Slashdot Mirror
Waymo Self-driving Cars Are Having Problems Turning Around Corners (siliconangle.com)
Alphabet's Waymo has long been regarded as the leader in autonomous vehicle development and technology, but all might not be as well as it seems at the company, according to a report published Tuesday. From a report: The Information quoted a number of unnamed Waymo insiders who claim the vehicles being used in the Arizona ride-hailing test have numerous problems. The test, which launched in November, is meant to be converted to a full commercial service later this year. The report claimed that the autonomous Chrysler Pacifica struggles to handle a number of driving tasks and even goes as far as annoying human drivers around them. Top among the problems is an apparent issue with turning left. "The Waymo vans have trouble with many unprotected left turns and with merging into heavy traffic in the Phoenix area, especially on highways," the report noted. "Sometimes, the vans don't understand basic road features, such as metered red and green lights that regulate the pace of cars merging onto freeways." If having problems turning left isn't bad enough, they also apparently on occasion have problems turning right. One woman claimed that she almost hit a Waymo vehicle as it suddenly stopped while trying to make a right turn.
Luckily she'd looked up from her phone just in time.
I really cannot fathom these monstrosities on the same road as myself. Time and time again it has been proven that a human driver is needed to intervene to keep these 'auto'mobiles in check. Yet, each one of these companies claims they'll be fully autonomous within a year or so, and each time it gets delayed again and again. If there's any indicator where we're at, just take a look at the newly released chat with the former Telsa worker. When will Silicon Valley and its ilk stop spreading false hopes and flat out lies to appease investors, it's starting to look like Trumpgate part 2. It is most likely at least a decade or 5 away from reality, maybe more so for the streets of London. Many more will die in the same likes as in the Tempe, AZ accident. It's going to take a revolutionary AI, one that hasn't been built yet to truly make this a reality. Robots are def not here to take our jobs or take over the world. If anything, it will create even more work for us humans, assuming we don't get ran over in the process...
It sounds like the car is too meek. Waiting for proper sized gaps in traffic including safety margins, before turning left / right / merging etc.
Instead they need to be more assertive, like Teslas. Grab any gap no matter how small, and expect other cars to slow down or swerve away. If they don't, blame the other driver, accuse them of being pedos while sobbing to the NYT.
More like Tesla.
When I was in high school, my sister had a friend who was deathly afraid of turning left from one busy street to another. She just didn't get the whole, "inch out until the light turns yellow, and you're sure oncoming traffic is gonna stop, and then complete your turn" thing. So, swear to god, she used to make three right hand turns instead. She drover her father's old '70 Buick Electra 225 4-door and that thing was like an aircraft carrier. But it had the first electric seats I ever saw and had the bucket seats instead of a bench in the front, which I though was cool.
In summary, as long as you can make a sufficient number of right-hand turns, you can get away without hanging a Louie.
You are welcome on my lawn.
OK ... So we have several problems
First the Waymo software is likely a bit buggy. No surprise there. It'll take several years to work through that Wait til they encounter some of the blinking red and yellow arrows recently installed on traffic signals around here. I don't have the slightest idea what they really mean. Neither does anyone else.. Neither, I'll bet, will Waymo. On top of which at some times on some days, the sensors trying to read the signals will be looking directly into the sun.
Second, the Waymo cars try to drive safely and legally. Whereas human drivers generally try to drive as quickly as possible without being delayed by accidents or police traffic stops.
Third, I expect, is that autonomous vehicles in general are likely going to have trouble with some forms of bad weather -- especially heavy snow which humans who like to stay out o0f ditches handle by driving quite slowly and keeping moving. This is likely not going to be apparent in testing in Sunnyvale or Phoenix.
You can't see ANYTHING from a car, You've got to get out of the goddamned contraption and walk...Edward Abbey
You may say I am a dreamer,
But I'm not the only one
I hope some day you'll join us
Then I will have won your gullible $
Woo Hoo;)
autonomous vehicles absolutely require infrastructure built into the roadways to support them, while those same roadways physically segregate self-driving vehicles from human-driven ones, pedestrians, bicycles, and others.
I live a few blocks from the Waymo/X building in Mountain View, and see about 5-10 of their Pacifica minivans every day on my commute.
Waymo cars tend to be slightly cautious and very defensive in traffic, but they behave well and are very clear and communicative drivers.
Certainly I see more idiot human drivers on the same commute, especially since texting and driving seemingly became required for driving.
I also see a number of other brands around here but the Waymo integration looks much slicker and less like a science experiment. Theyâ(TM)ve also been doing this since 2009, itâ(TM)s not like they just got started on real roads yesterday.
