- You can’t tell from the camera how visible she was peripherally. But based on posts upthread from someone who knows the spot, it’s not as inky lack as the video mak s it seem.
- Swerving is an option, as well as braking.
I’ve not seen the Tesla Model X crash, in which a Model X operating in ‘autopilot’ mode drove into the end of a concrete lane divider, mentioned on this board yet. The crash and resultant fire killed the driver. Granted this was not a fully autonomous vehicle, it does illustrate some of the difficulties with full automation.
A couple of Tesla owners have attempted to explain the behavior of the autopilot as following poor lane marking on the pavement:
[ul]This video shows a lane split, in which the autopilot begins following the wrong line and could potentially crash into a barrier.
[/ul][ul]Another similar video at exact crash location.
[/ul]
There are a lot of roads in the USA, especially in rural areas, that have poor, or are lacking markings altogether. It is concerning that the autopilot technology was allegedly relying on the lane markings. The lane markings in the video, while poor, are not at all out of the ordinary.
This is scary and it’s why I agree that fully autonomous cars are still a long way from reality.
I am further very disturbed by Tesla’s attitude expressed in their release. It feels like it is deflecting blame away from the autopilot system, because the driver should have been paying attention better, The lanes were poorly marked, and that the energy absorbing barrier had been damaged previously, (all true) but it seems clear that the autopilot system played a critical role in that crash.
Clearly, human drivers had hit the same divider in the past, but still, this human would have easily avoided driving into the end that barrier.
The same autopilot system that per your own cite gave multiple warnings to the driver that were ignored…
Are you similarly shocked when your fuel gauge reads empty and the engine stops.
If that is considered poor lane marking that could result in a crash, we will never have autonomous cars in Quebec.
The highway I commute on has no lane marking anywhere near that nice.
So surely not decades away for self driving cars.
Yeah, let’s see if they actually work without human drivers.
I’ll stick with my prediction from post #65: they won’t be ready for prime time before 2024 at the earliest.
It is a pretty big step but you know, caveats apply – it’s a geofenced taxi service in a sunny, Southwestern location. Even without engineers in the car I imagine it will be pretty tightly controlled. The key takeaway from that article is that consumer sales are still a dream away.
No, I would not be shocked if my fuel gauge read empty and the engine stopped, because it would have worked exactly as predicted.
I agree the dead driver shares some responsibility, but Tesla’s total lack of willingness to share any responsibility at all is despicable. By their own admission, Tesla’s autopilot system does not perform as expected in some routine situations and requires driver intervention. Tesla’s claimed belief that the general public would be able to understand that fact and to be able to responsibly operate such a vehicle is disingenuous.
Softbank and GM are investing a combined $3.35 billion into Cruise Automation to commercialize self-driving cars in 2019.
GM is already building its self-driving cars on a full-scale assembly line.
Actually, what the link says is, “The SoftBank investment will provide the necessary capital to “reach commercialization at scale in 2019,” according to GM.” Doesn’t say they’ll do it, just that they’ve got the capital for it, if it can be done. Gotta read these things carefully.
To quote your link, “[GM] now builds prototype autonomous Chevrolet Bolt EV electric cars on the same Michigan assembly line that customer cars are built on.” Whether or not you intended it, you made it sound like GM had a full-scale assembly line dedicated to production of self-driving cars.
Now I’m tired and cranky. 
(Bolding in both quotes is mine.)
ETA: Whatever rolls off that assembly line in 2019, it won’t be a fully autonomous vehicle. It’ll either be something like a somewhat improved version of Tesla’s Autopilot, or it’ll just be regular old cars, or it’ll be a prototype that’s not for sale.
While these are not cars that you or I can go and buy, GM’s plan is to have fully autonomous vehicles on the road next year. Another cite -
Now it may be that these vehicles (no steering wheel) intended for ride sharing will have a specific area that they will stay within. But they do intend on making them at fleet numbers by end of 2019. And that’s a GM 2019, not a Musk one.
I do wonder if they will pass through some of the technology to Bolts with steering wheels sold to general consumers.
So IF Waymo does indeed roll out some of its 62,000 decked out Chrysler as a ride hailing service (plan is to start in Phoenix and then San Francisco next), and GM does indeed roll out its own (by itself or in partnership) with its decked out no steering wheel Bolts by Cruze, both by end of 2019, does that count?
Not yet vehicles consumers can buy, and not yet for everywhere in all climates. Restricted well-mapped areas (defined metro inclusive of airports) covered by the services in very moderate climates very likely. Not snow. But not just an airport shuttle or delivery route either. And no drivers to “take over”.
How long, realistically, from those releases, assuming they happen and perform well, to consumer vehicles able to drive in Chicago snow?
That I couldn’t tell you. But if they can work in Phoenix, they’ll eventually manage Chicago and Pittsburgh, it’ll just be a matter of time.
But this does make for an interesting conundrum: it’s one thing to make prototypes that consumers can’t buy that (a) are ready for prime time in the Sun Belt only, and (b) don’t have any steering wheels.
But to actually sell them, GM would have to choose between (a) putting the steering wheel back in, and making clear that the driver has to take over if there is precipitation on the ground, or (b) waiting until their AVs can manage Chicago snow before selling the first unit.
After all, they can’t exactly sell a car with the qualification that it can only be used in certain parts of the country.
