood of accident significantly.

> just monitoring the system in case something goes wrong increases the likelihood of accident significantly.

That depends on the rate of system failure. I used an early form of autopilot and you quickly started to pay attention to where the system would fail. It was like cruse control where you look for situations where you can turn it on but never stop paying attention.

At the other end with really safe systems they are going to default to safer than a human driver. So, there might be a significantly deadlier middle ground, but no company has deployed anything like that yet.

> I used an early form of autopilot and you quickly started to pay attention to where the system would fail. It was like cruse control where you look for situations where you can turn it on but never stop paying attention.

You may do this, but I bet the majority of people do not. I know that Google has data about it's self-driving cars that they know their "safety drivers" simply don't pay attention. In Uber's case - here in the Phoenix area last year - this inattention helped to cause an arguably avoidable fatal accident. These were people supposedly trained not to do this - yet they did.

As far as cruise control is concerned - there've been more than a few accidents of people thinking it did more than it does, and not pay attention to the road in front of them. Tesla's autopilot system, while better than simple cruise control, has - because of this better system (and poor marketing) - caused people to think it does more than it can, and has led them to sleeping in their vehicles while it "drove", to being completely inattentive, etc.

It is pretty well known at this point that those "middle levels" of self-driving technologies are things that shouldn't be implemented, because people either misunderstand them, or misuse them, leading to often fatal accidents.

When I mean early I am talking a system that stops working every 60 seconds or so. The constant failure instilled a fair amount of paranoia, and did not give enough time to cause boredom.

Google’s cars are in that middle ground, but Google has also not released their system. Presumably they are going to wait until it’s as safe as the average driver before releasing it.

I've been wondering: why don't the self-driving car companies build a system that requires a person to drive it but assures the human can't crash the car? In other words, if the self-driving models predict that the human inputs will result in a crash it forcefully overrides the throttle, brakes, or steering wheel.

The sensation would be like driving one of those tracked cars at an amusement park. Try and slam the wheel hard to the right to steer into a guard rail? Computer would physically block it from turning past the point that you would leave your own lane.

This system would:

  * force human drivers to drive the car 100% of the time
  * allow real world testing and enhancement of self driving capabilities (data gathering, model backtesting, etc.)
  * be strictly better than both fully automated and existing non-automated systems

We have the beginnings of this with the new auto-stop capabilities that are showing up in new cars. Why not build this hybrid system as a stopgap until full level 5 automation is working?

Because now you have the worst of both worlds. The computer still has the final say in what the car does, but now you have a fallible human in the loop at all times as well. If the computer really can predict crashes better than the human, then have the computer drive the car. If it can't, then letting it override the human isn't a good idea.

Example: you slam the wheel to the right to steer into a guard rail. The computer overrides you and you remain in your lane. You flatten the toddler you were trying to avoid, that the computer didn't see.

> requires a person to drive it but assures the human can't crash the car?

Because the 'not crashing' part is the hard part, the rest is just navigation.

Uber knows the routes ahead of time. They can decide if the route and time of day is safe enough for the limited self driving technology they employ or send a human driver if their SDCs can't handle it.

Unfortunately too many people subscribe the idea of startups bringing disruptive changes into car transportation because other startups could do the same with entirely different problems. Just because Elon Musk created Paypal it does not mean that he will be able to jump the hoops to get self driving cars on the roads by end of this year. Many people who are intimately familiar with self driving car technologies predict that we need at least 10 years and several innovations before we can seriously consider this as an option.

Of course you're correct that they might not pull it off, but it's not like people are basing any confidence on Musk's Paypal background. Tesla already brought disruptive change to car transportation by making electric cars popular. And Musk brought a third big change by commercializing reusable rockets.

I wonder what the model of survival is for Uber's regional competition like Didi and Ola. I can't speak for China, but driving in India is way way difficult than in the US.