The Apple cars are maybe the most excessively equipped (they have a fleet of Lexus SUVs with a ton of equipment on the roof).
Iâ(TM)ve almost gotten into an accident with a drive.ai car once and I saw another (Toyota I think?) slam the breaks seemingly for no reason and a Mercedes (donâ(TM)t remember the brand) that refused to move and the driver clearly took over. The GM Cruise project is also extremely common to see on the street of San Francisco and they seem to behave quite well too.
I was never a fan of the tiny bubble cars Google/Waymo drove up and down the streets here, mostly because they were too slow and used to hold up traffic a bit. But those cars literally had no steering wheel and was an interesting project. Iâ(TM)ve seen the blue Waymo semi trailers around quite a bit too and they are impressive.
Anyway, based on my own experience coexisting with self driving cars in my city for the last 4-5 years since moving here I call total bullshit on this article.
Also i wish people would stop using the botched Uber project as a reference for safety. Those idiots did a lot of damage to the future of self driving cars. Itâ(TM)s amazing that they disabled the built in safety features of the Volvo.
I do own a Tesla and drive it in autopilot on the freeway and in stop/go traffic. I canâ(TM)t imagine having to suffer my commute without it. I arrive more rested and relaxed to work, and feel safer as a result going home after a long day when iâ(TM)m tired and less focused. Roadtrips are also more fun now that I can take longer glances at scenery and I find that I donâ(TM)t feel nearly as tired when I arrive.
that with more debugging, they're hoping to turn the corner on the problems.
Fascism: An authoritarian and nationalistic right-wing system of government and social organization. See also: NAZI's
It is not going to happen on regular roads as we know them. Instead some big corporation is going to build a new city (possibly around a new campus) where regular cars will be banned and all trafic will be autonomous and roads will be smart as well with sensors, broadcasts, and what not. It will be so much simpler (for the AI) and so much more convenient for the humans. And once the benefits are obvious, other cities will follow suit. Building a city from scratch was Walt Disney's dream btw.
The ideal solution would be to redesign the road system so that autonomous vehicles could happily work however that would require trillions of dollars
You may as well bring back tram systems
”One woman claimed that she almost hit a Waymo vehicle as it suddenly stopped while trying to make a right turn.”
If you almost hit someone because they stopped suddenly... that’s on you, not the other driver.
Don’t drive like an idiot.
#DeleteChrome
I'm not an ambi-turner. It's a problem I had since I was a baby. I can't turn left.
One woman claimed that she almost hit a Waymo vehicle as it suddenly stopped
Then don't tailgate. Idiot.
According to Top Gear, going around corners is a problem with US car design as they are only able to go blisteringly fast in a straight line :)
"The hands that help are better far than lips that pray." - Robert Ingersoll (1833-1899)
I'd like to hear Waymo's side of the story as I could imagine that the vehicle may have stopped during a right turn because it detected a hazard that was real (maybe a child running towards the road) or not real (paperbag flying towards the road). I also find the wording "the vans don't understand basic road features, such as metered red and green lights that regulate the pace of cars merging onto freeways" strange. Surely metered lights are not a basic road feature but something quite rare. I'm not saying that Waymo should not be able to handle those (surely they should!) but it does not seem to be a major failure either.
Ever heard of a curve, a fork, or a roundabout, dummy?
talking about LSD you don't want to be do 35-40 MPH on that no when it's open do 55+ MPH
responsible till they start cutting costs like uber and uber killed somebody
Oh brother. The difference is one out of a thousand humans may have trouble, *every* Waymo van will have trouble. The fact that this needs to be explained over and over is getting very tiring.
Laws are rules for the court, but merely a bottom bar to hit for life. Think beyond laws in your actions always.
Sadly, that was quite true up 'til 1986 when Ford released a well built, well engineered family sedan -- the Taurus. The reason is simple. The US auto industry was firmly centered in Detroit, Michigan. Detroit is built on an old lake bed, is flat as a billiard table, and the roads are pretty much all dead straight. There are some minor hills in the Western suburbs. But to experience a road that requires the driver to make an actual turn, you need to drive most of the way to Ann Arbor about 60 km to the West.
I genuinely don't think the designers in MoTown had the slightest idea what roads looked like on the coasts or in the mountain West.
You can't see ANYTHING from a car, You've got to get out of the goddamned contraption and walk...Edward Abbey
"Waymo Self-driving Cars Are Having Problems Turning Around Corners "
According to TFA, it should be "Waymo Self-driving Cars Are More Risk-Averse than Human Drivers."
