Legit rollercoaster level drama, cheating scandals, amazing achievements, and now the most impressive AI I've seen play a game. I agree one of the best parts is how well the community tells stories and gets us engaged when we don't even play the game. Probably my favorite reaction content too.
@@Brad-dx9fd if you want to see an even more impressive 'AI' Gaming bot - you should check out Seer from Rocket League. It even plays live against pro players and wins.
@@rasol136you would have to take the floor of a height modulo something (assuming they are equally spaced) as one big jump could ruin that model. Or maybe it just uses that as a shortcut 🤷
@@tom_skip3523 i mean a brain of a fruit fly is pretty impressive since it can use it perfectly. Yeah there might be better ais but not in trackmania (atleast i think there aren't)
Hey, quick thought on optimizing pixel rendering for AI vision. You mentioned that the AI has a relatively low resolution of images that it looks at to make decisions. What you could do is check the activation functions for each visual input and reduce the resolution on groups that have low activations and increase resolution on groups that have high activation. This is similar to selective focus in our natural vision where we may generally process the whole picture as we walk into a room, but may not notice that the bear on the table is holding a bottle of tequila until someone says "Hey, notice something strange about that bear?" Then, without moving closer to the bear, we look much more closely at the specific details to determine what we are looking at. Here's a computational example: Imagine the following 3x3 image (as text) A B C D E F G H I each frame will activate each pixel at a different amount, say 0 to 1. Here's an example activation array for the above: 0 1 0.5 0 0 1 0.2 0 1 Now, for the instances that have a low value, say less than 0.4, we will keep the resolution the same. But for higher activations, we may need to see more clearly so we have each pixel divide into another 3x3 that gives better resolution. In this case; B, C, F, I all appear to have high activation, meaning high importance, so we increase resolution to increase accuracy: C now looks like this: C.1 C.2 C.3 C.4 C.5 C.6 C.7 C.8 C.9 Same for all other pixel activations that are this high. You can further subdivide as low as you needed to get the most detailed activations to get a pixel perfect view of the tight turn or barrier, without having to strain the model with extremely high activation rendering on every frame. I don't know the exact architecture your AI is using, but this selective attention strategy works with most vision models. Let me know what you think!
Yea I feel like this would be great, especially since the screen more or less can look similar a lot of the time. Like the car is pretty much always in the same spot, taking up the same amount of space, which probably isn’t all the useful for the ai.
This sounds like a cool technique. Do you have any papers that apply this work? I've definitely seen the application of an attention mechanism, but never one that re-captures frames based on the attention weighting of the region
@@laurasisson1611 These people did something similar, and it seems like its working The name of the paper: 3D CNNs with Adaptive Temporal Feature Resolutions
Sounds like you found some nice potential for improving the AI! But just to understand the algorithm properly: the activation array would change dynamically during each step, which is the equivalent of „shifting focus“. But from a pc memory perspective (which he apparently already struggles with), would you have to save a full-scale image of each frame of the game to then be able to feed a small focus-part to the network? I think this might blow his current pc in terms of memory, even if the network itself only needs to be marginally larger (to now additionally cope with the focus area). Please let me know what you think!
Needless to say the ai work is amazing, but I think this video is brilliant in many other ways. I think you’ve made the cleanest and easiest to understand visualisation of the learning, with the waves of different coloured cars, and the train of WR cars to beat. Fantastic job on both the technical stuff and the video production!!
I second this! I'm a machine learning researcher and the amount of effort you've put into not only getting good results but also saving the data and displaying the data of the runs in an easy-to-understand way is extremely commendable! Well done!
Bruh... If I was picky I'd say that this is not exactly the same learning method, but I'll mark my words and eat my hat if you drive Deep Fear respawnless with AI v4.
