What Google is trying to do is an actual autopilot, which is much more difficult, I suppose.
When flying on instruments, air traffic control keeps planes apart. It's up to the pilot to program the autopilot correctly. (There are other collision avoidance systems but planes can fly on instruments without relying on them.)
But in the end they're not really comparable. Collision avoidance for airplanes is a different problem.
As long as they are telling people they have to be alert while using autopilot I don't see a problem with it. The 'common sense' assumption then would be that autopilot = assisted driving rather being an autonomous self-driving car. Therefore Google et al shouldn't be using the term autopilot when they mean autonomous self-driving vehicle.
Eventually as this tech becomes wide-spread the distinction will become common knowledge to people.
This is mostly just semantics... context is everything. The important thing is preventing people from dying. Which means looking beyond marketing material to educate people.
Driving already leads to mind wandering states; it is overwhelmingly likely that the passive aspect of autopilot will lead to mind wandering at even higher rates. Asking a less practiced driver to shift attention from internal to external states very quickly and then make complex judgments is simply not fair. A simple physics and statistics based model will be far more reliable. If it isn't, then Google's strategy of shooting straight to level 4 makes sense.
> The public's perception and 'common sense' understanding of technology is sadly very limited
Common sense is contextual and one of the more complicated aspects of cognition; it depends on the level of detail in the model being used to make inferences. A model's sophistication is dictated by internal preferences and goals. If most people's understanding of a technology is limited, then they're going to be doing what looks like averaging over distinct possibilities to a more informed model.
It doesn't help that if you know nothing about technical uses of the term autopilot but do know something about words (which will be the case for most) then in truth, it is the aviation industry that has misnomered.
It's not a matter of 'educating people', merely so they can use a popular term. It's just plain the wrong thing to do.
Doesn't this support my argument that this is the real meaning of the term? Which makes 'autopilot' distinct from 'self-driving' or 4th gen?
If you as a sailor can understand this then why can't drivers understand autopilot as a glorified cruise control?
The car issues obvious and repetitive warnings about what the technology is capable of and what the driver has to do, and turns off the Autopilot if the driver doesn't pay attention. I wouldn't call that "fine print".
Either I can fully rely on the system to get me from point A to point B, or I have to fully concentrate on driving. People aren't robots - they can't go from half-assed sorta-paying attention to a split-second life-saving reaction.
Why are you assuming that other manufacturers are not working towards full autonomy? I strongly doubt that everyone at Ford, GM, VW, Toyota, and Honda is asleep at the wheel... Especially when their luxury vehicles are all incrementally moving towards autonomy.
They certainly have a lot fewer PR pieces about how amazing their autonomous-but-not-really lane assist is.
Of course not, since only Tesla AutoPilot lets you sleep at the wheel
I think it's a fair call using this term.
By the way, I'm pretty sure completely autonomous parallel parking is already available retail.
I bet if I asked ten people in my office right now, less than half would say confidently that airplane autopilot is a turnkey solution.
"Less than half" is way too much.
Like I see Belgian drivers driving terribly every day and I'm pretty certain that they're much worse than British and German drivers in every way. But the only statistical thing I can go back to is the number of deaths [1], which could include other aspects rather than just stupid driving.
My brain wants to shout out that their's clearly a problem, but that 'clearly' is only on the stretch of road I see. It could be that Belgian drivers are really good everywhere else in Belgium.
[1]: https://en.wikipedia.org/wiki/List_of_countries_by_traffic-related_death_rateit is. Tesla was also beta-testing (last I saw a few months ago anyway) completely autonomous head-in/rear-in parking including finding the spot in a crowded parking lot.
I resent that you're accusing Tesla of intentionally misleading the public on what the feature does.
The car issues very obvious and repetitive warnings to the driver to explain the shortcomings of the technology and what the driver has to do to compensate. If the driver fails to act in accordance with the warnings, the car turns off the feature.
I mean, I dunno. To me, it looks as if Tesla has already gone above and beyond to communicate accurately, and the main reason they're pushing things even further is because of the intense media scrutiny on Tesla accidents (which the incumbent manufacturers probably love).
Whether or not they intend to mislead, in practice the name is misleading, and I think it is a mistake they could easily fix.
For example 'copilot': a competent, capable partner, but the ultimate responsibility still rests with the pilot.