According to Top Gear, going around corners is a problem with US car design as they are only able to go blisteringly fast in a straight line :)
There actually is a lot of truth to that. Compared to European cars in the Clarkson/Hammond/May era, many US cars had worse handling than a river barge. Up until this generation, the Ford Mustang had a live rear axle which really limited it's turning ability, it understeered more than my FWD Honda Integra, right up until the moment it oversteered (and not in a good way like my Nissan 200sx). At least when GM re-launched the Camaro, instead of bolting bits of Detroit together, they went to GM Holden Australia who knew about this thing called "multi-link suspension", even then the whole thing was based on a boat called the Holden Commordore.
I'm not sure if the current gen Mustang is any better as I haven't driven one but European cars are lighter and usually handle much better (weight is the enemy of good driving dynamics).
Calling someone a "hater" only means you can not rationally rebut their argument.
Sadly, that was quite true up 'til 1986 when Ford released a well built, well engineered family sedan -- the Taurus
LOL, get the fuck out of here... that was a joke, right?? If not... here's some history (hint: the Taurus does not have an impressive past):
: In 1983, the Audi 100 (aka "5000" in the U.S.) won the European Car of the Year award. The design was so revolutionary, even compared to the best-engineered vehicle currently available at the time (Benz's S-Class) that Ford panicked, launched an emergency development project to design an imitation Audi 100, copying the general looks and cab-forward design of the Audi (the 100/200/500 has a .30cd) but little else. Worse, since Ford didn't even bother attempting to duplicate the Audi's most effective available feature, all-wheel drive (to be fair to Ford, they knew they stood no chance), the least Ford could've done was equipped the Taurus with a proper longitudinal RWD drivetrain like a Benz, Beemer or water-cooled Porsche) for better adhesion and front-to-rear weight-balance... but no; the Taurus was designed to be as cheap to manufacture as possible... and considering how many of 'em you still see on the roads, it shows.
Somewhere in a parallel dimension exists an '87 factory Taurus with a port-injected 302 and a 5spd (rear) transaxle...
Chrysler minivans have always had a huge turning radius.
It's simply evidence that the hype for this tech is well in excess of reality.
Boing Boing had a link to the reporter who originally broke the story which actually has useful information - https://twitter.com/amir/statu...
They're having waymo problems than they expected, then? Now I understand where the name "waymo" comes from.
I've abandoned my search for truth; now I'm just looking for some useful delusions.
Waymo reported that they drove something like 37,000 miles without a human interaction in November 2017. How does this add up, now knowing that they can't even navigate a normal intersection properly?
Laws are rules for the court, but merely a bottom bar to hit for life. Think beyond laws in your actions always.
"The Waymo vans have trouble with many unprotected left turns"
I've said it before and i'll say it again. If they're having trouble making unprotected left turns, perhaps you should stop having Google Maps direct them to make unprotected left turns? And then PLEASE PROVIDE THAT UPDATE TO ME AS WELL!
Google Maps seems just fine for trips of moderate distance, but i've lost track of the number of times on shorter trips it has directed me to take side streets and then suggested that i make a left turn at an uncontrolled intersection onto a four+ lane road in Los Angeles in the middle of rush hour.
This Space Intentionally Left Blank
Yes, but every Waymo van will also stop having trouble when the issue is resolved as well. The humans aren't getting any better. For myself, I've never even heard of metered lights to control pacing other than a yellow to indicate hazard or slow.
It is important to Turn Left.
I'm a good cook. I'm a fantastic eater. - Steven Brust
We don't have those metered lights here (god knows we could use them, nobody knows how to zipper merge), but my self adaptive driving software (my brain) can instantly figure out how to do them.
It can also deal with other oddities, such as an emergency vehicle that requries me to move off of the roadway to get past, construction workers operating stop-and-slow signs. And potholes, I think Uber and Argo paid to have the road outside my office repaved instead of coding their systems to avoid the potholes. I guess there is some good to the limited design.
Will AI be able to be more adaptive than humans? Probably, some day. Throw enough processing power at a problem and you can do speculative prediction and investigate multiple contingencies at once. But that is not this day. I see self-driving cars every day (my office is near both Uber and Argo's testing facilities) I have never once seen a self-driving car actually dealing with an unusual condition. Every time there is a stopped vehicle on the road, a closure, or some other condition, the driver has been in control.
I'm a good cook. I'm a fantastic eater. - Steven Brust
The metered lights are not uncommon, they're used all over the country for freeway ramps. To avoid excessive congestion on the freeway during rush hour, one is not allowed to proceed past the metered point until one has one's own green light. The lights are normally red, momentarily turn green for one car to proceed, then revert to red.