Actually the market they are most interested in is, I think, not consumers owning vehicles, but that ride hailing fleet application. And I am getting the sense that they are less of the mind to sell these vehicles to companies that do the ride hailing service than they are of one to be that service with vehicles they produce. That’s the urgency to get there as soon as possible. They want to disrupt the disruptor and displace Uber and Lyft.
Plenty of market to capture where there aint as much snow while the technology continues to improve.
That said, they’d do fine with the technology in a steering wheel installed Bolt that identifies clearly and with reasonable notice when it can and cannot do the driving for you … even in Chicago the days that the roads are not clear enough are fairly few, and it may be that for a route travelled before as its regular commuting route the vehicle could use other markers to know where the lane markers are supposed to be.
Interpreting that sentence any other way than that Softbank gave GM a bunch of money so that GM would be able to execute its plan and intention to commercialize self-driving cars in 2019 is tortured reading to support your denialism. You seem to be willfully reading the clause “necessary capital to reach commercialization” as being only some colorful, abstract quantity of money that says nothing about how the money will actually be used. Under your reading, the article would be equally informative if it had said that “The SoftBank investment will provide the necessary capital for GM to acquire 100 million wheels of Shullsberg Creamery Baby Swiss Cheese in 2019.”
I don’t know if they will accomplish it but even if GM is a year late, their accomplishments will contradict the idea that self-driving cars are decades away.
That’s neither what I said and nor what I intended. Assembly-line manufacture means that GM isn’t just hand-building prototypes but that they have figured out how to produce them in an efficient way on one of their assembly lines. Since that assembly line won’t be in the same configuration in decades, it suggests that GM doesn’t intend to wait decades to begin producing self-driving cars at scale on an assembly line. It also means that GM can crank out high volumes of them quickly by substituting for the production of other models on the same line. That’s how the modern mass manufacture of automobiles works. This suggests that GM’s plan is to begin mass production of self-driving cars before this assembly line is entirely reconfigured and it is consistent with the idea that they will do it next year.
They said they intend to commercialize self-driving cars. That means use them in commerce. That doesn’t necessarily mean selling the cars to the public but it does mean that they won’t be just prototypes. I suspect that it means using self-driving cars in a car service or as delivery vehicles. In fact, GM previously announced that they would operate a self-driving car service in 2019. Their announcement of getting capital from Softbank is consistent with sticking to that timeline.
That’s the key thing for me. Automakers have been saying for several years that we might have self-driving cars by 2022 or so. Unlike shady flying car makers who have been saying for decades that flying cars are just five years away, automakers seem to be making announcements with dates and sticking fairly closely to schedules that make sense if they plan to actually start selling self-driving cars by those dates. They are sticking with calendar dates even as those dates keep getting closer. Since these automakers are public companies, they are responsible for not lying to their investors about when these plans might come to fruition. Their statements are forward-looking statements so they aren’t to be taken as gospel but, at the time the automakers (and Alphabet) make these statements, they must believe them. I’m beginning to believe them too.
Self-driving cars only drive where they are programmed to drive and that could very well mean a car that only works in certain parts of the country. That’s fine for a local car sharing service or a local delivery service. I anticipate that the earliest self-driving cars will be ring-fenced to places where the weather is predictable and the road-mapping is sufficiently detailed. So, the earliest self driving cars will work only in a particular metro area. Later ones will work in any of those metro areas, some new metro areas, and the highways that join them. Eventually, self-driving cars will work anywhere in America. Somewhere in between, automakers will solve the weather problem well enough to make them a viable replacement for ordinary cars.
Another option; when a self-driving car is unable to navigate on its own, hand over control to a remote human operator.
(hijack)
As a motorcyclist, I am worried that some selfish, lazy sonofabitch “driver” using autopilot is going to plow into the back of me while I’m stopped to make a left turn. If it will do it to a pedestrian walking a bike, it will do it to me.
As a bicyclist I am more scared every freakin’ day that some distracted sonofabitch driver is going to plow into me, at any moment … which a true L4 autonomous vehicle would not do. FWIW the Uber crash happened to no small as a result of a decision Uber made to NOT trust the vehicle to autobrake in an emergency but to require the (distractable) human to do so - they had the emergency autobraking deliberately turned off. That was a crazy choice.
I think the progress will be faster than Tired and Cranky predicts.
I don’t believe that a car that can go absolutely everywhere a human can go now is coming any time soon. I think “autonomous cars” will exist soon, but I very much see them saying “Uh, Bob, wake up! Switching this back to you” a LOT.
Your whole post was excellent, T&C, but I want to focus on this part.
I don’t think most people understand how fast an AI can learn something. Once an AI begins a task and has a foothold, it doesn’t take the AI long to excel at that task.
AlphaGo Zero taught itself to play Go:
The folks at OpenAI had one train itself to play DOTA 2, one of the most popular player-vs.-player games known as MOBAs. With two weeks of training, OpenAI was able to crush the #1 player in the world so badly that after just 2 matches, he declined to play again, saying that he could not win. And that was with OpenAI being trained by playing against itself., not against humans.
Alpha Zero became the best chess player in the world in just four hours of self-play.
In the scenario you describe, T&C, assuming that the AI in each car is actually a hive mind, the AI would learn to drive in the known areas within days; expansion of that area would increase rapidly.