Self Driving Car thread

[NervousEnergy]

dh87[/url]":34jac9yf]The companies will either carry insurance or self-insure. There seems very little chance that a company or an individual engineer could be found criminally liable. For insurance purposes, the only thing that matters is the statistics. I agree with some earlier posters who think that many individuals will not want to give up driving, ceding decisions to the car, even if it's statistically safer.

I agree that will take quite a while and a lot of persuading and culture shift, since psychology trumps statistics. The statistical safety effect is an objective thing... over time, you can easily show that X number of folks using SD cars are alive today that wouldn't have been with human piloted vehicles, but that's an abstract. People have problems assigning weight to abstracts, and they sure don't have an emotional impact or make the evening news. That one in a million fluke technical problem, though, that drives a family of 4 off a bridge? EVERYONE is going to focus in on that, and damn the fact that for every 1 person a SD car kills it saves 1000.

I don't know we can fix that.

 

[NervousEnergy]

NavyGothic[/url]":1ueoxec4]

Wudan Master[/url]":1ueoxec4]How would it work if a human has to drive the 1st and last miles of the trucks journey?

It might be workable to have drivers behave like old harbour pilots - the auto-truck drives itself to a holding facility on the city outskirts, and a local driver collects it for the more complicated delivery and / or loading.

I've thought about this idea as well, but then you get into the situation where the automated part of long-haul trucking is competing much more directly with rail. They do now, of course, but trucking is still more profitable/efficient in many instances than rail due to the point-to-point nature of the delivery. If you've got to start building large 'handoff stations' for automated trucks and possibly automated vehicle only lanes on most major interstates, then the value proposition for long-haul trucking gets a lot harder compared to rail, which already has this in place.

Reading this discussion, I'm leaning more toward not seeing self-driving cars in any significance till the surface infrastructure can better support them, and that's a lot more money. I'd love to be wrong about that, because believe me, I'd be one of the first in line to buy a vehicle that I could sleep on the way to work in.

Like everything else with the automobile, I'm thinking it's going to sneak it's way in. Lane drift warnings, blind side radar, and front collision radar/lidar/optical are all already a thing and being advertised on current (albeit higher end) models. I'd bet that some of the first regulatory issues to fall are the limitations that keep front collision avoidance from being allowed to stop your car completely. IIRC, the Audi A7 from the Ars review could (and would) actively slow your car without your input if it detected an approaching front collision, but wouldn't stop it completely if necessary due to legal restrictions. I can see these kinds of things happening first, where car automation WILL be allowed breaking control all the way to a dead stop to avoid a front collision, and even disallow a lane change if you're going to collide with someone already in the lane (though that's technically tougher.) This kind of automation combined with adaptive cruise control leaves the driver in charge of navigation and general operation, but takes most emergency responses out of their hands.

Once that kind of tech is ubiquitous in even mid-lower tier car trim levels (or even mandated safety equipment), then the next steps of actual navigation gets easier. I still think it's going to take some adaption on roadway construction, though. Bike lanes in metro areas have gotten popular over the last decade... how long till we see Google Lanes? Yes, complete with sponsored branding.

 

[NervousEnergy]

Without citations, I'm having a hard time taking the way this thread has gone seriously. People talking about reaction times in hundreds of milliseconds? Precision guiding of vehicles inside the stopping radius to consciously avoid hitting one target instead of another? Ridiculous. Perhaps if your population is nothing but NASCAR pro drivers, but even then I'd be skeptical that Jeff Gordon could pull off the dog vs. child trick more than half the time.

Look around on the road on your way in to work. The elderly lady in the Land Cruiser *may* be able to avoid a train if she sees it more than half a mile off. The guy in the suit in the i335 would have a better chance if he wasn't paying more attention to his cell phone. The teenager in the old Camaro might think he's Jeff Gordon, but he's going to be travelling too fast and be too inexperienced to avoid hitting anything. Jim Bob in his F250? He might be able to swerve around a battleship... maybe.

80% of accidents involving a collision of some kind involve inattention, and every citation I can find using that figure is from 10 years ago, pre-smart phone. I'd bet at least a dollar that the number is higher now with the rise of smart phones and complex in-dash systems. You're not going to choose between the kid and the dog, you're statistically an order of magnitude more likely to plow over both, and only realize it when you feel the bumps.

A front collision warning system that can apply full breaking down to a complete stop will beat any human driver the vast majority of the time, and save a huge number of lives. Fully autonomous cars may be further away than I'd like for many reasons, but I can see much higher automation coming to mainstream cars in the next 5-10 years. There's no reason not to allow the car to stop itself if it detects an object it's otherwise going to collide with. I'd go much further and have the car disallow steering input that's going to cause a side collision as well, or steering input that will cause a rollover. Eventually cars will be more like modern airliners... the systems will prevent you from doing anything catastrophically stupid without a lot of effort on your part to get around it.

 

[NervousEnergy]

NetMasterOC3[/url]":3jgdg5pk]

Alamout[/url]":3jgdg5pk]
Trains and planes are pretty dang automated already.

Trains, at least in North America, are not even close to "pretty dang automated". At best the "automation" is limited to stopping the train via dead man switch or signal overrun. There is some level of over-speed protection on the NEC, I think, but nowhere else in any significance in the US. Hell, there's still 100s if not thousands of miles of track that's not even controlled by CTC and is using either ABS or track warrants and no signalling at all.

That said, on one hand trains would potentially be much easier to automate. Avoidance is a non-factor, trains are on rails after all. On the other, the problems you have to address are considerably different from either planes or cars and there's likely little if any real carryover in terms of knowledge or tech. Moreover, unlike say long haul trucking, railroads, again at least in the US, are heavily unionized.