That surprises me is that these are a challenge for the cars. The control signal lights are not placed in exactly the same place everywhere (some older freeways place them above, while most place them to the left of the left lane or the right of the right lane) but they're time-activated and they function the same way everywhere. Worst case Waymo should be able to map-out where they're at and should also be able to establish rules that govern when they're in-effect, how the light in this zone work, and where to look to confirm no-go versus go.
Do not look into laser with remaining eye.
Are you maybe thinking of the Ford Taunus? I can see Ford Europe being concerned with European competition (and thus Ford Europe producing the Sierra), but in the US market the Europeans weren't big players - especially the notorious Audi 5000. There's no doubt that it was inspired by European styling, but Ford already was making the Sierra in Europe - which I think predates the Audi 5000 slightly. Ford was indeed in a panic, but it was more driven by Chrysler's K-car, GM's A-bodies, and the Japanese imports, most with front-wheel drive.
The styling is very consistent with other, preexisting Ford cars:
1982 Ford Sierra
1984 Ford Mustang
1986 Ford Taurus
W..w..W - Willy Waterloo washes Warren Wiggins who is washing Waldo Woo.
I also avoid left turns across double lanes if there's no traffic light, no center lane, and visibility is questionable. I'll still do it if in a hurry or traffic is otherwise light, but those are hair-raising. There's a lot of jerks on the road who drive too fast, are texting, etc, creating risks. If something goes wrong, there's not a lot of ways out.
I suppose one gets better over time when they do such turns often, just like any task, and by mostly avoiding them I'm not getting enough practice. But I'm at the age where I say, screwit, my approach has kept me alive for decades. A bot does have the ability to "look" both ways at once, though, so it has an advantage over a lone driver.
Table-ized A.I.
The ideal solution would be to redesign the road system ...
Given that there's nearly 9 million lane-miles of roads in the US alone, I'm going to come down firmly on the side of "that's not the ideal solution".
My previous comment aside, I actually see the concept working for personal flying vehicles before cars. You can just build the infrastructure and interconnection much more easily. Most of what makes piloting difficult for humans is trivial for traditional algorithms and the road conditions and human signaling infrastructure along with complex maze of state and city codes don't exist and or are less of an obstacle because the FAA can preempt them. The exceptions already transmit signals on RF.
blinking arrows are a new thing and it's a kind of equivalent
These issues trouble human drivers too.
Only the ones we shouldn't have given licenses to.
Up until this generation, the Ford Mustang had a live rear axle
Didn't the Mustang gain independent rear suspension several generations back... or was it just Cobra variants? The Camaro, on the other hand, did just recently lose the compound live axle...
but European cars are lighter
Definitely not accurate...
Its probably a problem differentiating a normal light from a metered light, and behaving as if it will remain green for more than a split second. A human can easily read the sign that says 1 car per green. A computer can't.
I still have more fans than freaks. WTF is wrong with you people?
My scenario wasn't a branch staying on the road, but just temporarily bending in the way of the road so that the driver, human or bot, temporarily pauses until they understand what's happening.
But in general, the bot-driver will probably lack "common sense" in many situations, and present "odd" behavior, at least from a human perspective.
For example, it may slam on the breaks because a plastic bag is tumbling across the road. Most human drivers would probably slow down a bit but not outright stop because the humans know a plastic bag poses no risk.
Bot programmers may add a "plastic bag" subroutine to mirror such behavior, but some new situation will come up they didn't anticipate. Over time, the bot-car will get better, though, as more sub-cases are added to its "experience bank".
I don't believe lack of common sense is a reason to give up on bot-cars, for they do better than humans on other metrics. They don't have to be perfect, just better than human drivers on average. They just "fail differently", to mis-borrow Apple's old motto.
Table-ized A.I.
Cars need to have their safety and guidance systems onboard and not rely on a possibly broken or misconfigured signal from outside the vehicle.
You don't get off the hook for causing an accident just because the traffic lights were not working.
If these were trains with very limited options, what you are saying would work fine. With unlimited possibilities and having to contend with flawed human drivers, the AI will need to be adaptable.
I'm a good cook. I'm a fantastic eater. - Steven Brust
"Cars need to have their safety and guidance systems onboard and not rely on a possibly broken or misconfigured signal from outside the vehicle."
Why not have both? AI learns after all, so long as there are people driving through that space when the signal is working your car will have gained what their car learned from that experience, even if you are approaching it for the first time with a down signal. It isn't as if Aircraft don't rely on external signals.
"You don't get off the hook for causing an accident just because the traffic lights were not working."
