I love GTAV so much, it's the most real single player world I've ever experienced. I once rear-ended an NPC because they suddenly stopped in the middle of the street. I got out of my car (because it's GTA and you have to shoot everyone, duh) I saw the NPC had stopped because he hit a cat with his car!<p>I'd be lying if I said OpenAI Universe / GTAV wasn't one of the primary drivers in me trying to learn ML.<p>Anyone know what kind of hardware I will need to play with this? The AMI page on github mentions an AWS g2.2xlarge (iirc a GTX 1060 is slightly better than this?). And it seems like GTA is actually running on a different system you use VNC to share the screen/input? Any estimates about whether a 1060 or 1080 on a beefy gaming PC could handle both at once?
Ok, I am <i>so</i> glad that we're not training AI to "win" at GTA V. That is not an android I want to meet on the street :-)<p>The transitive property of teaching an AI to drive given a world simulation, makes me wonder if you can train a network to recognize speech by feeding War and Peace through a text to speech program and then sending the audio data generated into an RNN. Where is the ground truth in such a system which you can calibrate your results by?
This is cool.<p>Next step is obviously teaching it how to kill people in the game and get away from police.<p>I'd like to see the AI go on a rampage and win :).<p>That should be enough to wipe off a lot of smiles off our idealistic faces :)
Who's ready for some "Deep Pimping"?<p><a href="http://gta.wikia.com/wiki/Pimping" rel="nofollow">http://gta.wikia.com/wiki/Pimping</a><p>ABSTRACT: We explore the adage "pimpin' ain't easy" in the context of deep reinforcement learning in the OpenAI environment GTAV. We train a Deep Q-Network (DQN) to optimize player cash flows. Over the course of training, our network learns to keep these hoes in line, and discipline misbehaving customers.
This is incredible! Huge fan of GTA V / Rockstar as well as Open.AI<p>Will be super interesting to see how all of the randomness in GTA V driving influences autonomous driving AIs.<p>I could see them being a bit more "cautious" in order to deal with all of the crazy drivers (and events) sharing the roads of Los Santos.
I wonder if AI is where games can make use of many cores. If the game logic can not be reasonibly parallelized anymore, the unused computing capacity can be used to add "intelligent" agents into the game. They don't need to do revolutionary things, just being more complex, more "reasonable" and less predictable than the usual agents would be enough. I think that has the potential to make some games more interesting.<p>Imagine a game like Warcraft 3 where the units actually seem to have a personality and follow "their guts" in battles rather than following simple patterns like attacking the nearest enemy.<p>Maybe I'm ignorant about the current state of the art and games already do that. Do you know examples?
It looks like it's not just using the camera/vnc images as input, but also uses depth data to simulate LIDAR input (<a href="http://deepdrive.io/" rel="nofollow">http://deepdrive.io/</a>).<p>I wonder, what makes it hard to rely on solely the 2 cameras to calculate/guess the depth, similarly to the way us non-robotic drivers do it, when it comes to machine sensing.<p>EDIT: found this: <a href="https://web.stanford.edu/class/ee368/Project_Autumn_1516/Reports/Appiah_Bandaru.pdf" rel="nofollow">https://web.stanford.edu/class/ee368/Project_Autumn_1516/Rep...</a> on the topic. Is there anything else I should read up on?<p>EDIT: found lots of articles: <a href="https://www.google.com/webhp?sourceid=chrome-instant&ion=1&espv=2&ie=UTF-8#q=optical+flow+depth+estimation" rel="nofollow">https://www.google.com/webhp?sourceid=chrome-instant&ion=1&e...</a>
Open world games seem like a great place to work with AI. What has traditionally prevented more intelligence, in my mind, was the ability of AI to understand and integrate with its environment. For example, if I say I threw a baseball, you can visualize that. In a virtual world, the AI can "see" and potentially interact with every object to increase understanding. Great experiment and I look forward to seeing how far you can take it!
Off-topic, but does anyone know whether it's the GTA engine used to record these videos:<p><a href="https://www.youtube.com/watch?v=T5kooJeXGiI" rel="nofollow">https://www.youtube.com/watch?v=T5kooJeXGiI</a><p>There's a ton of them on YouTube, but I can't figure out how they make them.
This is super cool. Are there any good resources/how-to modify GTA V for things like this? I'm looking in to GTA main website and wikies but info for developers looks thin.<p>Here's few links I found so far:<p><a href="https://en.wikipedia.org/wiki/Modding_in_Grand_Theft_Auto" rel="nofollow">https://en.wikipedia.org/wiki/Modding_in_Grand_Theft_Auto</a><p><a href="http://www.dev-c.com/gtav/scripthookv/" rel="nofollow">http://www.dev-c.com/gtav/scripthookv/</a><p><a href="https://www.youtube.com/watch?v=wS8oJTHqf8Q" rel="nofollow">https://www.youtube.com/watch?v=wS8oJTHqf8Q</a><p><a href="https://www.reddit.com/r/GrandTheftAutoV_PC/comments/38wct0/guide_to_modding_gta_5_safely_and_effectively/" rel="nofollow">https://www.reddit.com/r/GrandTheftAutoV_PC/comments/38wct0/...</a><p><a href="https://steamcommunity.com/sharedfiles/filedetails/?id=558079253" rel="nofollow">https://steamcommunity.com/sharedfiles/filedetails/?id=55807...</a><p><a href="http://gtaforums.com/forum/109-tutorials/" rel="nofollow">http://gtaforums.com/forum/109-tutorials/</a><p>Also I'm reading that GTA folks went out of their way to prevent modding. So may be everything you do with GTA is a hack and there is no supported APIs for developers?
This feels like an important step in the democratization of AI.<p>We have algorithms, data, and computing power which are unevenly distributed. Providing easy-to-access training data for a self-driving car feels like a huge step in democratizing the data aspect of car AI.
It would be really great to see OpenAI sponsor AI contests, the # of people contributing / progressing on specific goals would be enormous.<p>This is really awesome, great work to those involved.
too bad GTA V is not open-source for researchers. I don't know how customisable it is, but e.g. experiencing snow/ice or a failure like spontaneous pressure loss in one of the tires would be interesting to simulate.
> artificially slowed to 8FPS<p>i'm sure this is to offer relief to the nn but does this further simulate 'reality'?<p>are our autonomous vehicles seeing at 8 frames per second?
Just a heads up to anybody who tries to get this running, you may have to request permission from Amazon to get the required instance going (g2.2xlarge). Apparently they are making sure people aren't trying to set up botnets and bitcoin mining. Just sent in my request, we'll see what happens.
This would make an interesting plot for a movie. Someday hacks GTAV ai into Google self driving cars. Suddenly "do no evil" company becomes "fuck the world up" and man tries to fight cars.<p>Anything that has wheels is dangerous to humans!
Imagine using this to train and to "unit" test a self driving car. With realistic graphics and the ability to control events, I wonder how useful this could be. (e.g. simulate near accident situations)
It looks like it has driving support but I read the post and can I make a pedestrian agent? It doesn't appear that I could. This looks like something I would like to play with and an awesome project.
now the interesting part: put it in front of ethical problem (take one from <a href="http://moralmachine.mit.edu/" rel="nofollow">http://moralmachine.mit.edu/</a>) and lets see what it does
I'm probably one of the most critical people of the new AI-safety craze, but even I wonder if training AI on games like GTA is a good thing to do... if it is something I'd be hesitant to let my kids play, should we really be training our mechanical offspring on it?