@@linesight-rlhow come it struggles with full speed? As a human that naturally prefers the full speed elements of trackmania I feel like surely it should be easier because there’s less complex inputs? Just curious
@@tylercorrin1615 My guess based on the information given in this video, FS might be harder to train because of the low resolution screenshots and low fps the AI has available. If you look at 5:50, you can see how hard it is to see far ahead, which is critical in high speeds, because what is far ahead will be much closer much earlier in high speeds compared to low speeds.
It would be really interesting instead of 0-shot learning to see how few-shot learning would work. Let the general AI fine tune on the new map for a few minutes to see how much it improves. Maybe we can see the AI play Random Map Challenge at some point?
@@cyanhacker It can, but as you see the tchniques here applied are generally guiding the AI a lot and the neural network is small. With techniques that give the AI abilities to look ahead, it could very well surpass and even plan its own route faster than humans. It woudn'T need to have human reqard shaping, insteadyou could employ PPO for optimizing the reward function along a general goal. so props to the author but a fruitfly brain is notgoing to cut it. I would love to see a breakdown of the Neual network used currently and maybe with some added hardware getting that up to speed already beats the challenge
Honestly, a map like Deep Fear seems perfect for an AI. The only scary thing about it is its length. For a human, the problem with Deep Fear is remembering the track, focus, precise input and reaction time. The AI has no issues with any of these things.
And I think the original commenter must have a misunderstanding about how it works, since they specify "respawnless," because there's no reason a fully trained version would need to respawn, it would just do the right inputs the first time.
@@SamuelBoshier Its possible some sections are easier to do with the standing respawn, every checkpoint on the map is respawnable and has boosters, doing it segmented like that eliminates some variation. The AI could potentially learn to do it like that instead.
Trackmania AI youtube videos are the reason why I am so invested into learning Neural networking now. This amazes me. You're amazing. I'm already subscribed, but if I could, I'd subscribe twice.
You can... Sub #2 = Patreon But there's MORE!... Sub #3 = GoFundMe - Contribute towards PC upgrades. Sub #4 = Secret OnlyFans 🤫- Can't imagine of what 'Content' could possibly be posted here that's Trackmania themed? AI can. 😉 /s
@@pahom2You do realise that what u said just makes you look silly, right? If you’d the slightest idea how this AI works, you’d know that it is definitely a neural network. The genetic algorithm bit is (one part of) how it’s trained, not how it runs. Saying it’s one thing or the other just doesn’t make any sense; it’s not even wrong, it’s just daft.
If this doesn't go viral, I'm going to be so sad. The AI just gets so interesting every time. Can't wait for the time that FWO itself will have trouble beating it.
I really doubt they'll recruit AI. Jealous as they are there's no way that they'd welcome someone better than them at the game. I mean, how many years has it been since TAS entered the WR scene and they don't even consider recruiting him to the team? He's been crushing every map he plays and they keep denying him membership
The field of reinforcement learning is so varied and has so much potential that has yet to be seen. Thank you for showing what a mid-size gaming rig can do and cutting down a little of the hype of scale we're seeing with AI right now. The next big breakethrough is in RL!
Man, I only watched 10m and I am already amazing of the quality of the video and the quality of the AI. Holy shit, this is the best AI I saw not supported by money. So inspirational!
As a developer, this video has me worried about my job security. You probably think I'm referring to AI taking my job, but no. I'm talking about people like Agade and pb4 massively raising the programming bar! I feel pretty bad putting my dopey little pet projects on a resume after watching stuff like this!