Trains are heavily automated in the sense that they need a fraction of the crew they used to. My father worked for the Southern Pacific for nearly 40 years before retiring, and watched a lot of it. One of the biggest impacts was removing the crewed caboose and replacing it with what literally looks like a small box, an EOT device. Labor fought it tooth and nail, but the redundancy argument was too good. Today the personnel costs of rail per ton of freight moved is a tiny fraction of any other transport industry, to the point where automating away the rest of the train crew will have a poor ROI unless such automation (and liability) gets very cheap.

As noted, I expect cars to move closer to trains and airliners for quite a long time before they're completely automated, outside of certain geographically tightly constrained areas (a downtown that was completely rebuilt for Google Cars and no other vehicles could be made to work very well.) Allow the driver to operate and navigate the vehicle, but disallow control inputs that could definitely cause a crash and initiate inputs required to avoid one. Full braking for the obstacle in your path until you maneuver around it. Side swerving (if space is available) or braking (if not) to avoid someone swerving into your lane. Disallowing a full wheel turn on the highway at 70 if it would result in a side collision or rollover. Etc..

Poorly designed automation and poor pilot training/execution doomed flight 447, but similar automation (the computer decides what inputs are safe, which are disallowed, and which are initiated automatically) would save tens of thousands every year on roadways.

Or we could go the other way and mandate that all passenger cars need full roll cages, multi-point restraints, and HANs devices, and you'd also see fatalities and even serious injuries plummet. Tough to sell the American public on that one, though. Creeping safety automation is easier to get away with.

 

[NervousEnergy]

NYC already has a very efficient above-ground people mover system... it's the enormous fleet of yellow cabs. I remember the first time I visited NYC and was totally shocked at how more than half the cars in view at any time (and sometimes nearly all of them) were cabs. Hold my hand up, and within seconds (if not instantly) a cab was pulling up right in front of me, and minutes later I was at the next destination (and $15 lighter.) Replace those with Google-style pods in a more controlled street environment and you'd change the nature of surface life in NY a tremendous amount. It'd sure be quieter at least.

One thing that gets emphasized a lot in research on self-driving cars that the public rarely seems to grasp is how a busy traffic pattern full of slower automated cars will typically get people to their destinations much faster than a faster but less consistent traffic pattern of human drivers.

Of course, if we're wishing, I'd rather ban ICE completely from NYC, dome it, and replace all the streets with variable-speed moving pedways. But then we'd probably have to replace all the cops with Sandmen...

 

[NervousEnergy]

river-wind[/url]":3f648nmd]Little steps on the availability, but the lane departure warning and blind spot collision detection systems on the 2015 Subaru I had for a week as a loaner both worked very well. The design was good and unobtrusive, and it didn't misfire when driving near jersey walls. It even alerted me to a parent with a kid walking behind me in a parking lot that I couldn't see due to a parked car next to me.

I was impressed.

What about camera lens cleanliness, I wonder. Google's system can't work in rain or snow due to the impact on visual distinctness of objects; if in everyday use these cameras get dirt on them, they'd stop working. How best can the user know this is an issue and effectively clean the lens/cover?

The last sounds more like a hardware maintenance issue than anything. I can see the same kind of tech you see with the in-car cameras at Indy/NASCAR events: a clear lens cover that rotates when it gets dirty, with the dirty section rotating through a cleaning brush before it comes back around.

This is going to be the tech that pushes the public closer to full automation, even if it may ultimately delay such full automation due to being 'good enough'. The next step is the collision detection system not allowing steering input that results in a collision, and braking to a full stop if necessary to avoid a crash. Neither is legal (AFAIK) right now... the Ars article on the Audi A8/S8 stated that the car could initiate breaking / slow down if it detected an object in path, but it couldn't stop completely to avoid the collision due to legal constraints on the scope of automation. I'm betting that's the first thing to fall. People will still drive and navigate the car, but over the next decade they'll increasingly restrict inputs that will cause a crash, and initiate inputs to avoid one.

That's a lot easier than navigating roadscapes that to date haven't been designed for easy automatic visual interpretation, especially in construction zones.

 

[NervousEnergy]

pluckedkiwi[/url]":3q0ri717]Not everyone has had good experiences with cabs. I have been in a few where the driving was excessively dangerous or where the cabby spent the trip yelling at someone on the phone the whole time, and I'm sure I'm not the only one.

Many people go for the teller rather than the ATM because the ATM is not as quick or easy for making various transactions. If it is for a simple withdraw, those people are probably just techno phobic. A self-driving car might need you to type in your destination, but that shouldn't be much of a problem.
Many self-checkout lines have poor interfaces and functioning - I don't blame people for not wanting endless 'put the item on the belt, not recognized, try again' garbage. Like anything else, if the UX is bad, adoption will be poor. Think of the difference between the original blackberry and the iPhone - user experience makes all the difference.

While maintenance of the auto-cab will be higher, the lack of labor costs for the cabby should more than compensate, such that auto-cabs will be significantly cheaper as well as being safer, more reliable, and not going to cheat you.

Hell yes. It's been nearly a decade since I was regularly traveling to NYC, and the cabs didn't especially bother me, but my wife still tells stories of the first time I put her in a cab going from Penn Station to the Met. 5 minute trip, and she couldn't watch more than a few seconds of it. Missing pedestrians, cabs, and curbs by inches constantly. I thought it was a virtuoso driving performance (that, to the cabbie, was utterly mundane), but she hated it with a passion. We did the subway more after that. If she could get into a driverless car that moved at a more sedate pace, with none of the... personal driving peculiarities of a manned pilot, she'd pay MORE for the privilege. If it was actually cheaper then so much the better.

 

[NervousEnergy]

Couple of things come to mind on the liability front:

- I think it's obvious they'll need at least some legal/regulatory framework that doesn't exist yet, or is still nascent. Not sure what that will be, and I hope it doesn't go too far protecting the car companies at the expense of the consumer, but they're going to need something. It might be that insurance companies wind up footing the bill for cybernetic mistakes, in the expectation that their rarity will be more than balanced by the huge drop in accident costs due to removing the accident-prone wetware pilot.