No, you don't get off the hook when using your cell phone or drinking either but people do. What I have noticed with self driving vehicles is people seem to expect them to perform flawlessly or be deemed unsafe even though thousands of human lives are lost due to how flawed human drivers are each year. If the self driving cars result in the same or fewer casualties (without any need to fail in the same places the humans do) they are a success.
"With unlimited possibilities and having to contend with flawed human drivers, the AI will need to be adaptable."
The AI does need to get better, I'm not disputing that at all. I'm just suggesting there are ways to improve some of these situations and in the case of traffic light signaling it's been shown that automating the humans speed through the lights to avoid full stops would be much more efficient anyway.
Ideal and realistic aren't always the same thing. It would be the ideal solution, it will also never happen. Any plan that requires redesigning the roads is a non starter as nobody is going to "fix" every back road in every country. Realistically suggestions like this are just an admission of how terribly far away we still are from actual self driving cars.
Didn't the Mustang gain independent rear suspension several generations back... or was it just Cobra variants? The Camaro, on the other hand, did just recently lose the compound live axle...
It was only the Cobra, and it was only for a couple of years. Then they dropped IRS again for some years, and then they brought out the new Mustang with IRS. All models in the latest generation have it.
"You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
I actually see the concept working for personal flying vehicles before cars.
Personal flying vehicles have a much simpler set of problems to deal with, atmospheric conditions aside. Hobbyist drones can already handle wind buffeting and the like, so the only part remaining to be solved is traffic control. You can't reasonably expect aircraft to always be able to sense one another and react appropriately, so there must be some kind of central authority which manages it. And the load will be too great for humans to cope with, so it will have to be automated.
"You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
"as there are people driving through that space when the signal is working your car will have gained what their car learned from that experience"
It doesn't follow. An 'AI' may learn how to deal with that specific situation, but then have to start all over when a slightly different situation comes up.
You mean like when Alexa learns how to understand a new speaker? Oh right, she does that on the fly. AI is not clever programming to appear smart anymore, AI are shifting and self-programming networks that learn abstract patterns more or less the same way the brain works.
We already have auto-pilot. In visible conditions aircraft can detect each other as well as cars. I think the big advantage of this arena is that you don't have to design the way we have with cars and current flight systems where each vehicle is an island. The line of sight is fantastic for this use case, GPS, altimeters, etc can be used to get a very good idea of location right off, when vehicles are in proximity they can employ shorter range mesh networks and flocking algorithms and actually improve their collective ability to buffer wind and other atmospheric disruption. I suspect in normal conditions, and with quite a bit of polish, you'd be able to produce vehicles that can pack quite densely in three dimensional space. You don't exactly see a lot of bird-on-bird collisions even with groups of hundreds or thousands.
We can do most of that with traditional algorithms and light AI. But if we got serious we could do a better job on this AI than you might think. We are able to fully simulate a mouse brain using a super computer, not your typical synthetic neuron made to steal the concept of how one works but an actual replica of the known physical function and have shown mouse like behavior as a result (there are previous papers published on this). I have a strong feeling we could do the same with a bird brain and use it to train up a more computationally efficient Artificial Neural Net instances with flocks flying in a simulator.
The atmospheric conditions are largely already handled in auto-pilot systems but personal vehicles would be a different class because the weight and design of the vehicle impact these things a lot. Hobbiest drones are a worst case scenario. Personal vehicles would have to be somewhere between.
Google has this thing called "Maps" that has basically the entire country and large swaths of the world mapped-out. It would not be a challenge to use data from this to identify where these kinds of controls are.
Do not look into laser with remaining eye.
It would be a huge challenge. First off, lights change. New ones go up. Secondly, GPS is not always accurate. The bounce around skyscrapers is particularly bad, I can be off several blocks randomly. Its very easy to be off by the hundred feet difference between a freeway onramp and an intersection light nearby. Thirdly- maps doesn't have this info. Fourthly- there's multiple types of metered lights. Some require not only 1 car per green, but alternating lanes.
Yeah, this kind of stuff isn't going to happen in the next few years. More like next few decades.
I still have more fans than freaks. WTF is wrong with you people?
First, I agree cars need to talk to each other and to fixed roadside beacons or stations if there is any chance of them becoming safe and reliable.
"You mean like when Alexa learns how to understand a new speaker?"
No, I mean that AIs can't competently merge onto highways or negotiate left turns. The complexity is currently way beyond their capabilities.
"AI is not clever programming to appear smart anymore, AI are shifting and self-programming networks that learn abstract patterns more or less the same way the brain works."
There is a very very big difference between how 'neural nets' and the brain process information, so AIs are not shifting from 'clever programming' to functioning 'more or less the same way the brain' does.