@@Justin73791 well it was a day like any other when i discovered AGI would arrive in 2030. i woke up at 5, went for a run, cold shower, alright team lifts. i stuck around to run some drills with coach after and then headed back for food. then i read 90 pages of the fifth book on gaius julius caesar i pirated this year and had half a pound of beef with half a stick of butter. alright then i went to my graduate topics in poptropical geometry class and dozed off, i already know the material anyways. walking back to my penthouse i stopped at a busy crosswalk and remembered that yielding is the motion the tao, so i yielded to my intrusive thoughts and walked across anyways (like that one scene in gattaca.) well i downed my 2 PM half gallon of milk and then checked if kai trump was 18 yet. damn not yet. hell of a swing though in conclusion AGI is coming in 2030
@@archsys307but it's always 6 years in the future. :) in 6 years it will still be 6 years in the future. same with fusion always being 30 years in the future
@@Decodeish1 go look at current graphs of various AI progress metrics over the past decade hardware is the growth determining factor right now… it could actually happen much sooner isnt that crazy in any case go look at some predictions from smart people in the field i saw a list, bunch of openai guys, other founders, ai researchers, etc all rather knowledgeable and intelligent people, 2/3 of them put AGI pre 2032 (in fact a ton in the 20s) and 90% of them by 2040 AGI as in rivaling an average, 50th percentile, skilled worker. nothing crazy, but it’s gonna be escape velocity extrapolating from progress patterns of fusion or say SD cars doesnt hold up to the fundamentals graphs and the resounding, damn near unanimous predictions from experts
I feel like such AI can be both a good thing and bad thing for the world recording competition. It showed how some minor improvements can be made in places human players might not think of, but also it might led most people to just blindly copy the AI's approach, like how it went in Go. One of the biggest charm of watching world records run is the creative ways players figured out to gain even the thinest margin. But anyway, this video is such a joy to watch, absolutely great work!
Honestly, I am interested in how AI can be used in TASing. TASing requires so many human inputs and knowledge of what saves time or at least thought possible so it can be attempted via TAS. The thought of AI being a tool that can be trained to help test theoretically possible shortcuts or how fast something can truly be makes me wonder about how far it can be pushed. Yes AI needs to be trained, and so still requires at least a standard level of the aforementioned knowledge, but still I think it could be a massive tool for TAS's. Especially of AI can be developed for other games where TASing is heavily reliant on human play + savestates due to a lack of other available compatible tools.
@Ezz_Fr That’s definitely true, but there’s a catch: AI can try more times in days than human can in years. When you know that someone plays the game so much more times than you do, I think it’s hard to think that you yourself can even find a better way with the your ‘super inefficient’ approach. After all, it takes too much time and luck for a human player to find, test, and execute an approach to perfection.
In the world of Chess AI has overtaken human for a long time now. AI vs AI is what generates progress and humans are studying AI tactics to learn new strategies. We are a long way from that in physics based games. Chess is just a discrete decision matrix. The video game genre that could probably be optimized the fastest would be 2D fighting. It's discrete.
@@christopherlperezcruz1507 Any real-time player vs player mutiplayer game would get easily dominated by an AI simply because of inhuman consistency and reactions. Per example, all an AI needs to dominate in CS:GO is to be able to identify the enemy and click on their head as fast as possible since the AI will be able to kill the human before the human brain even has time to send the order to shoot back + some basic patroling to find the enemies and pathfinding to get to the bombsites when the bomb is planted. The only thing the humans can do to fight back at that point is try to coordinate with granades and stuff, but the AI could also easily learn to run away from explosive/fire granades and look away from flashbangs at the exact frame where they explode. The only videogame genres where the humans could come close are either very complex turn based games where there is a lot of time to think about what to do and too many options for an AI to calculate everything even for a single turn, or time attack games like trackmania where the human can just keep farming the same map thousands of times to get close to the AI until they reach the human limit.
I mean technically it can, it discovered the barrier jump on its own. the issue is that the scoutting and labbing that it does is very very inefficient as it doesn't actually know what its looking for. a human could say "i think this might be faster" and try something new, but a neural net simply doesn't have the capability to do something like that, it can only make small adjustments.
This was awesome, not just the AI, but also really well put together and presented. Having recently started video editing myself, this is so much above what I currently could reasonably do Looking forward to your next video :)
Dude, you're killing it with this style! I love the balance between technical and simple. The editing makes it very engaging, but honestly I would watch it unedited. You're doing something never-before seen, and doing it very well. I love your voice, too. Great video!