- The sensor suites and resultant data gathering will make any accident extremely transparent, as we've already seen in Cali. I'd be willing to bet that not only will the vast, vast majority of accidents be the fault of other human piloted vehicles, but that it'll also be extremely obvious that it was so. Having virtually all the ambiguity of who did what in an accident removed is a big deal for fault and financial responsibility.

 

[NervousEnergy]

jungwolf[/url]":39peozcn]I easily imagine it taking that long for a person to go from not paying attention to orienting themselves enough to make an informed decision. Especially since it is a situation that the car can't handle, probably a complex situation or one where it is hard to see what's happening.

It depends on what state you assume the person is coming from. If they've completely checked out of the driving experience and are texting away on their phone or reading, then yes it's going to take a subjective eternity to check back in, figure out what the car is doing and what everyone around you is doing, and react. That's going to be an order of magnitude too long.

On the other hand, it's hard to imagine an emergency where the automation isn't going to do a better job of keeping you alive than you will. Collision avoidance is one of the things a computer is simply going to be light years better at under almost any circumstances. I say 'almost' out of a reflexive need not to post in absolutes, but I'm having a hard time coming up with edge cases where it won't be better.

It gets trickier when you consider the non-emergency handoff. The car is about to exit off of the highway or major thoroughfare into a complex urban roadscape with construction and imperfect maps, and decides it can't handle it. Even then, though, it should be an easy thing for the car to simply pull over and stop if you don't take control in a timely fashion.

I'm not too keen on autopilots that aren't fully baked, which includes a road infrastructure (especially construction zones) that's been modified to cater to them. I *am* keen on collision avoidance systems. I'd love to see cars in the transition period to fully autonomous mode behave more like an Airbus liner in normal law. You navigate the car, but the computer decides if it's safe from a basic 'am I going to hit something' perspective. Car stops in front of you (or any other object is detected) but you don't hit the brake: the car does. You try to change lanes into an occupied lane: the steering doesn't accept the input. The vast, vast majority of accidents today are due to inattentive driving, and this level of avoidance automation would save countless lives without tackling the much larger problems of autonomous navigation.

 

[NervousEnergy]

Alamout[/url]":1ufio7n0]
Then again, "totally fucks up in certain extreme situations" is a perfectly acceptable level of performance, since it's still way better than people.

There's going to be a perception problem here, though, since an automation failure is likely (at least some of the time) to produce an accident that a human driver would never have precipitated. 'Car drives screaming family over a bridge out for construction at 70 MPH in front of dozens of horrified witnesses!' is an all too likely possibility, especially in the early days. No matter that the cars would save 1000 people from accidents for the 1 they cause that wouldn't have happened with a human driver, the 1 is going to get a ton of attention and airplay. The safety aspect is going to need some time to seep into the cultural psyche. Same thing with aircraft... they're several orders of magnitude safer than cars, but air crashes still command a horrid fascination far out of proportion to their impact, and largely because the people involved in them usually never have anything to do with it. The thought of screwing up and driving yourself into a brick wall at 70 is easily accepted by almost anyone, but the thought of being driven into that wall while you can't do anything about it is something else entirely. We've got to figure out how to get past that.

Simple mechanical breakdowns are also something that needs some planning, but on reflection I can't see how the automation wouldn't do a better job of that as well. Something fails catastrophically like a front wheel blowout and I'd much rather have a computer figuring out how to get the car to the side of the road without a loss of control and rolling. It can let other cars around it know that it's got limited control and stay clear, and give AAA a tow call for you once you stop moving.

 

[NervousEnergy]

Syonyk[/url]":2u3mp55d]We'll have ceded control for the normal cases, and have drivers with no fallback experience to deal with the weird stuff.

We've got maybe 10 years to go down this path fully or there will be significant issues with handing control back to humans. Possibly less. That new drivers generally can't afford vehicles with heavy assistance now is a helpful for now, but that won't be the case forever.

Or we accept that people won't be able to get around in bad weather, and quite probably if the cell network is down.

But that depends on your definition of 'weird stuff'.

I see it more like the evolution of flight control in a typical Airbus model. In normal law, you don't fly the airplane, or at least not in any traditional sense. You tell it where you want to go, and it decides exactly how to do it... where to put the flight control surfaces, engine throttle, flaps, etc. Cars will get to the same place. The difficult, higher-brain function of navigating a changeable world with construction, bad weather, and all that will be the last thing to go, and may never go completely. Full control over the wheels and throttle, however, will be a thing of the past. The car simply won't intentionally collide with something no matter how hard you try and steer it into an occupied lane, or how hard you've got the throttle down while watching your cell phone and traffic stops ahead. Absent some catastrophic mechanical/logical failure, you won't be allowed to crash. You'll still have to drive/navigate the car, however, in a lot of situations.

This change alone - auto braking and disallowed steering inputs that would otherwise cause a collision will save tens of thousands every year. We're talking tens of Billions (hundreds?) in dollar savings to society as a consequence of all those lives not lost.

 

[NervousEnergy]

Sigh... I shouldn't have posted the Airbus analogy, as I should have realized it would be taken that way, rather than as an overall descriptor. Yes, I'd much rather fly on a Boeing aircraft as well due to the same issue.

Pilots, however, are generally trained to a much higher degree on doing their job than your average driver. We are so bad at it, collectively, that allowing all drivers 'ultimate authority of control' is going to kill many, many more people than not.

 

[NervousEnergy]

redleader[/url]":39hi9s4z]
They can communicate with each other their intentions (e.g. 'I intend to merge'), coordinate actions ('are you ok with me drafting behind you'), and provide information about conditions ('possible obstruction beyond the next turn'). All of this can be done without trust. As I said before, one generally does not completely trust the output of individual sensors, but instead treats them in a bayesian fashion depending on how reliable they are thought to be. The same approach (but much lower weighting) can be used for sensors you do not control. Completely trusting anything should not happen in a system that lives depend on.

