Let’s Talk About Self-Driving Cars

Edit: I would gladly accept more failure in uncommon edge cases if it means reduced failure in normal operation (ie: drunk guy slams in to a house at 1 AM on a clear calm night)

http://www.autoblog.com/2017/01/10/nissan-self-driving-cars-...

> For example, Sierhuis described a situation where a temporary construction site, something like a crew repairing a traffic light, would confuse the car. Not only would the traffic signal be missing for the AI to process, but we would also have a worker waving cars through the intersection. It's a prime, real-world example of a challenge that is basically impossible to program into software. With SAM, when a self-driving Nissan approaches a situation like this, it stops and sends a signal to a human operator. This operator, called a Mobility Manager, checks the camera feeds from the car, high-definition maps, and whatever else he can, and then plots a new route for the car to take. This information is then sent to all the other Nissan AI machines in the area so that they don't have a problem when they approach the same intersection.

Obviously "do not stop in the middle on the road" will be high on the priorities of self driving algorithms...

If you're going to look at your phone at a red, know the light timings or look at the transverse pedestrian crossing signal or transverse yellow light so you know how long you have.

I do think there's the potential to enable full autonomy in well-defined use cases while requiring manual control elsewhere. (Still problematic--how do you enforce?)

But the idea that a driver can take control on short notice from an automated system is a non-starter. Maybe you can assume there's an adult with supervisory control in the vehicle but you certainly can't assume they're prepared to deal with urgent emergencies.

That sounds about right. The ability to read/write/etc. while driving on the highway would actually be a huge win for a lot of people. And, equally important, it seems achievable in, perhaps, a single digit number of years.

But there would need to be an unambiguous mode change from "you can do Facebook now" to "you're driving" with plenty of time to make that shift.

This isn't as sexy as roboUber, shared cars, etc. but I'm inclined to agree that getting the last 10% of the way to unsupervised self-driving in essentially all situations is going to take a very long time.

Serious question.

In fact some vehicles already have front crash prevention systems and stability control systems that automatically brake and cut the throttle (2 control axes).

http://www.iihs.org/iihs/topics/t/automation-and-crash-avoid...

So manufacturers could build upon that baseline by adding more sensor inputs and control axes for collision avoidance. Instead of just braking when a large obstacle appears in front the vehicle could steer around obstacles, exert force on the steering wheel to prevent running off the road, limit throttle inputs when approaching curves, etc. Some aircraft autopilots already have similar functionality by automatically climbing to avoid terrain.

Obviously, a lot of people still die in traffic accidents. But the bar is actually still fairly high for automated systems. I also suspect (though I don't have the numbers) that even if automation were just for highway driving that would, by itself, significantly reduce the fatalities.

But...

It isn't just deaths that need to be prevented. It is a combination of total damages, delays, deaths etc. If self driving cars take forever to make a turn or have other AI failures like driving too slow, it won't be acceptable to a large number of people.

In addition, you will see a situation in the future where partially automated cars Level 2 or 3 are safer than current driving modes. That means that in order to go from level 3 to level 4, you need the car to be safer than level 3.

I want a fully automated self driving car so badly, but I suspect I have a long way before I can hop in a car with no steering wheel and sip an IPA while being rushed home from a tough day at the office.

You also can't compare deaths per million miles if the self driving cars don't operate in dangerous conditions. You'd have to compare apple miles to apple miles and orange miles to orange miles. A lot of people die in car wrecks in bad conditions.

A car that only kills .5 people per million miles, but doesn't work in the rain, snow, dark, whatever, likely isn't safer than a person.

Autonomous cars absolutely have the potential to be much safer than human driven cars, but their innovations don't change many other aspects of car oriented transportation. It remains questionable then whether our cities will be all that different if the cars on the road are autonomously driven or not.

In this sort of discussion I feel it must be added that it's fully possible to build walkable, bike friendly, cities right now with current technology. We just choose not to do so. We don't need new technologies to achieve this.

Driving actually doesn't suck that bad in the states, even in crowded cities like LA, but self driving cars could be transformative in a city like Beijing.

I do agree it is a problem that we share roads, and many people don't take the risks seriously. People's poor judgement, e.g. using apps on their phone while going 70mph in a 3,000+ lb vehicle is disturbing. Motorcyclists experience things like this vividly, and often daily. Maybe fully automated cars are the only answer to that though, because relying on improving everyone's judgement, or policing it, might be the least realistic solution conceived (for anything).

Of course we want to make driving safer, and in that way autonomous capabilities are a ground breaking safety feature. Other cars probably need to be safer too, since we are a very, very long way away from even a simple majority of autonomous cars. Maybe one day we'll see the "end game" of zero-ish road fatalities, but that relatively distant possibility is not a reason to essentially outlaw driving. Keep in mind autonomous cars are also very expensive, and will be for some time. If the "outlaw" part came a bit too early, it could end up looking a lot like "People too poor to buy and maintain a brand-new car are not allowed to drive on public roads".

There are plenty more reasonable steps to be taken before that, e.g. since teenager drivers account for a disproportionate number of accidents/injuries, so maybe we don't need high-school kids driving. Or, maybe we need to require extra licensing for a housewife that wants to pilot a 5,500lb Escalade at highway speeds. Maybe people should go to jail instead of paying a tiny fine if they are caught using Social Media while driving.

The list goes on, without copping to the naive, even juvenile, conclusion that outlawing human drivers is logical because autonomous cars are safer.

Specifically for motorcycles, as many experienced riders would say, sport bikes especially can hardly even be used on public roads anyways. My old bike red-lined 1st gear at ~96mph, i.e. there isn't a public road where it is legal to "use" the first of 6 gears. They're made for a track. On the other hand, RE: safety, in the overwhelming majority of cases the safety risk on a motorcycle is to the rider not the other drivers on the road. Motorcycle accidents where the rider was at fault and a vehicle occupant was killed are likely less probable than death via choking on a bagel. So a outlawing human drivers in this case sounds more like taking away an individual's choice rather than controlling any legitimate risk to public safety.

As for cars, I agree that it will probably at some point make sense to separate autonomous cars from human drivers. There is not much way around the fact that you can kill other people easily in a car, and that a great number of people would justifiably want to opt-out of that risk if they had the option. Someone's enjoyment of driving won't be nearly enough to justify otherwise that when the time comes.

I mean that's the whole selling point...and it would basically take that level 5 ability (as demonstrated in many articles and studies) to give you the "I can kick back and watch TV on my highway commute" ability.

Hell you can install an in-dash DVD player now, but that's not the vision of self driving cars.

Looks like it's hosted on Vimeo... so I don't think so, at least without hackery.

Edit: But also like that cool double/triple pendulum thing (https://www.youtube.com/watch?v=cyN-CRNrb3E) you don't necessarily have to be able to predict the future if you are simply monitoring the present state at a high enough sampling rate. Note that this thing is sampling at 1ms. Not sure what kind of sampling would be necessary for cars, certainly a function of velocity and braking power.

Example: a highway has one side closed for roadwork. It's dark and it's snowing. At the start of the construction work there is a sign that says "wait at this point for escort car". The text might be partially covered in snow. When the escort car arrives it leads the queue of waiting cars past the road construction, in the oncoming lane.

A human driver understands that the right thing to do is wait at the sign. If the sign isn't readable because of snow, the driver just gets out and brushes off the snow.

The human driver understands which car is the escort car because of its markings and behavior.

The human driver understands the reason the escort car is there is because traffic has to take turns in the same lane and there are no traffic lights. Even when a human driver encounters this for the first time, it will realize what's going on (it confused me the first time).

I'm not sure it would even be possible to teach an autonomous agent to adapt to any unknown situation where the information is lacking, incomplete or conflicting.

There's no point in discussing L5 realistically, nobody is anywhere close to pulling it off.

So it seems like they're pretty close to L4.

The question has big implications... the post mentioned the follow-up question of whether we need stoplights anymore. We don't need stoplights anymore if all existing cars disappeared all at once, and all car manufacturers go straight to level 5 at the same time. We do need stoplights for the time being if for some reason that doesn't happen.

Probably, yes.

> Yeah, sensors will be added to cars (like seat belts were) and cars will be aware of other vehicles.

High end cars have had auto steering and auto throttle for at least a couple of years. One of my friends took me in a ride two years ago in a Mercedes that keeps a safe distance between the nearest car, and steers itself on curvy roads. You can set cruise control and go from Sacramento to Reno on I-80 with little intervention.

I drove most of the new compact SUVs on the market over the weekend, and bought a 2017 Honda CRV. All of them have sensors to warn about nearby traffic. Half of them have some form of automated steering, usually called "Lane departure mitigation". The CRV will try to steer itself to stay in the lane. (It's not great, I wouldn't trust it for very long, but it is there.) It has "Active Cruise Control" where it follows the car in front with a preset time gap you can set. I let it drive me through the city yesterday, it comes to stop at a light as long as there's a car in front of me. (Which, incidentally makes it seem dangerous to use this way since it doesn't read stoplights.) The car has anti-collision automation, it will slam on the brakes and possibly steer if a crash is imminent, and it has "Vehicle stability assist" mean anti-roll and anti-slide measures that will take over brakes or steering or both.

The CRV is a mass market car, not something high end or experimental, so yeah I think a high and growing level of automation is on the road already and will not go away.

Fully driverless, I'd agree, that seems unlikely right now, and I wouldn't be surprised either if lawmakers or the public decided to make that illegal for now.

They're just not very popular.

They're (all?) electric which you sort of want in this application. But that means that they typically only mow up to 1/4 acre or so--which isn't a lot of work to do with just a standard mower. And they're expensive ($1K+).

So you have an expensive and not very good lawnmower that can only cut lawns that are generally easy to do anyway. And, besides, if you're not very price-sensitive and don't want to cut your lawn, there's no absence of people who will do it for you and do other trimming etc. at the same time. So there's just not a lot of market for the product.

Full autonomous Uber's - so many orders of magnitude that I don't even want to speculate on magnitudes.

I'm pretty sceptic so I put "mostly autonomous vehicles" within reach, but "driverless taxi" several decades away. We could make "autonomous safe regions" on the map where the mostly autonomous cars could work well. But we just can't make them work well in the normal world for humans because of the edge cases.

The last fraction of a percent of scenarios require humans or something very nearly human. You need to be able to improvise, learn, read half-hidden text on road signs and deduce from context what it means etc.

It's not acceptable to just handle 99.9 % of the situations for a driverless car, then you end up with driverless cars stranded everywhere.

I think the last step from "mostly autonomous" to "fully autonomous" is comparable to going from an ai recognizing all the objects in the famous funny Obama-scale photo, to having an au explain why it's funny. We don't even fully understand that problem, much less have any idea of how to solve it.

http://karpathy.github.io/2012/10/22/state-of-computer-visio...

Obviously you don't have to go to those places.

1) Society tolerates a small amount of alcohol for most drivers.

2) Driving while drunk impairs your physical and mental ability to drive safely and only gets worse as you drink more. Obviously a manual driver has slower reaction times than an autonomous system but they are within a level that society tolerates and far better than someone who is drunk.

There are also risks with autonomous vehicles - the potential for equipment/sensor failure, software being unable to adequately handle bad weather, the potential for exploitation by hackers etc.

I'm currently in the process of buying a new vehicle and have been test driving vehicles with various sensors designed to prevent accidents (specifically the Subaru eyesight system). One of my first thoughts was that it would be easy to become reliant on the technology and have a serious accident when a sensor fails (particularly for blind spot detection). I'm still likely going to buy a vehicle with that functionality but am not sure how much I will trust it.

To me, drinking(alcohol) is a pointless activity, with literally no upside. It costs money, doesn't feel good, I don't see the social aspect of it, and the danger it poses to people is well documented - it's a "negative" for society in my book. But, at the same time, I am fully happy to acknowledge that there are people who enjoy it, who like to party, and I would never say that drinking should be banned.

I know a few people who treat driving like I treat drinking - they see it as boring, without any benefit, with negative consequences well documented and they clearly think that society would be better off without it. They think that the day that manual drivers are banned can't come soon enough.

As mentioned above - we are not robots. Not everyone enjoys the same things. And that's great.

- Yielding to an emergency vehicle with sirens on.

- Moving backwards to a safe and large enough spot when the route is too narrow to fit self-driving car and oncoming huge lorry (and there is no line marking the limit between road and ravine).

- Upon instructions from authority, recognize that the highway is closed due to an accident and, no matter what the driving code says, you actually have to make a U-turn on the highway and follow the crowd. Alternatively, just take that route (yes, the one with the large no-entry sign at the beginning) or that narrow path in the wood (yes, it exists, even if Google Maps isn't aware of it). At the bare minimum, park yourself off the road and let the others move on.

- Verify whether a queue is forming behind you. Listen to the honkers, they may be right. When you are an obstacle to the most part of traffic, moving to the side and letting others pass from time to time is sincerely appreciated.

Cars == freedom.

Freedom == control

Freedom as in own Geographic control. When. Where and how we get places with that freedom.

We control our freedom.

Before furnaces we burned wood. Furnaces helped. We accepted them.

But cars are different. An extension and expression of our freedom. Out health. Cars have been our best (& worst sometimes) friend as we used our freedoms. There is no substitute for this machine that affords is freedom control.

Risking the loss Is where my gut says people will fight this.

I would also feel comfortable assuming that even if they put an effort into driving at similar times as other people, they would still skew away from scenarios that engineers find boring. Good luck getting them to drive in 2mph stop and go on 880 or 101 during rush hour.

Seeing them on the road commonly is not a guarantee that their driving patterns are representative. Their driving patterns are deliberately designed to not be representative...not because it helps their metrics, but because it is the only feasible path for development. Back when the first DARPA challenge was around, the only objective was to make it from point A to point B in an empty, flat, and mostly obstructionless desert (and everybody failed). By the time of the last challenge, they had to navigate a suburban-ish area devoid of people but populated with demarcated roads and carefully placed obstructions. Now they are on relatively slow city streets with actual people but controlled times/conditions.

So this phenomenon is obviously not new...SDC engineers have known this from the beginning. But now we have a new class of SDC boosters that don't actively involve themselves with development and don't seem to realize that their metrics are representative of the current state of SDCs and not a guarantee that can reliably extrapolate.

For example city X decides to change some traffic rule on date Y. Each autonomous vehicle will need to receive an update by date Y. What happens if a vehicle hasn't been updated soon enough?

One solution is a mandatory update scheme. Every day a vehicle checks if it has received its daily update, if not the vehicle is disabled. Another more long term solution could be close co-ordination with cities to determine which upcoming traffic rule changes warrant mandatory (vehicle cannot be enabled without update) updates.

These corner cases may seem trivial, but with zero tolerance for error across hundreds of thousands of cities careful strategies will need to be developed.