They are both fully aware of the importance of the human awareness implications not expecting continuous input from the human driver and choosing to encourage it.
Humans fundamentally can't stay attentive when their attention has no reactive feedback loop and they're only using the "stay attentive" (even though we know that it's essentially impossible at a psychological level) argument to shield themselves from legal liability.
It's both fair play because they knew in advance (and have since proven) that they can beat the odds against a human driver, and have simultaneously built in their legal defense to mitigate their downside. "We told you to pay attention. Your death is not on our hands."
[1] there are examples where this did not happen, but they're considered a failure.
You can still use an autopilot when flying visually but then it's up to you to watch for traffic. This is using it like cruise control.
Yes, perfectly fine. If Teslas autopilot can rely on external traffic control to provide that information, call it autopilot. As long as it can't don't.
CMU is working on this (seriously).
Some can land, in optimal conditions, but can't taxi.
If you put a plane on autopilot and then go read a book for some hours, most likely nothing bad is going to happen (particularly on IFR, but even on VFR, realistically).
Tesla's autopilot is more advanced than aircraft autopilots, I'd say, yet more dangerous (for now) because of all these darned cars and trucks and pedestrians.
In that sense, "Autopilot" was an unfortunate and overly ambitious name.
Autopilot (before Tesla) is a feature on modern aircraft. If you'll notice, aircraft always have an attentive person behind the wheel (if it's a commercial flight, two).
A plane can fly itself under normal conditions (highways for Tesla) and many can land themselves now (Model S can park / summon feature).
Autopilot couldn't be more apt.
> "oh, crap, there is a truck stopped 20 feet in front of you that I failed to detect at highway speed.
To appreciate your differences in measurement. 5 seconds in a plane is something about a mile away. A high performance car can still stop from 60-0 at 110ft; which is about 1 second. In context of time; it takes at least .7 second for the human to respond.
(The car or terrain monitoring autopilot can react in <.1 s; which means if the car sees the obstacle at 110 feet away it can stop in time, while, if a car pulls out in the highway in front of a human it would take closer to 200ft to stop.)
What happens if the driver expects the car to react upon seeing the obstacle but it fails to do so? Will the driver have enough of the remaining time to react?
The >=.7 second human reaction time (with a planned/known reaction such as brakes or jerking steering wheel) is the very reason why Autopilots are safer.
> TCAS Resolution Advisory will give an attentive pilot a good 5 seconds to react in the worst case
5 seconds in order to avoid violating intruder's airspace, not to avoid actually hitting the intruder. AFAIK there's even then some additional buffer zone.
Also, reading the list of actual TCAS advisories sheds additional light of the substantial differences:
https://en.wikipedia.org/wiki/Traffic_collision_avoidance_sy...
Differences from Tesla / cars:
1. Almost all advisories actually tell the pilot what the problem is and what to do, not just that there's a problem. Tesla Autopilot basically has one advisory: "TA".
2. Almost all of the advisories resolve collisions by making one-dimensional maneuvers. Having an extra dimension allows you to substantially simplify the collision avoidance problem.
3. It's a safe bet (although not assumed) that both planes involved in a TCAS event are receiving advisories and cooperating with ground ATC.
Though I guess Google/Uber are probably getting significantly higher resolution data than Tesla/MobileEye/etc since you can't stream all sensor data back to HQ over LTE.
I think a lot of these articles are driven by the fact that Google has been working on this for so long and yet we have upstarts that are getting cars on the road in a fraction of the time.
It's very possible that Google overestimated the problems with Level 3 Autonomy and the public backlash associated with it and building L3 tech as an on-ramp towards L4 may turn out to be the correct path to market and funding further development effort.
Given that Google doesn't make cars, I don't really see how they would go about commercializing L3, even though it's well within their capacity from an innovation standpoint.
I really don't think they expected the auto industry to respond so aggressively with their own proprietary autonomous driving OSs, determined to shut Google out.
At this stage their only option is to get to six sigma level 4 in a way that leaves everyone else scratching their heads wondering how Google did that. It's going to take a lot longer than Google anticipated.
Now Google could probably release level 4 quite soon in a limited capacity that's more an amusement park ride than a relable, profitable transportation service, but it's probably smarter for them to keep their eye on the prize and keep plugging away at it. They're in too deep to throw in the towel.