I've got to agree with HeadlessCow here... Not sure how any of this is useful if the car is 100% making it's own decisions (as it should be) on things like safe following distance and braking. There's ALWAYS a 'possible obstruction beyond the next turn', from the point of view of the driving system. Intent to merge is even more problematic... when teaching my daughter to drive it took me ages to tell her that even though she has a legal obligation to signal turns and can be pulled over if she doesn't, she should NEVER trust that signal on other cars and *never* do ANYTHING that relies on them actually following through for her to not get hit. Even under the best of circumstances with an automated car the occupants could suddenly give the vehicle a different destination that results in it aborting a signaled turn or lane change. Signals are to help prepare the slow, inefficient human wetware of changing conditions. They're completely unnecessary and even potentially dangerous with an autonomous vehicle that maintains strict avoidance standards based on its own sensors.

Automating route optimization through Waze or Google Maps is a totally different thing, and I'd expect a much greater degree of trust on the part of the vehicle here. Though I'd still want some form of 'opt out' when the car wants to change routes to a given destination... it's one thing for Google to tell me the 'fastest route has changed' and for me to actually want to take it up on its idea if I know the new route isn't going through an area I want to drive through.

 

[NervousEnergy]

redleader[/url]":260x45q7]I don't think V2V could realistically work like is being assumed. Any message sent out has to be relative to the sender's location, which receiver only knows approximately (if you can even see it, in which case you may not even know where a sender is even approximately). You'd get a message that is something like "problem ahead +/-45 degrees range 0-500 meters" and then have to determine where you think it came from using your own sensors. In that case you would not stop if you received a message, at least not until you could figure out where it came from and if it is applicable to your current trajectory. If there was nothing there, you might go slower while continuing to look, but would not stop. Being able to reduce the speed of traffic by 5 or 10 mph would be annoying, but probably not catastrophic. Or at least not nearly as dangerous as someone compromising some other aspect of the car's controls/sensors.

Agreed... having the car logic slow the vehicle if V2V notifications show a wreck or other obstruction ahead would, IMHO, be the extent of action you'd see taken. If V2V comms could actually cause cars to pull off the road and stop it would be a very juicy target for sabotage.

An attacker would probably get more effect out of hacking Waze or Google Maps to cause a bunch of vehicles to take a jammed route than V2V.

 

[NervousEnergy]

jungwolf[/url]":3ib4v341]

DeedlitCryogenic[/url]":3ib4v341]

wco81[/url]":3ib4v341]For instance, there's this one big intersection where a lot of cars take a turn. So the lane leading to the turn lanes always back up and if just stay in line, you might go through 2 or 3 cycles of red lights before you get to make the turn.

So I'm used to just driving along the next lane and jumping in where there's a gap, as opposed to getting at the back of the queue.

My expectation is that when you program a route into an SDC, it knows it has to make a turn at these intersections so it will stay in the turn lane well ahead of the intersection, resulting in the SDC taking longer than experienced drivers who are a bit more aggressive.

That kind of experience might sour some people on SDCs, at least initially.

If they even notice. Add a wifi hotspot to the car, and quite a few people will just be using their laptop/tablet and not really pay attention to much of their trip.

Seriously. I can either 1) save 10% transit time driving manually or 2) be surprised when I get home because I've been engrossed in reading.

99.9% self-driving car that goes 55 on the highway, 35 on major streets and 20 on residential? Done.

Oh hell yes... and I'd be happy with even a bit slower than that. Once people get over the 'white knuckle' experience of being in a car with no driver, I'd bet 90% of them sleep, 5% work, 3% read, 1% have sex, and the remaining 1% bitch about how slow the car is.

 

[NervousEnergy]

New to me: The 15% not wearing seat belts/shoulder harnesses are about half of the crash deaths in the US.
The percentage of the 2.3 million treated in emergency rooms isn't known.
https://www.cdc.gov/motorvehiclesafety/ ... facts.html

Maybe autonomous cars can be programmed to stop if anyone is not strapped in.

Had been checking on the death rate versus vehicle speed.
Pretty much nobody dies in a crash when their vehicle is going less than 35 mph IF they are strapped in and there is an air bag.

I'm not sure how seat belt use will shake out in the SDC future. For the next 10-15 years while the roads are still filled with cars driven by unpredictable humans, it would make sense to enforce seat belts in SDCs, though I wonder if the buying public might balk at technical enforcement with movement lockout.

Go out 20+ years where the manually driven fleet has been mostly mechanically deprecated and/or license restricted to death (especially in highly urban areas) and I'd expect seat belts usage to suffer greatly. It will likely always be a good idea to be wearing one, as mechanical failures and freak accidents will still cause the occasional crash even if it's 1% of what we suffer through now. But if safety behavior is as good as it looks like it may be and those pesky humans get taken completely out of the decision loop it'll be really easy for most people to ditch the SB altogether.

 

[NervousEnergy]

A mixed environment leads to all the problems we have now, with a lot more miles being driven.

I think that's unduly pessimistic. Replace half the human drivers with machine drivers in any given situation and you should get a MORE than half reduction in accidents. SDCs shouldn't only be a lot better at not hitting something, but also avoiding being hit by something.

 

[NervousEnergy]

If it becomes a problem, I'd expect quick passage of laws to make intentional interference with an SDC to be worth heavy fines. The sensor suites on L4 cars in 4-5 years time should make identifying someone doing this a trivial issue.

 

[NervousEnergy]

It still seems the real value from the SDC taxi angle is in the app. How easy is it to order (uber?) an SDC to get you? How fast? How about options customization and what does that do for the lead time to get your vehicle (such as more seats, car seats, extra cargo room, etc.)? Will the app have price tiers based on whether or not you're willing to share the ride, and how much extra time overhead you're willing to live with?

