Slashdot Mirror
NVIDIA To Install Computers In Cars To Teach Them How To Drive
jfruh writes: NVIDIA has unveiled the Drive PX, a $10,000 computer that will be installed in cars and gather data about how to react to driving obstacles. "Driving is not about detecting, driving is a learned behavior," said Jen Hsun Huang, CEO of NVIDIA. The data collected by Drive PXes will be shared, allowing cars to learn the right and wrong reactions to different situations, essentially figuring out what to do from experience rather than a rigid set of pre-defined situations.
Is that you?
I recognize that analyzing lots of data across lots of cars, drivers, and routes might yield useful knowledge. I'll bet there are even insights that no single human driver could ever gain.
But an awful lot of driving behavior comes from things that have nothing to do with anything this computer can monitor -- specifically, the driver's thought processes. If I slam on the brakes suddenly because I remember something I forgot at home, what will the computer make of that?/p?
... in order to learn behavior you need to know where things are in space (aka detecting that they are there). What a dumb statement.
Only is this a success if the self driving car can appropriately extend a middle finger to other motorists who are giving them road rage.
The computer isn't learning from experience it is being programmed by a different method. basically they are copying other drivers reactions to a set of obstacles so that the programmers don't have to create all those rules themselves.
Think of it this way instead of manually programming a replacement robot arm on an assembly line they are copying the program code over directly to save time. This isn't a bad thing. However it is far from learning.
i thought once I was found, but it was only a dream.
a behavior is: observable activity in a human or animal. the aggregate of responses to internal and external stimuli. a stereotyped, species-specific activity, as a courtship dance or startle reflex.
Driving is the appliaction of education, study, and practice of a specific set of rules and regulations pertaining to a specific type of motorized vehicle. the reason we have different licenses for motorcycles, cars, and semi tractor trailers is more than enough to discredit the idea that driving is a behavior itself. You could classify things like tailgating and jackrabbit starts as poor behaviors associated with operating a motor vehicle, but driving is far more of a learned skill than an expression of ones inability to cancel a turn signal.
if we distilled driving to behavior, as nvidia insists, we would have a car that refused to turn its headlights on in the rain and couldnt properly allow vehicles to overtake while passing. It would go 25 miles over the speed limit, ignore school zones, and divert 100% of its resources to text messages and cellular communication at random intervals. It would occasionally ignore green lights and red lights, and it would tailgate and merge without signaling almost religiously. it would erroneously yield right of way at a roundabout, it would ignore speed limits in construction zones, and it would short-stop at continuous merge lanes and wait for traffic to pass before entering. In short, it would be the single most dangerous thing on any tarmac or asphalt since Gary Busey.
Good people go to bed earlier.
Even for the often flawed human drivers, this rings true. It seems one of the more common single vehicle highway accidents is the slight drift off the road followed by the panicked, aggressive over-correction... experience teaches us to gradually bring the vehicle back in line by fighting the gut-reaction to hurry.
Happiness in intelligent people is the rarest thing I know.
Ernest Hemingway
Accelerate and flip the obscenity bit.
He aint buying
What are they going to learn? How to not pay attention; how to not allow other vehicles to merge; how to force their way in when not allowed to merge; how to tailgate; how to brake check when others are tailgating; how to not use turn signals; what type of actions from other vehicles should cause them to rage; how to rage properly; how to ignore all the signs leading up to your exit and then cut across three lanes to take it at the last second; how to drive slow in the fast lane; how to pass when there isn't really room; how close they can get to a bicycle without actually hitting it; or hitting it, either way; ... etc..
"Driving is not about detecting, driving is a learned behavior," said Jen Hsun Huang, CEO of NVIDIA
It's about learning what to do with what you have detected, from position to collision detection is critical, then knowing what to do with that data is also critical.
"If any question why we died, Tell them because our fathers lied."
Great, just great, where I live the rules of the road are treated as a "guide" line. A million ass hats experience of bad driving here will outweigh the few who actually driver responsibly, so I will end up with computers tailgating me and going through red lights.
There are three kinds of falsehood: the first is a 'fib,' the second is a downright lie, and the third is statistics.
This is an indispensable aspect of the road to automated driving, so to speak. Yes, learning human behaviour behind the wheel would be detrimental if it were applied directly to the AI's own driving algorithms, but consider that it's going to need to anticipate the actions of all the human drivers around it. Yes, I would like my self-driving car to have a very fine-tuned set of expectations for the idiocy it'll encounter.
Is this just a way to keep a robot-car manufacturer from specifically assigning weights to various bad outcomes and possibly avoiding lawsuits?
Suppose a crash looks imminent. Whose life is more valuable? Instead of programming for this specifically, the manufacturer uses algorithms developed by obserations. Then the manufacturer could argue that it's not to blame when one person dies instead of another.
In any case, this sounds like a great way to teach a computer how to drive badly.
No one is a perfect driver - we don't want to teach our mistakes.
People make correct or safe driving decisions based on inputs that cannot always be well measured - we don't want to teach incomplete rules.
Ford has unveiled a car prototype designed to be parked inside a desktop computer. Its purpose is to teach cars how to compute.
speed limits set to low is an issues and 85%+ of traffic is going over it.
Also some roads that are set to 55-65 people see that as the old 55 law and go over it. Now if the roads where posted at 70 (some of them used to be there before the old 55 law) then less people will be going way over the limit. Also get rid of the 45 work zones when no one is working / in places where a 55 work zone is good.
>"Driving is not about detecting"
This man should never be allowed within thirty feet of a car, much less drive one or design one that drives itself!
When and where will these cars be tested on the roads, so I can know when to stay home? Sounds like a horrible plan.
At least computers never eat or apply makeup while driving. But can they text? How many threads?
I still don't trust the idea of giving 3000+ pound vehicles autonomy. Thankfully, they still need humans to pump gas.
Remember kids, if you're not paying for the service, YOU ARE THE PRODUCT THAT IS BEING SOLD.
Teach a bunch of chips to drive?
They're Asian!
PS. Do I get first stereotype at least?
Put it in the car with the Stig... That might make me consider a self driving car.
Slow Down Cowboy!
nVidia requires you to wait between each successful usage of the brakes to allow everyone a fair chance at avoiding accidents.
It's been 4 minutes since you last successfully used your brakes.
What sensors would these test automobiles be equipped with?
The "crash avoidance" radar front and rear as being installed on mid market and luxury vehicles?
The visual sensors of Subaru EyeSight
Unique radars or lidars as used on the Google autonomous vehicle?
What GPS sensors, to what accuracy?
The article is very short on facts.
One thing I fear about driverless vehicles are implementations based on outcome of learning algorithms that will be sub-optimal, impossible to understand, verify or debug.
Would much rather see a model driven approach with enough simulation capability to support an objective function causing the least amount of havoc regardless of the situation. As a practical matter count of possible reactions to a given event is limited.
Green light: go, red light: stop, orange light: go very, very fast!
Cause my Quadro crashes all the time!
If their experiment is successful, their computers should be able to give feedback fairly quickly as to what the driver should do before they actually do it. Advanced braking warning, correction if you get distracted/drowsy and start to drift out of your lane, speed corrections if the speed limit is changed or the road conditions changed etc etc. This idea is really cool and all, but they need to pack in a ton of sensors for any meaningful learning on the computer's part... once these sensors are there and the computer has some experience with decent drivers, the next step is to see how much better it can identify a bad situation and how much faster it can make the correct decision and if it can do it at a reliable rate. I'm not sure I want a computer trying to be my annoying Back Seat Driver.
will the cars only control 3/4ths of the tires, with the last tire "slowly kicking in" when roads get slick?
NO MAN! It is just a way that the cars can legitimately claim that we didn't land on the moon, that the CEO of its corporation is actually an Illuminati reptilian, and that the fluoride in its cooling water is there so that it can be remotely controlled by HAARP. Even suggesting that it's for computer-derived life-worth calculations is squarely in the loony toon conspiracy nutter camp and you should feel bad for being so dumb. Liability shifting.....goddamned moran....
Learning by observation tends to miss important exceptions when they're not observed.