Best trackmania video I've seen in a long time Amazing job, huge effort and it's so nice to see how far your ai is coming. It's like watching your child learning to ride a bike or something
Maybe not something zero shot but more like how we currently use llms Maybe have a model that is trained on a ton of different tracks(base model) and then fine tune it for the one it's currently attempting. This could lead to a large reduction in training time. Possibly even something use something like a Lora adapter so that only a few parameters have to be trained, and the rest can be frozen. This would mean the hardware requirements could be reduced while actually having a larger base model(no gradients and other training parameters in VRAM).
Just freeze everything except the last couple layers and then train on the track for a few iterations. This can be done easily in keras or pytorch or whatever hes using.
@@mithril_leaf well, that is how humans do it as well. People spend hours learning a track trying different lines and grinding. We just start out with more generalized knowledge and then take longer to optimize. A human trying a track ten times will get a better time than an AI without prior knowledge but given enough attempts the AI will do better.
@@TheSuperappelflapthe problem is humans have general knowledge about how to play trackmania. Based on how he set up his AI, the AI does not have any transferable knowledge between courses. He is training it to follow the red line he already defined ahead of time for that specific track. This approach is more akin to brute forcing a single map to find the optimal inputs, rather than creating an AI general knowledge for trackmania that can play on any track.
@@brunch. Right, this is not a generalized system. But it is a first step in that direction. The next step would be to take this system and use it as a component in a larger system that uses computer vision and some general knowledge about the limited amount of different blocks in the game to find the red line by itself.
This is one of the most interesting, well presented and entertaining videos I've watched on RUclips in a long time, and I've never played the game! I have however kept up with a few videos coming out about the game over the last year. What an exciting prospect; bravo!
Damn you for beating Rollin's 56.86 before we released the video!! Also nice run by the way :) I have no idea about the viability of a D06 run with 2 laps intended way + 1 lap cut, what do you think about it?
@@linesight-rl ahah sorry about that :D In my opinion it would give a huge advantage to the AI to do that 3rd lap jump, it clearly doesn t need for the 2nd one but at least, since you have to release otherwise, the last jump is very doable for the AI and would add an extra 1 or 2 seconds to the ai
@@loupphok what do you mean by "doable by AI"? robot always do perfect inputs, yes, but with current approach discovery is questionable. getting 'some' jump is not a problem for AI; problem that human can understand 'oh I see i can make it eventually' while AI punished over and over at mistake and not progressing on discovery. So... in theory AI can discover jump but as DL engineer I feel it will be very unlikely. Especially other 'brother competitors' of current model will finish the map. Discovery of jump only can found by AI if there are no competitor model that progress further IMO. (like with gamma anywhere greater than zero ofc)
I have been watching this journey since you started posting these. I remember the first try, the model couldn't even finish the map. It's amazing to see the progress in the model and also how you have adapted to training the model. This is really freaking awesome.
Great vid! Your presentation has improved a lot and the idea of multiple cars of the same run is a BANGER idea! Can't wait for you to put AI to even more maps.
Fantastic video and astonishing AI! I didn't think it would get this good this quickly. Having one system capable of learning the different tracks is impressive. I'd love to see a single general model playing next. Im looking forward to the code release 👌
One of the main reasons I love this project so much is maybe one of the reasons you chose Trackmania. The deterministic nature of the mechanics is perfect for reinforcement learning and the progress is just so cool to watch. The game is already set up for viewing that progress perfectly. Watching AI behave in such a fluid way, navigating obstacles faster and more precise than human drivers is inspiring. The game is so competitive, so AI getting the W is just more impressive. This really showcases what we will see in the future with games. Imagine if you just created an environment and allowed the AI to figure out the best way to survive, navigate, handle different combat scenarios, etc. