Slashdot Mirror
Ex-Uber Engineer Claims a Self-Driving Car Drove Him Coast-To-Coast (theguardian.com)
"Anthony Levandowski, the controversial engineer at the heart of a lawsuit between Uber and Waymo, claims to have built an automated car that drove from San Francisco to New York without any human intervention," reports the Guardian. Levandowski told the Guardian that he completed the 3,099-mile journey on October 30th using a modified Toyota Prius, which "used only video cameras, computers and basic digital maps." From the report: Levandowski told the Guardian that, although he was sitting in the driver's seat the entire time, he did not touch the steering wheels or pedals, aside from planned stops to rest and refuel. "If there was nobody in the car, it would have worked," he said. If true, this would be the longest recorded road journey of an autonomous vehicle without a human having to take control. Elon Musk has repeatedly promised, and repeatedly delayed, one of his Tesla cars making a similar journey. A time-lapse video of the drive, released to coincide with the launch of Levandowski's latest startup, Pronto.AI, did not immediately reveal anything to contradict his claim. But Levandowski has little store of trust on which to draw.
I still want my flying car!
A bullet may have your name on it, but artillery is addressed to " Whom It May concern"
This guy is mostly famous for being a big liar and a thief. Not buying it. Also not sure why anyone would care about this.
"Nothing went wrong this time so nothing can go wrong"
This so-called 'engineer' is a danger to himself and others and having admitted what he's done they should find something to prosecute him for.
Just because Uber's car had some software disabled that would have prevented collusion with a pedestrian, how does the mean this guy lacks trust?
Uber's cars were working generally OK - on city streets mind you - until they were pulled, because of one accident.
I actually don't find it very hard to believe this could be done myself, because you are talking about almost all city, or near highway driving which is generally straightforward. Pretty much no obstacles, maybe some construction zones which are usually pretty well marked on highways.
I don't even find it hard to believe an individual with a lot of skill and programming/AI/electronics knowledge could have pulled this off.
"There is more worth loving than we have strength to love." - Brian Jay Stanley
Self-driving works the vast majority of the time. How many attempts were made (by him and/or others) that we're not hearing about because they had to be aborted? Just doing it once is not exactly Lewis & Clark territory here.
How is the Riemann zeta function like Trump rallies? Both have an endless number of trivial zeros.
He doesn't lack a store of trust because of the accident, he lacks the store of trust because he's a thief.
So a guy who is highly desired by TWO companies for his autonomous driving knowledge cannot be "trusted" to have managed to drive cross country in one?
I wouldn't trust him very much in terms of a contract, but he obviously has the TECHNICAL skills to do what he claims. And there is video of course.
Like I said, what he did was not even that TECHNICALLY challenging, so it's more likely than not he did what he claims - despite how he may have treated employers he worked for.
"There is more worth loving than we have strength to love." - Brian Jay Stanley
Levandowski told the Guardian that, although he was sitting in the driver's seat the entire time, he did not touch the steering wheels or pedals, aside from planned stops to rest and refuel.
So not any different from driving with cruise control and lane assist, then.
huh? driving is not the pinnacle of human achievement. This is a very doable task. God, I hope that in 5 years we are laughing about how people used to have to drive cars themselves. There will be a lot more of us laughing too since many of us won't be dead from the car crash epidemic that plagues us currently since humans are pretty terrible drivers.
Really? Very doable? You should launch a startup then. Make billions!
In 5 years you'll have to keep on dreaming. They are so ridiculously far from making this work, two years in. They are not on the verge of multiple breakthroughs.
Laws are rules for the court, but merely a bottom bar to hit for life. Think beyond laws in your actions always.
I may be wrong, but I have the feeling that running coast to coast through highway is easier than running in residential areas, where pedestrians can pop up at any time.
I can tell you that this isn't really a test, even if he did do it.
That is exactly what I am saying, it's not really that much of a test of todays autonomous driving cars because pretty much all highway driving is easy.
I do think the car probably drove itself to the gas stations and hotels used - that's a bit more impressive, but also like I said, roads near highways tend to be pretty wide, and clearly marked.
Pure city driving or really bad weather is more where things get interesting I think.
Even though you and I know this is easy though, it should be a real wake-up call for lots of self-driving car naysayers.
"There is more worth loving than we have strength to love." - Brian Jay Stanley
I use comma.ai and their opensource platform OpenPilot. It works. It's here today. It does not know how to handle crazy events, like a deer running back and forth across the street (it will try to avoid and then it will stop and won't go until the driver takes over).