I can see the SDC taxi base vehicle experience pretty well commoditized, where the app-centric service delivery is the high-end value add.

 

[NervousEnergy]

I can see the SDC taxi base vehicle experience pretty well commoditized, where the app-centric service delivery is the high-end value add.

Maybe, but this isn't anyone's pitch that I've seen. Most of the cost of a taxi is the driver. If you can get rid of the driver, you can cut costs well below Uber, while still being very profitable, which means you'll attract a lot more riders, which means even more profits and even lower costs. Everyone seems to be aiming at lower prices with the goal of growing the customer base of the current taxi market. Value add is probably still some decades away once the engineering is really worked out.

I'd strongly disagree with 'decades' away. I'd be skeptical of years. Value add will come *fast* when SDCs are mass-market viable. As you point out, the main cost is the driver, and without them the barriers to entry for an SDC ride-sharing service will be very low since it's all capital... you don't need drivers, you just need the initial capital for the cars (and you can start small), a maintenance / recharging facility (assuming those aren't ubiquitous at that point), and the app/service. Value adds to differentiate SDC ride sharing will happen almost day one.

- Click the box to include a child seat or two.
- Click the box to have your vehicle show up to take you to work with your choice of Starbucks drink in the cup holder
- Select a wider time frame for pickup/destination arrival to lower your cost (priority scheduling.)
- Select sharing with other travelers for lower cost, complete with a slider to give your range variability and if you can be dropped off last

And that's just off the top of my head. If you're not thinking of how you can attract people to your SDC taxi service with that kind of flexibility I guarantee your competitors are. You can offer almost any service level... from nearly your own car experience (just you, straight to your destination, no stops, coffee in the cup holder, max price) to an SDC minivan with 7 passengers (lowest cost). Flexible experience tailored to price vs. convenience.

Hell, I need to see if I can sneak an app patent in... :devious:

 

[NervousEnergy]

If the public use self driving car takes off I bet the whole notion of what cars are supposed to look like and feel on the inside will be turned on its head. Likely you'd see vehicles that are built more rugged inside and out and a lot less sexy to look at since nobody will have their identity tied to it.

This is a significant point regarding design. Consumers who ditch the notion of a personal automobile and go 100% 'taxi' style SDC will become completely unconcerned about the appearance of the vehicle, and pay by the experience. For your everyday commute to work the SDC can look like a bad 70's sci-fi notion of a completely featureless pod. The car as fashion and socioeconomic symbol will fade for them. For a nice weekend date out they'll pay a bit more for something upscale and stylish. For a camping trip something rugged, spacious, and long-range. The ability to alter how you pay for transportation by going to the single experience model instead of a massive multi-year loan commitment for one vehicle will dramatically alter how and what kind of SDCs are made.

One significant negative issue for suburban locations is that if significant numbers of suburbanites ditch their cars and go for the SDC taxi then you'll be doubling the number of vehicle miles driven on those roads, as cars have to drive into and out of the neighborhood both morning and afternoon. SDCs should be *massively* better at handling congestion (and not causing it in the first place), but only when you've got a critical mass of them making up the vehicles on the road.

 

[NervousEnergy]

I believe the outcome everyone wants it to charge a toll on all non-residents who drive down the street. As long as you don't pay a toll to get to work.

I *think* you're being sarcastic, but if not, I'm in one of those zones where there is a "cutoff" that runs through my neighborhood, and is *full* of cars during commute hours. We also have several schools local to here, and they just happen to walk to school in the same streets. We have the crossing guard lights, but I've also seen kids scared shitless because some asshole on his phone and sipping his Starbucks with his other hand, while steering with his knees and doing 40 in a 25, just didn't happen to see the kid.

There has to be a better algorithm than that.

We have a similar problem. Major construction on the highway (114) outside of our town, and people going through town to get around it, right through a major school zone. The response was to put in the fancy 'flashing lights' school crossing with lights at both ends and along the walkway when someone pushes the button, in addition to a heavy police presence. It's a major ticket factory through there, and School Zone tickets are painful. A citation in either a school or construction zone (with workers present) is double the fine and cannot be resolved with defensive driving or deferred adjudication. Pay up or go to court.

It's not perfect, and people are starting to ask about more technical solutions. Conditional speed bumps (only rise up during school zone hours) are on the wish list, but cost a fortune.

 

[NervousEnergy]

One possibility in the GM/Ford boardrooms: "Tesla doesn't need Waymo so we sure as shit don't."

I suppose it's reasonable to assume that SDC will be a universal thing for any car maker that wants to remain relevant over the next decade. With that in mind, investing that kind of money now - which as Frennzy points out they can totally afford - means that they might come up with something cool on their own but it could be, since they are later to the game than others and progress is moving quick, around $1B is the amount to catch up and at least properly understand what the others are up to. If you're going to license SDC tech from another company and incorporate it across some significant percentage of your fleet you should at least know how it works.

I hope that's the case. Tesla seems at least adept at software design, and Apple of course owes a significant chunk of it's existence to excellence in software design, but when I think of The Big Automotive Three and software in the same breath I think of the complete disaster that is the infotainment system in any modern vehicle (prior to Apple/Google taking that over.) It's always seemed to me that with enough money you could buy all the expertise you need in a given field, but in practice that's just not true.

On the Apple front, though, SDC control still seems an odd endeavor for them that they have little to no corporate experience with. It's not like they're big in industrial controls or embedded systems, or much of anything else that would seem to give them an institutional edge here. UI supremacy is great but it's the tip of the iceberg for SDC control systems.

 

[NervousEnergy]

AEB is planned to be standard equipment in the US by 2022.

http://www.iihs.org/iihs/news/desktopne ... w-vehicles

The question for me is how quickly will these features become redundant with level 4+ self-driving capability?

