Grand Theft Auto V Is Being Used To Help Teach Self-Driving Cars (bloomberg.com) 57
An anonymous reader quotes Bloomberg:
In the race to the autonomous revolution, developers have realized there aren't enough hours in a day to clock the real-world miles needed to teach cars how to drive themselves. Which is why Grand Theft Auto V is in the mix... Last year, scientists from Darmstadt University of Technology in Germany and Intel Labs developed a way to pull visual information from Grand Theft Auto V. Now some researchers are deriving algorithms from GTAV software that's been tweaked for use in the burgeoning self-driving sector. The latest in the franchise from publisher Rockstar Games Inc. is just about as good as reality, with 262 types of vehicles, more than 1,000 different unpredictable pedestrians and animals, 14 weather conditions and countless bridges, traffic signals, tunnels and intersections...
The idea isn't that the highways and byways of the fictional city of Los Santos would ever be a substitute for bona fide asphalt. But the game "is the richest virtual environment that we could extract data from," said Alain Kornhauser, a Princeton University professor of operations research and financial engineering who advises the Princeton Autonomous Vehicle Engineering team.
The idea isn't that the highways and byways of the fictional city of Los Santos would ever be a substitute for bona fide asphalt. But the game "is the richest virtual environment that we could extract data from," said Alain Kornhauser, a Princeton University professor of operations research and financial engineering who advises the Princeton Autonomous Vehicle Engineering team.
Re: (Score:2)
Hard reverse and hard right when a cop walks up to your door to give you a ticket! Then burn off over his body.
Re: Hmmmm Maybe Not? (Score:2)
Re: (Score:1)
Re: (Score:1)
Re: (Score:2)
?
That hasn't been a bonus since the first GTA (sadly)
Re: (Score:1)
Because you've not updated your videocard since you bought GTA3?
https://en.wikipedia.org/wiki/... [wikipedia.org]
The problem is depth perception (Score:2, Informative)
Your eyes are far better at matching light frequencies between both eyes to get the depth mapping correct. Your standard camera can only distinguish 24 bits of light frequency. At that level you get somewhat of a depth map but not a very good one.
Lasers try to get around that limitation by using a frequency the camera can easily pick up and compare between the two images. If you could use the whole image and any frequency, you'd be a lot better off.
That's ultimately the challenge: getting cameras that ar
Re: (Score:3)
Your eyes don't have enough parallax for depth perception to be that accurate in the ranges needed for driving. Each pixel does not have to be independently matched.
Which isn't to gloss over the difficulties to 3d vision in real time. The depth perception part is solved by two meters of parallax.
Re: (Score:2)
Research Paper (Score:2)
https://cias.rit.edu/media/upl... [rit.edu]
Luminescence can be used to create a depth map using two stereoscopic images.
3D movies and VR headsets don't work for everyone and create headaches for a variety of reasons, one probably being the lack of bit depth.
Re: (Score:2)
They'd need to render 48-64 bit color to emulate what might be possible in the real world to get accurate depth information.
Or use radar.
Re: (Score:2)
Your eyes are far better at matching light frequencies between both eyes to get the depth mapping correct. Your standard camera can only distinguish 24 bits of light frequency. At that level you get somewhat of a depth map but not a very good one.
No. You don't get a depth map from one camera. That's not something they do.
Lasers try to get around that limitation by using a frequency the camera can easily pick up and compare between the two images.
LIDAR [lidar-uk.com], how does it work!11!!!?/?!?1?? (hint: it doesn't use a camera)
Re: (Score:2)
You can get a depth from a single camera if the object or scene is suitably lit.
People do it with IR. But it's pretty crap, and it can be fooled by some surfaces and materials. And when you talk about what people are actually looking at doing in cars for full autonomy, it's combining normal visual cameras with lidar and radar.
Re: (Score:2)
That is why cars use black and white for that. ... facepalm.
The light frequency is completely unrelated to a depth map.