What is not here is autopilot for the masses. A competent driver is required to be at the controls, that is why a system like OpenPilot will never be a platform. Too many idiots raises the cost of liability. However crowd-sourced insurance is sure to come soon
It didn't use lasers? Sometimes a big "spot" in the road is merely discoloration, spilled paint, or a reflection. Such could easily fool camera-driven AI to slam on the breaks, risking a rear-end collision. Lasers can verify such a spot is "flat" in a more direct manner.
I've had close calls myself over mistaken identity because reflections etc. confused (human) stereo vision, being one eye may catch a reflection that the other eye doesn't, and the brain thus misaligns the pair of images, or at least gets an ambiguous result, which should result in breaking until "solved".
It's possible that "camera-only" AI is good enough 99.9% of the time, but I'd prefer lasers to double-check, because 99.9% is not good enough with cars.
It's kind of comparable to, "I eat bacon 3 times day, and don't have health problems; therefore bacon is safe."
Table-ized A.I.
So if it only used video camera's it was not driving at night.
Nor was it being placed in situations were there were indecipherable objects in the roadway.
Video can't judge object size or reliably get distance to an object unless it has some references.
For example, a hefty bad in the middle of the road. Now if the road is straight and the lines are dashed or even spaced so that the system can estimate paralav then it could figure out the extent of the hefty bag partially. It will not be able to separate distance along the road and height. If the hefty bag is stationary then as the car approaches it "might" be able to sort those two out. But hefty bags usually are moving and are dark.
Consider a semi truck crossing the intersection in front of you. Now here things get tougher. The road is visible under the truck on the other side of the truck. So that provides false infonrmation making it seem like the bottom of the truck starts at a position much farther away. And it also block the horizon view of the road so you lose the size and plavement information that holds.
thus you get to crash into the truck.
If the side of the truck is also painted to look like say a jungle or a road or a person or dazzle camoflage things might get even crazier.
In short what he did was rigged and reckless even if it was happily Wreckless.
Some drink at the fountain of knowledge. Others just gargle.
It's basically first-gen Tesla auto-pilot.
There's no footage of it doing anything but staying in one lane.
Nothing of it pulling in to a gas station, nothing of it navigating an interchange.
It seems to not even be able to change lanes.
Yeah, it's basically first-gen tesla autopilot.
it can stay in a lane and react to changes in traffic speed.
What if it had 2 cameras in the same direction?
Wow, sent an e-mail as suggested when clicking on "use classic" banner, and got a fast response that addressed my msg
it doesn't help much for distant objects.
Some drink at the fountain of knowledge. Others just gargle.
I would love to do this.
It would not be difficult to build an autonomous vehicle that could drive coast-to-coast on I-80. The autonomous systems company I worked for in the oughts could have built such a vehicle in the 1970s. Except of course for those tricky bits at the endpoints, San Francisco and New York City, I'd suggest it not trying to go through the Chicago metro area at rush hour either even though I-80 is fairly far south. So yeah, possible, easy even - except for the hard part.
Coast to coast is all highway driving. Still impressive, but I would think that getting around in a congested city would be more challenging.
On the highway you have no bicyclists, or dogs, jumping in front of you. No traffic lights, no stop signs. Nobody running stop signs or traffic lights.
Video can't judge object size or reliably get distance to an object unless it has some references.
It should provide at least the same possibility to judge those as a human does with only visual input, which seems to be enough to drive a car from coast to coast.
However, I find it rather unlikely that this guy has *algorithms* that can make those judgements to the same level as a human.
Or he should be prosecuted for reckless endangerment.
What if it had 2 cameras in the same direction?
it doesn't help much for distant objects.
So your objection is that it can't do something that humans can't do, and that somehow renders it unable to drive? With greater stereo separation, you get greater depth perception. Our eyes are just a few inches apart. In a car, you can place the cameras feet apart. An AV's ability to determine depth from cameras alone can be superior to ours, and ours is good enough for driving. QED, this is not a real limitation.
I'm not saying he has done it, mind you. I'm saying it's conceivable.
I can drive with one eye closed with almost zero degradation in driving quality, because of the time I've spent playing driving video games where you get no depth information anyway, and have to infer it all from the apparently changing sizes of objects. But even if multiple cameras were necessary to estimate distance, we don't need to know precisely how close distant objects are, specifically because they are distant. Range to them is irrelevant until they are nearby.
"You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
Interesting. I was not aware of this. Thanks!
"Don't meddle in the affairs of a patent dragon, for thou art tasty and good with ketchup." ~ohcrapitssteve
In 1995 Carnegie Melon drove cross country no-hands in a automated Pontiac minivan. Not quite as automated, as it was one camera facing front, a gps (for speed only, no good enough maps available), and a 486 computer. They claim in the journal (ahh, the days before blogs) that the car drove 2800 of 2850 miles across country.