A couple of years ago in this debate I was of the firm opinion that true self-driving automation was 10+ years away on any US public road, but that in the interim cars would become more like an Airbus in normal law - the driver 'drives' the car by telling it where to go (you could use a joystick pointer if you didn't want a wheel), but the final word on if the car would actually go there would depend on inputs from it's sensors. The concept of direct wheel, brake, and throttle control through mechanical linkage would be gone... you'd use those three interfaces to tell the car where you want to go, but the computer was the final arbiter and wouldn't let you hit something, drive off a cliff, roll it, etc.

Once it became clear that every major player was pouring tens of billions into true self-driving automation I wasn't so sure of that anymore. It seemed that level of sensor capability would just feed the computer pilot in the first place, taking the human completely out of the loop even for high-level navigation. Now I'm wavering back that way.

 

[NervousEnergy]

Tesla forcefully blames driver for fatal crash.

https://meincmagazine.com/cars/2018/04/te ... t-perfect/

Guess they don't want any suggestion that auto-pilot is dangerous but is it good PR to get into a dispute with the survivors of the deceased?

Legally they're probably covered by whatever terms that people who activate the auto-pilot package agree to.

Maybe a quiet trial that absolves them of liability would have been the best move.

The problem is that a 'quiet trial' leaves the big, bold headlines of "AUTOPILOT KILLED HIM" dominating the news cycle without any rebuttal. Tesla has made the decision that they can't risk that kind of heavy brand blackening without trying to do something about it.

Whether or not that's a good decision is more than I can say.

 

[NervousEnergy]

That might go much further to improve safety, in the short term, than aiming for certain levels of "autonomy."

But would we (in a general sense) actually be willing to pay in terms of extra training, licensing fees, and general hassle?

Eventually this issue will fix itself - as cars become obviously better/safer at driving than humans then the barriers to letting accident-prone human drivers mix it up with the machines will rise, mainly due to public demand. In 30 years I'd expect drivers licenses to start to look more like general aviation pilots licenses. Much more of a luxury rather than a need. That may also evolve regionally even in the US... I wouldn't be surprised to see municipalities like Manhattan or downtown Chicago/LA move to far more restrictive requirements ahead of rural Montana.

 

[NervousEnergy]

Eventually this issue will fix itself - as cars become obviously better/safer at driving than humans...

Assuming that's possible, sure. I'm still skeptical that a computer can do everything a human can in terms of driving.

why, are you under the impression that we're really good at it?

Given that we do as well as we do with fairly minimal training, in spite of myriad distractions and multitasking with other processes, with limited sensing capability, etc., we actually seem to be pretty good at it.

We're really good at ambiguity and edge cases. We're really good at figuring out a confusing mess of a roadway under partial construction. We're good at correctly getting around high visual interference from rain, snow, or fog. We suck at paying attention in the normal course of driving when the situation *isn't* challenging.

Level 5 will likely require some significant changes to how we build and maintain roads. You can half-ass it all and trust the human wetware to just 'figure it out' on the fly, but that won't work as well with machine intelligence, or what passes for it.

Here is a question I struggle with: how will the public react to vastly safer cars that still occasionally crash in ways that are emotionally horrifying because no human would do that? Say in 10 years L5 cars are available that, nationally, would reduce our vehicular death rate from ~40,000 per year by two orders of magnitude to ~400 per year, but those 400 die in ways like the recent crash in California - driving into a solid barrier at high speed, or off a cliff, or something similar due to either a freak mis-interpretation of the environment by the electronics or some sort of physical/electronic failure of the system that included a failure to pull of the road with the 'call service' light. Cue cell phone video of screaming family being blithely driven over the rail on Cali's Highway 1 or some such. By any rational analysis any sane person would take the risk of being one of those 400 over being one of the 40,000, but I have a bad feeling the sensationalism of those kinds of crashes will have them loom over the public consciousness with far more weight than they should get.

 

[NervousEnergy]

By any rational analysis any sane person would take the risk of being one of those 400 over being one of the 40,000, but I have a bad feeling the sensationalism of those kinds of crashes will have them loom over the public consciousness with far more weight than they should get.

Unless you're actually a person who trains for adverse situations and tests/trains those skills. I'll take my chances with the 40k. Since, you know, distributions aren't even. And will completely resent the choice being taken away.

Sounds like another self-* issue.

That still may not be a rational decision on your part... I'd have to dive deeper into reported accident statistics to see if there's any tracking of fatalities for those completely uninvolved in the initial fault. Distracted drivers crossing into oncoming traffic on a divided road probably being the most common, but living in a huge city for most of my life with millions of drivers on the road every day you see the craziest things weekly on the news. Driver hitting a guard rail and having the airborne car flipping into multiple vehicles on the other side and such. Drunk wrong-way drivers speeding down the highway. I have no idea what the numbers are for people dying in their cars due to sudden, unavoidable circumstances from another driver screwing up, but I'd be surprised if it was less than 400 a year across the US just from the Dallas evening news.

What degree of statistical assurance would it take to get you to trust the machine? At 400 deaths per year given the enormous number of man-miles driven in the US I'd have to think your odds of dying to be so low from a vehicular crash as to not justify that time spent training/testing a high level of defensive driving ability.

 

[NervousEnergy]

What degree of statistical assurance would it take to get you to trust the machine?

The correct analysis isn't statistical, but situational first. Bottom up, not top down. You need to cover all the common edge cases. Within my 3 mile, within city limits, commute, within the last year

2-4 inches of snow on the road.
Unmarked single vehicle width road up to my house with the ability to take turns to use that single vehicle width road.
flooding of 6-12 inches in the turning lane due to fall leaf plugging of drains.
cyclists ignoring road rules and going on the wrong signals, while blocked by a concrete divider from the road until the last minute.
multiple train track crossings where the tracks are exposed and the road surface severely potholed.
a queue of semis taking over the middle divider to get into shipping terminals, combined with a poorly marked track with one of the main bike divider routes.

