cancel
Showing results for 
Search instead for 
Did you mean: 

NVIDIA AI turns reality into virtual reality (one step closer to the matrix)

LZoltowski
Champion
Seems that an idea I wrote on the proverbial napkin a few years back that posed the following question:

"I wonder if you could get AI to look at GTA V game footage, and then after being trained what it should look like in real life, could the AI apply reality, onto the 3d game engine"

Well, folks, NVIDIA researchers are on it, after training AI on roads, buildings, cars etc they have managed to make the AI generate "reality" on top of a rudimentary Unreal Engine environment. This is all early stages, but the possibilities are endless, and a true precursor to the matrix future.

Just think about it, in the future there won't be ray tracing, raster or any other "traditional" graphics engines, or cuda cores, there will be Neural Networks that generate images indistinguishable from reality running on Tensor cores, 

https://www.youtube.com/watch?v=ayPqjPekn7g

https://www.engadget.com/2018/12/03/nvidia-ai-video-to-video-synthesis/
Core i7-7700k @ 4.9 Ghz | 32 GB DDR4 Corsair Vengeance @ 3000Mhz | 2x 1TB Samsung Evo | 2x 4GB WD Black
ASUS MAXIMUS IX HERO | MSI AERO GTX 1080 OC @ 2000Mhz | Corsair Carbide Series 400C White (RGB FTW!) 

Be kind to one another 🙂
25 REPLIES 25

kojack
MVP
MVP




Perhaps computing cost is too great, I wonder if GPU's having Tensor Cores being more common in the future will really give us AI powered NPC's. 


Not if the tensor cores are also being used for ray tracing and dlss. 🙂


It also depends on the actual AI techniques being used. Neural Networks seem the most suited to Tensor, but there's a lot of other techniques that are useful in games that might not fit as well (while I've done a fair bit of GPU coding and a LOT of CPU coding, I don't know much about the tensor core model).


Since we're talking about AI and graphics, here's a certain moderator's avatar recreated by a line drawing genetic algorithm (specifically an asexual reproducing hill climber) that I wrote. 40000 generations (only a few minutes) and a genome of around 4700 lines. (I set it to cap off adding new lines at 5000, at that point it is only allowed to delete or move existing ones).

qnpcne7fbz0l.jpg

While genetic algorithms themselves aren't directly AI, they are a common method of growing neural networks to solve problems that don't have a known solution to train against. Most of those demos you may of seen of physics bodies (creatures, bipeds, etc) learning to walk are done with genetic algorithms breeding neural networks.


One of the big issues of game AI is that more advanced versions can be rather unpredictable and very hard to test. Bethesda ran into that with their Radiant AI in Oblivion. It was really cool, but so many unexpected (and game breaking) things happened in testing than they had to disable it before release.
Author: Oculus Monitor,  Auto Oculus Touch,  Forum Dark Mode, Phantom Touch Remover,  X-Plane Fixer
Hardware: Threadripper 1950x, MSI Gaming Trio 2080TI, Asrock X399 Taich
Headsets: Wrap 1200VR, DK1, DK2, CV1, Rift-S, GearVR, Go, Quest, Quest 2, Reverb G2

JohnnyDioxin
Expert Trustee
Case in point:

People have complained in more than a few flight sims about the "Golden AK47" - the single shot from a rifle that brings down their multi-million pound fighter, yet recently I have been reading about unofficial RAF ops in Oman, and one of their aircraft was brought down by a single AK47 bullet - which went through one of the fuel lines.

i5 9600k @4.5GHz; 16GB DDR4 3200; 6xSSD; RTX2080ti; Gigabyte Z390D Mobo
Rift CV1; Index; Quest; Quest 2

Anonymous
Not applicable
What they REALLY need to do for neural networks is base everything on a superior brain. I will quite happily donate my brain to gaming science when I pop my clogs so that they have something top drawer to work with. B)

LZoltowski
Champion

snowdog said:

What they REALLY need to do for neural networks is base everything on a superior brain. I will quite happily donate my brain to gaming science when I pop my clogs so that they have something top drawer to work with. B)



Image result for steve martin brain
Core i7-7700k @ 4.9 Ghz | 32 GB DDR4 Corsair Vengeance @ 3000Mhz | 2x 1TB Samsung Evo | 2x 4GB WD Black
ASUS MAXIMUS IX HERO | MSI AERO GTX 1080 OC @ 2000Mhz | Corsair Carbide Series 400C White (RGB FTW!) 

Be kind to one another 🙂

Jake_Dragon
Adventurer
You know your going to have to buy a new PC right 😄 

shadowfrogger
Heroic Explorer
Yeah. There are lots of issues with machine learning A.I.  Most of these problems, Devs will have to create work around and tools to help create human like A.I.  Or sloppy and beatable A.I depending on the game type.  E.g, in FPS or platform games. The most efficient way to move is micromovements which can't be done by a player.  You may have to introduce some sort of reaction time into the learning code.

Other possible ways are to beta test and record massive amounts of human gameplay and try use that data to recreate a workable npc. But that would mainly only work for movement and shooting patterns. (Which still is useful)
I think this type of AI will be the future of all games as eventually it will be easier to set the rules up and get it working for your game rather then manually coding everything.  But I don't think the tools to really do this well on a AAA game exist yet. So we will probably see small implementations and improvements until the first AAA game manages to pull off some AI that is truly ground breaking and reamarkable.  Then there will be no looking back.

Visit my amateur homegrown indie game company website! http://www.gaming-disorder.com/