Plenty of the startups in this space aren't making cars, but buying cars and installing aftermarket kits. Otto is a great example. Or they could have done what MobileEye did and sell L3 tech to car companies to fund further development.
I don't find the auto response particularly aggressive; they've waited 7 years to even start anything. The barrier to entry in that time has gotten so much lower due to deep learning that geohot can build a car that has some level of autonomy in his garage, and I think those advances were what really caught google flatfooted.
Can Google keep funding a pie in the sky research project for another 20 years, which is roughly the timeline they've said about L4 autonomy.
Yes
My suspicion is that they are way ahead of competitors in terms of data from testing self driving cars and realize there are still so many unsolved problems that true autonomous vehicles aren't going to happen nearly as soon as competitors claim.
For instance, I see self-driving Google cars on a daily basis here in central Austin. There are a ton of zero visibility turns where you have to slowly edge out and hope no one is coming because visibility is blocked by a hedge.
How can self-driving cars handle that scenario without delegating to manual intervention?
[1] This one police radar seems to also nicely work through stone http://www.camero-tech.com/product.php?ID=38
One should not assume that simply because a human can't see through it, doesn't mean a computer with the right kinds of sensors can't.
One possible way: http://web.mit.edu/~velten/www/corner/
Discussing your specific example, I wonder if autonomous vehicles might learn to avoid blinds given enough crash data.
Granted this can happen now with people calling in but they will forget to do so, won't have images to show for backup, etc. Roads can be made safer this way.
Second, it will come when cars finally talk to each other, which will become mandatory in a few years.
It's going to get to a point where they say it's all or nothing, we can't rely on people to stay attentive when the car is doing 90% of the driving.
That is exactly the sort of situation self-driving cars are expected to be perfect, or at least superior, to humans at handling. If they're no better than a human driver or even a bit worse when edge-cases occur than what's the point?
[0]: http://www.mercurynews.com/2015/01/23/chp-tackles-surge-in-s...
That said, I think GP will be proven right in time. Risk compensation[0] is a well-documented phenomenon and this seems highly likely to trigger that sort of behaviour.
One of the biggest mistakes you can make is to believe that the news is an accurate reflection of reality. The opposite is actually true, in the sense that what makes it to be high profile news is something unusual and surprising.
They might get political pressure to consider high attention events instead of the statistics, but they are partly in the business of pushing back against that kind of pressure.
Make the autopilot ring alarm when uncertainty reaches some point and that's it. Tesla already can do it, but doesn't do it for some reason.
HTC/Valve and Occulus have produced a high quality product and then you have the trashy phoneVR which isn't even comparable. Is the prevalence of a substandard VR behind the recent slump in sales? It's fairly probable (with other factors).
Not to mention the youtube video of those two guys sleeping while the car is driving. (can't find a link, sorry)
Personally I'll be Will Smith in the movie "I, Robot" and will never trust them.
This means that Uber can implement 'mostly there' driveless vehicles in a patchwork fashion before everything is perfect.
If Ubers engineers can master a circuit for driverless cars that runs from, say, the airport to downtown and back, but only when the weather is good and traffic is light, they can do it because they've got the human drivers available to pick up any slack in the service.
With Google, if they're only taking passengers partway to their destination or if the system is down 20% of the time, it's just going to piss people off.
Do you mean Tesla? Is quite pointless for Uber to have a car that almost drives itself. They are not car makers trying to sell a better product, they need to get rid of the driver to get any money out of it.
If you are going from the airport to downtown you don't want your car to stop working because it is starting to rain and wait until a human driver comes along.
I think that in that sense Uber is competing directly against Google and not, for example, Tesla. And it looks like Google's technology is more mature, even that it is still not good enough.
For Google is a long time investment, I think that Uber is quite wrong if they think that is a short time investment for them.
Tesla can sell cars with autopilot even when it doesn't work all the time, but they would have very hard time selling autopilot if it only worked on very specific cities or "tracks."
In contrast, Uber can start rolling out autopilot which works very reliably but only on a specific and well-understood "track." They can start phasing it into the market by dispatching autonomous vehicles only when you request a route that they're very confident on. In the medium term, they can dispense with having a driver on such routes at all.