Also, assurance that when the machine fails, it 'fails to safe' and doesn't try to drop the whole load of poop in my lap.

I'll gopro my 3 mile drive for ya if you like, on a rainy day.

It may end up that there are some conditions or regions just won't be able to be managed, and thus the SDC won't drive there. It may take well maintained roads and extra effort to ensure they're viable to drive - wether's that's premapping, better signage, better maintenance of roads etc, that will rule out entire areas. That leaves plenty of places where they could well bee a viable, safer alternative to human operated cars, and from all indications that will happen.

This may well be how it evolves. I can small, dense, tightly controlled urban areas eventually enforcing L5 control. Manhattan would get a lot of benefit once the tech is ready, rural Kansas not as much from an overall congestion and traffic control perspective.

My original question on public acceptance was posited 10-15 years from now when L5 tech would presumably be much closer to fully solved, as it's got to be very well developed to reduce road accident deaths by 99%. My concern is that even in such a rosy future, those 400 deaths (1% of the current total) may be things the public has a hard time with emotionally, since they're likely to be accidents a human wouldn't have had.

Edit: Nod to Happysin.

 

[NervousEnergy]

To pre-empt the obvious "a rear-end accident is always and entirely the fault of the following car's driver": While this is a good thing to teach in driving schools and serves as a general guideline it is not universally true.

Even more so, it's only generally true because it's hard to determine who is at fault, therefore it defaults to the rear driver. That's not the case for a self-driving vehicle, because there are copious logs of what it was thinking and why it made its decision..

Well, and the fact that the following driver NOT being at fault is a pretty edge case. Insurance fraud 'swoop and stop' behavior comes to mind, but that's all I can think of. If you're not paying enough attention to the car in front of you it's obviously your fault, but even if you are and still can't stop in time you're almost certainly guilty of either tailgating (and in most states the legal definition of tailgating is pretty far back at highway speeds) or having faulty brakes, both of which are also your fault.

 

[NervousEnergy]

I turned most of that off tbh. Less situational awareness than just turning my head. Especially since if I spot that person approaching at speed it'll beat the sensor. Only modern nicety that I think is pretty essential is the rear camera for parking. I'd like a 'top view' 360 for the front as well.

That kinda blows my mind. My rear camera and cross-traffic warning sees cars I can't easily see when backing out of a space, due to the extremely wide angle lens on the back that's not impeded by the huge SUV's parked on either side of me and blocking my view. The system telling me when I'm getting a corner too close when parking is also invaluable (system looks at 8 cardinal points around the car.) More information is always better, IMHO.

Potential insufferable attitudes aside, I had a different take on the school's purpose. I saw them as cultivating administrators\technicians for whatever habitats might be in orbit around some body or another, or on the ground on some body that isn't earth. Not far fetched to think Elon wants to position his children (and children of close associates) in prime places to take advantage of his life's work.

Yeah, those kids are going to grow up to be mostly useless outside their exact niche, and probably will be insufferable to be around as a bonus.

I don't know if all of these comments dissing on Ad Adstra are serious or just trolling. If serious, then... wow. Y'all don't remember how utterly boring and useless 90% of elementary & grade school were?? These kids are going to run rings around traditionally educated high-achieving students in initiative, problem solving, creative approaches, initiative... yeah I said it twice, since initiative is one of the main things the 'memorize all this stuff tonight so you can blow out the state STAR test tomorrow' method of teaching crushes out of kids. I'd have given an important but redundant body part to have gone to a school like that, instead of my 'spend 5 minutes paying attention so I can ace the homework, then the next 50 minutes drawing pictures of spaceships and dreaming of stories' elementary-middle school experience. Off-topic, but I'm shocked at the responses to that education story in this thread.

Edit: ++ to Redleader, who posted while I was composing.

 

[NervousEnergy]

I don't know if all of these comments dissing on Ad Adstra are serious or just trolling.

I'm not sure why you included my remarks as dissing the school. I wasn't at all, merely commenting on one of the (probably many) reasons they have for starting the school.

Yep, my apologies... I grabbed the wrong quote.

 

[NervousEnergy]

1. You're the sapient person in that situation? The car isn't (yet)?
2. In quite a lot of vehicles, the captain is responsible for the wellbeing of everyone onboard. If they hand off control to a first officer, you best believe they're still accountable to the results. "I wasn't the one flying/piloting at the time" isn't an excuse.

And to be clear, I'm not saying Autopilot is perfect (clearly not, or they wouldn't be offering updates) or Tesla is fully blameless in how it's used. I'm just saying I think they're on the correct side of the curve.

Sapient is an artificially high bar. What's needed is SAE level 5, which is not sapience. Maybe you could get by with level 4 as long as the vehicle responds safely in the situations that it can't handle. Levels between 2 and 4 shouldn't be released for public use ever and that's where Tesla's Autopilot has been operating, pushing the boundary between level 3 and level 4 because it can be expected that a human who's not driving the vehicle can't take over at a second's notice. There are other companies doing the same and it shouldn't be allowed. You're either in continuous command of the car or you shouldn't be expected to respond in

I think we're going to have to recognize that SAE 4 or 5 is going to require as much or more attention to the roadway infrastructure as it does to the cars.

 

[NervousEnergy]

I think automation is going to get really good at avoiding (or reducing the impact of) the common big accidents, which is where the majority of people die.

Fortunately, this argument is going to be easy to settle with the statistics. I think 2019 will provide good data (if only from the Model 3 fleet) to make statistically valid distinctions on the relative safety of the systems.

I agree, but I think you may have to wait till 2020 or 2021 to start to get a statistically significant volume of data.