The distance of your eyes is
Re:The problem is depth perception (Score:4, Funny)
Found the DeVry biology grad.
Research Paper (Score:2)
https://cias.rit.edu/media/upl... [rit.edu]
Luminescence can be used to create a depth map using two offset images.
3D movies and VR headsets don't work for everyone and create headaches for a variety of reasons, one probably being the lack of bit depth. It's much harder to judge depth when the colors are so close together.
Re: (Score:2)
Human eyes aren't digital cameras, you pillock.
Re: (Score:2)
Re: (Score:2)
The probem with that is that they are most likely doing some kind of reinforcement learning. That requires not just an input, but also an ability to respond to the input, and then be scored on the response.
A game that simulates driving like this might be an excellent way to get a baseline level of training for the AI, since it can experience many more combinations of situations in simulation than it can in real life driving.
Re: (Score:2)
Re: (Score:2)
Really what's needed is lidar. Lidar can build a 3D image that would actually be suitable to let the game AI roam around in. No one is going to do proper self driving cars without a serious lidar solution. The ones available right now (mostly Velodyne) aren't that great but, the stuff that's in the pipeline is pretty fucking amazing. You can even get doppler information out of newer lidar approaches and that's like the greatest thing since sliced bread for self driving cars.
Re: (Score:2)
Your eyes are far better at matching light frequencies between both eyes to get the depth mapping correct. Your standard camera can only distinguish 24 bits of light frequency. At that level you get somewhat of a depth map but not a very good one.
Waymo uses LIDAR, not visual light cameras. It gets an extremely accurate depth map, far more accurate than any human could, because LIDAR measures the time it takes light to reach the "seen" object and bounce back to the receptor.
In a 3D mapped world, all the depth information is 100% accurate.
Which is only slightly better than LIDAR-derived depth information.
Hoe (Score:5, Funny)
Re: Hoe (Score:2)
Oh, they will stop for them alright...
Re: (Score:2)
At least I know that my autonomous vehicle will have had training on how to deal with a carjacker or with someone trying to run it off the road.
Re: (Score:2)
Unfortunately, this normally means hopping onto the sidewalk and driving off at full speed with no concern about how many nuns and small children you run over.
Re: (Score:2)
Depends on driving mode engaged; "Libertarian" she's probably OK, "Gangsta" or "Fundamentalist Christian" not so much
Re: (Score:2)
First the former, then the latter. Free health bonus!
Should we train our AI in GTA world? (Score:1)
This is how Sky Net, et al become cynical
It's all fun and games (Score:5, Funny)
remember cops are bad in that game (Score:2)
remember cops are bad in that game
Re: (Score:2)
Based on the Waymo vs Uber Law Suit... (Score:2)
Cool (Score:2)
So they're teaching their cars how to find prostitutes and avoid the cops?
I read using GTA V and immediately thought of.... (Score:1)
Rockstar Not too Happy with this (Score:1)
It turns out that Rockstar is not so happy about them using their software with out asking first:
http://mashable.com/2017/04/21... [mashable.com]
“We welcome discussions about the use of our technology to help further academic research, but it’s obviously not appropriate for corporations to take our work and use it for their own financial interests or for researchers to distribute unlicensed copies of our code as part of their work without first seeking our permission,” the company said.
I suggested training on sims to Alain a decade ago (Score:2)
I worked with Alain Kornhauser about thirty years ago, first taking his robotics course as an undergraduate, later managing his robotics lab as an employee, and then again even later (briefly) as a grad student tangentially as part of a group doing self-driving car research focused mainly on a neural networks approach. I had also been hanging around Red Whittaker's group making the first ALVAN (Autonomous Land Vehicle) around 1986 before going back to Princeton to work as an employee.
While I did not contrib
Not sure that's a good idea (Score:2)
NPC cars in GTA V usually drive quite "neatly" and predictably. Not sure it will be an adequate representation of real life where you have morons doing stupid things every once in a while.
Unless they are using the online version of the game, in which chase there will probably be too much chaos on the streets.