One thing that Uber might try is a self-driving car "delivery" service. People might not be confident enough to ride in a self-driving car yet, but you could have an app deliver a car for you to drive to your destination (maybe w/ driver assist).
For Tesla, the incremental strategy is partial autonomy "anywhere." For Uber, it's probably full autonomy in specific locations. That's a huge advantage because it can seamlessly mesh with their existing app—they can simply start serving a percentage of Uber rides with autonomous vehicles.
Cynically, I think Google is realizing that they need an incremental adoption plan and that's why they're launching their own ridesharing platform.
There are lots of places where it doesn't rain very much.
They could also be running only when the weather forecast says that it will definitely not rain.
If there are problem locations where it is difficult to drive, simply don't run them in those locations!
Also, they could have self driving being opt in for a discount. If there are problems, well you signed HP for it. Don't want to deal with them, then don't take the 90 percent discount!
If it can't get from A to B without a human behind the wheel, it's not autonomous.
How about if it can achieve this 99.9% of the time, but needs help very occasionally? How many nines do you need? Humans are not at 100% overall. There's no clear definition of autonomy.
http://bogbit.com/vancouver-man-sacrifices-self-to-save-wife...
Would you rather have surgery from a human with a 90% success rate but a 0.0001% of a heroic save, or a robot with a 98% success rate, with the bonus that the robot will only be getting better with more adoption.
[0] https://en.wikipedia.org/wiki/List_of_motor_vehicle_deaths_i...
This blog post pretty much sums it up: https://electrek.co/2016/04/11/google-self-driving-car-tesla...
They'll also have the equiv of a waze killer when the model 3 hits and there are a lot more Tesla vehicles on the road. As it stands today, I am literally unable to come to work and back without seeing at least 2-3 Teslas (model S or X) on the road in Chicago. Tesla is on their way to be a bit of a creepy monopolist, but I'd rather them do it than say General Motors.
I think people are handwaving away the true difficulty of door-to-door (as opposed to lane-keeping) driver-less cars that can operate safely without humans being attentive all the time.
I'm pretty sure anyone sane realizes that the current production shipping autopilot hardware is not capable of fully autonomous driving ie: 100% humanless. I also think that no one is handwaving anything as it is a problem that as of yet, has never been solved. However, data is an enabler for this and having more of it means Tesla is uniquely positioned to crack that nut before anyone else. Teslas are collecting data even when AP isn't enabled. That doesn't mean the cameras are off or the radar is off (they aren't). This simple fact gives Tesla a leg up over virtually all of the competition and until Uber has the same or superior tech on all of the uber fleet, they'll be lagging.
So... When the tech is capable of full autonomy, Tesla will be sitting on an absolute treasure trove of fleet data, which they'll be able to exploit.
Human drivers have actively-pointed stereo vision sensors, IMU, and haptic feedback from the steering wheel and pedals. That's it. We know it can be done with only those sensors.
On the other hand one could argue that putting these sub-par systems out there teach us a lot we couldn't learn in a closed environment.
Also as far as I am aware, Google have had their self-driving cars driving on the road for year now.
Cars have always been coevolutionary with infrastructure and, hell, society as a whole.
Why solve a problem for a point in time and its assumptions... the hard way?
By adopting adaptive and coevolutionary timelines, self-driving cars and a society that uses them can "organically" (sorry for the buzzword) adopt them.
And ignores multiple levels of useful intermediate achievements:
stage 1: self-driving on superhighways for the long haul: a wayyyy easier task with much more controlled information, markings, protocols, risks, etc. And with massive payoffs in consumer and business transportation.
stage 2: more extreme weather events, highways besides superhighways, handle commuting and congestion
after that: progressively more local streets and conditions.
But you are correct in that other companies design their cars to drive themselves, and Google's just making an overglorified Google Maps client with collision detection. And the reality that maps will never be accurate actually renders Google's methodology a complete nonstarter for ever being a real product.
Gradually introducing self-driving features lets the competition gradually gain consumer trust and confidence in the technology, as well as giving them more data for future improvements, and it also gives them a revenue stream to fund future development.
When I think of companies that are good at execution, Google isn't on the list.
To me it seems so hard as to be almost impossible to jump directly from no computer assistance to fully autonomous driving, without first moving through (and learning from) all of the semi-autonomous steps in between.
Tesla on regular highways with sleepy consumers behind the wheel.