A lot of this can be resolved by how the vehicle responds when it thinks it's in over it's head. If you're going 70 MPH on the highway and it suddenly decides to completely dump all control in the drivers lap as it's approaching a confusing intersection, then yes that's a terrible design which will cause accidents. If the vehicle immediately slows as fast as it safely can and pulls over and says YOUR TURN!, then that's different.

The big issue I see is emotion vs. statistics. If L3 cars significantly reduce injury and fatality accidents, but do produce the occasional news-headline accident where the car kills someone by doing something even a drunk on his/her phone wouldn't do, then I have a bad feeling the emotional response to the rare but spectacular automation failure accident will drag on acceptance when it's not rational to do so.

 

[NervousEnergy]

Translation: The car should be able to drive you anywhere, but the human must take over without notice if the vehicle decides it's in a situation it can't handle.

That sucks and it needs to not do that. It should be trivial to have the 'situation I can't handle' routine be 'STOP', not 'Take over at 70 MPH'.

 

[NervousEnergy]

That is very much not the safest option in many situations. This is why most of us are saying that L3 is going to be dangerous as fuck.

Assuming that your human driver isn't paying attention, under what situations would it NOT be the safest option? Yes, there's an increased risk of a rear-end collision if the person behind the SDC isn't paying attention themselves, but that's small potatoes compared to continuing forward at highway speeds toward a possible collision with an AI that's totally gone home and a driver who's literally asleep.

The best general purpose reaction to an AI with a panic attack is to see a safe route to pull over outside a traffic lane and stop. If it can't see such a path, or has suffered some catastrophic technical failure, is to emergency stop. Yes, that's dangerous, but nowhere near as dangerous as proceeding forward blind and out of control.

Your whole contention is that right now people have to be ready to take over full driving capabilities in a time frame measured in tenths of a second, and that this is just flat out impossible even for a fairly alert driver. I contend that you have to assume the human driver is 100% incapable of emergency response. They're asleep, drunk, or playing Pokemon Go. If the SDC has confidence in it's sensors and can assume it's not about to collide with something then the reaction to loss of confidence in it's decision making should be gentle braking and maneuver to the side of the road and stop. If the AI doesn't have confidence in the sensors it should be emergency full stop. If the human at any time signals to the AI that he/she can assume control, only then should the AI abort which of the two emergency routines it's activated.

 

[NervousEnergy]

They got a pass from a criminal perspective, I believe, though criminal negligence might not have been a stretch. I'm pretty sure they're getting the living daylights sued out them in civil court, though. Gross negligence should be a slam dunk.

 

[NervousEnergy]

They got a pass from a criminal perspective, I believe, though criminal negligence might not have been a stretch. I'm pretty sure they're getting the living daylights sued out them in civil court, though. Gross negligence should be a slam dunk.

Sued by who, they have already settled with the family.

Ahh... yep, just looked it up. Nevermind. Given how quickly that got resolved it must have been a fat check.

 

[NervousEnergy]

Would a GPS failure really be all that catastrophic, though? Assuming even minimal steps were taken to mitigate it. 'You are here' physical (or Wifi) markers that interface with the maps stored by the car would help quite a bit. If the car was periodically assured where it was with regards to it's maps and what it saw/confirmed by visual comparison with map features then it should be pretty difficult to get lost, unless the maps were really out of date.

And if the onboard maps are that out of date then the car will have navigation problems that GPS will only partially help with.

 

[NervousEnergy]

The essential argument is that the cost of running an autonomous taxi business with an 8% ROI is higher than the cost of owning a private car, so either investors accept lower margins in order to ??? or the business isn’t viable in most cases.

So the argument is that a private individual can buy a car at a borrowing cost of, say, 4%. While a business has a required ROI of 8%, so private ownership is cheaper?

Seems like that assumes a one-to-one correspondence between private cars and taxis. That is, for every private car that would have otherwise been owned, there is now a taxi instead. That almost certainly would not be how a taxi-heavy transportation system would work. As Shavano pointed out, the taxi fleet only needs to be able to handle peak traffic, while there are many more privately-owned cars than are being driven during peak traffic. That means that anyone who does not need to travel at peak times can get rid of their privately owned car without an additional taxi having to be added to the taxi fleet.

EDIT: And that doesn't even factor in the distinct possibility that a taxi-heavy system would result in more ride-sharing. One of the big reasons people don't like to carpool is that then you're stuck at work without a car. If you're already taking a taxi to/from work, then that argument against ride-sharing simply becomes irrelevant. There are other reasons why every taxi trip wouldn't be a ride-share, of course. Some people are willing to pay more simply not to have to share (whether all the time or just one day when they are in a hurry). Some people will be going somewhere that the ride-sharing algorithm is unable to match another rider. Etc.

EDIT2: Just to be clear, also, I conceded that I could be misunderstanding the point. I'm just reacting to how I interpreted your comment.

I agree with this... you don't have to capture the (really massive in the long term) cost savings of building most dwellings (homes, apartments, duplexes, etc.) without garages or driveways to get a taxi fleet to a lower cost. Even individuals calling a SD car to take them, and them alone, in at 8 and home at 5 would be cheaper due to consolidated parking/storage and massive economies of scale in maintenance. The Uber of the future will likely have a sliding scale of cost depending on how many other people you're willing to commute with, and how much time you're willing to commit to the trip. An Uber Black Cadillac SD picking you up alone and taking you straight to work probably wouldn't be much cheaper than owning your own in terms of direct cost, but you'd still save the maintenance, insurance, and storage. If you're willing to travel with 2-3 others to similar destinations then the cost would drop a lot. Ride in in a minivan with 6-7 people and the cost drops even more, with further sliding scales for getting priority drop off or willing to go last.

Once the SD car is a pure utility then it's all about how tailored you can make the service model. If every single one of the 80-100 million commuters in the US absolutely demanded a solitary, direct trip back and forth to work at peak time then the savings over private ownership certainly slims down, but even then it's still there. Get a good number of people to choose lower cost sharing and the savings really add up.