"Hi, this is your car. I've discovered that the road is icy, and am currently in the middle of a 360. Would you like to (a) take over, (b) call an ambulance, or (c) retry?"
Too late! You've crashed.
You have no faith in AI. The car will always be better at driving than the human. It would be much better at handling a 360 than a human ever could. Actually humans are very bad at handling 360s unless you are a pro.
In order for any AI to handle any situation, it must first be programmed to do so. If that situation was an unknown, you get undefined behavior. Undefined behavior in a car traveling down the road is not a good thing to place a bet on... especially your life.
This is why majority of sane driverless vehicle talk has concluded a human driver must always be in the driver seat, ready to take control. There's an impossible volume of failure modes a driverless car must contend with, everything from someone pointing an infrared laser at the distance sensors, to complete power failure, to a sudden rockslide, a bird darting in front of the car, lines missing from a windy and hilly road, etc.
In normal, daily mundane driving - driverless cars are expected to excel. Sitting in traffic is mind numbing, and mistakes happen when driver's minds wonder. In "active" driving situations, a human brain will almost always excel.
For many human drivers on the roads right now, the response to a sudden weird situation is either going to be "brake heavily" or "swerve dangerously."
You don't have to get "undefined behavior" in response to an unknown situation. The AI can simply default to braking to a stop in the absence of any better programmed alternative. This will already put it on par with most drivers. Better, in fact, because the AI can begin braking instantly.
I personally find that the majority of driverless vehicle talk vastly overestimates the average human driver, and demands unreasonable perfection from the computers.
Next up is per-wheel braking. Not just stability control, but actual directed thrust. The computer can control each wheel independently. Humans just have one brake pedal. Four wheel steering might even make a comeback.
When people say that it doesn't work, I really think they're just saying that they can't stand having other people get ahead by breaking the rules, and they'd rather drive more dangerously than let that happen.
Or a combination of both, potentially causing the tires to slip and the car to leave the roadway!
Would add also something that requires an actual driver to take the wheel periodically to not get lazy and forget what it's like to drive (as apparently what has happened with some airline pilots?)
The list of situations a car has to react to is not infinite. You can program a reasonable response for a broader list of situations than you could for a human. And really, a human would fare very badly in your situations so a computer could hardly do worse.
Not exactly. The situations we're discussing are more-or-less unique, having their own environment variables to contend with.
Not to mention, there are scenarios where a driverless car will never be able to operate (such as off roads, or back-country/non-mapped areas, or when road lines are covered by snow or just not there due to road construction, etc...).
Roads are anything but a constant. I cannot foresee any future where someone can get into a vehicle, in the back seat, tell the vehicle where to go, and then promptly take a nap.
Even with high-grade multi-million dollar vehicles such as military drones, they still require two humans to be at the controls at any time (even though in the air, collision opportunities are almost non-existent). The sensors and computers in a mere $20,000 vehicle aren't going to compete with this stuff.
A human will always have to pay attention to the road and be ready to take control. This aspect almost completely nullifies the desire for a driverless vehicle.
And people who do research in this area have concluded that this doesn't work. People let their attention drift (even if they don't actively perform some other task) and, even if people are more or less paying attention, they haven't been driving and switching modes in that way is difficult.
We're definitely reaching the point with assistive driving systems where there probably needs to be more discussion about how much assistance makes sense if no one is willing to fully turn over control.
No, that's not how AI nor computers work. It's quite trivial to have a default behavior like "brake", the notion that a new situation will confuse the computer is just you watching too many movies.
I don't think this is true yet. We can interpret the incoming data incredibly well in ways the machine vision can't yet match. This is absolutely vital for driving. The difference between an empty trash bag in the road and a toddler in the road is a matter of vital importance, and correctly identifying the difference matters a lot.
Humans area actually capable of being incredibly good at driving. The problem is we are also incredibly unreliable.
I'm sure I've stopped for empty trash bags before because I wasn't sure if they contained something solid. That doesn't mean I failed as a driver. Everything went fine. Same if it's a computer but it might stop more often than a human.
Two engineers are arguing: a mechanical engineer and a software engineer.
The software engineer brags:
If we were the ones making cars, they'd go twice as fast as they do now and they'd use half the fuel!
The mechanical engineer replies:
Yes, but if you'd want to break in case of an emergency you'd get a pop-up saying: Are you sure you want to break? Yes/No/Cancel.
The mechanical engineer says, "It's a loose belt. Let's tighten that and we'll be all set."
The chemical engineer says, "No, it's fuel contamination. We need to drain the fuel, filter it, and then we can be on our way."
The software engineer says, "No, all we have to do is turn the car off, get out, shut all the doors, then get back in and turn the car back on."
The irony is that I now own a car for which "reboot it" is a valid troubleshooting step....
At a talk to an audience of avionics programmers, the speaker at one point asks: "Tell me truthfully; how many of you would be comfortable flying in a plane where your team had done the software?"
Onlt a couple of people put their hands up. The speaker targets someone at the front. "You have that much confidence in your work?"
"I'd be fine," they reply. "If my team wrote the software, the plane would never even leave the ground."
The technological solution may be that there is no (or vanishingly small numbers) situation where emergency transfer is likely to improve outcomes. So you just don't do it. Non-emergency (i.e. I'm going to park now and hand off control to you) are a separate case.
This may plausibly be the best technical solution - how marketable it is, that's another story.
I wonder if even at the really primitive state of SDC if they aren't better (in terms of safety) overall than human drivers.
Object avoidance and control in slippery conditions is a vastly easier problem than a 100% self driving car. So, if the default is a car that safely slows down to a stop while beeping loudly then hand-offs are a non issue.
Remember, the state of the art in human driving still kills 32,000+ people per year in the US and something like 1+ million worldwide. Beating that seems very doable.
EX: You show up at a concert and parking attendants are directing people to an open field. Car goes WTF, stops and hands off.
The idea that humans are any better at handling situations they've not previously encountered is dubious.
Humans also can fail spectacularly. But they also can just try something, realize quickly that it isn't working, and switch to trying something else. An AI could do the same, of course. But humans are better (at the moment at least) at eyeballing the situation and realizing that the current approach isn't working. (Or so it seems to me.)
This might be true in X years but right now human ability to apply existing knowledge to brand new situations far surpasses that of AI
This is still facing regulatory hurdles, obviously.
I've been driving a Tesla with “Autopilot” for five months now. Since then I've probably logged about 10,000 miles of use of “Autopilot” features – lots of two-hour trips between Eastern and Western MA; a MA<->NC road trip and some shorter road trips; and lots of 30-40 minute commutes.
Prior to getting the feature, I was concerned about handoff. I was concerned that it would be difficult to force myself to pay attention during those times that I wasn't directly responsible for the steering and acceleration.
Instead, I've found that it's easy to pay attention, but that I shift to a more strategic mode: scanning further ahead and around for hazards. It's actual easier to maintain situational awareness than when my attention is split between what's going on in the next 10s–1m, and more tactical tasks such as manually tuning my lane position, speed, and follow distance. Driving with automation feels (I told you this was subjective) safer than manual driving.
Just as in manual driving you keep your eye open for certain kinds of situations (passing a pedestrian or car on the shoulder, erratic driver ahead) and are poised to go into action, I'm especially primed to take the wheel when these risk factors add up (as well as other areas specific to this particular vehicle automation: its uncertainty about which way to steer around a small traffic island at the end of a turn; its failure to see stopped – as opposed to slow – cars just after a sharp turn or lane change; its penchant – mostly fixed now – for taking highway exits). And being ready to shift more cognitive resources over to driving is such a natural part of being a driver in any case, that I imagine this is a common experience.
When autopilot is turned off or I'm in a car that doesn't have it, it feels riskier not to have that second set of “eyes” on the road.
There still could be risk compensation: I could be driving faster because I feel safer, for example. I don't really trust anyone's attempts to figure out a priori whether I or the population of current or potential drivers of automated vehicles are subject to this; instead, I hope we see statistics soon.
And – I have a few decades practice driving non-automated vehicles, that I can pretty effortless fall back to. I've swerved around emerging accidents a few times, and handled drivers coming into my lane and other near misses; I'm decades from having to consciously think about where to put my foot to brake. I do wonder about the hand-off to people who are just now getting their drivers licenses and are going to have almost all their experience in automated vehicles is going to work.
We may be entering a (hopefully brief) pair of uncanny valleys of vehicle automation. (1) Where younger drivers don't have enough practice to deal with crises. (This has always been true of new drivers, but they'll no longer be learning at one year of experience per year; and the handoff requirement is new). And (2) the parent poster's concern, that drivers in general will zone out when the car is driving. My anecdotal evidence, above, is that this isn't yet a problem with the state of the art, but that's just a matter of whether this is the year that happens.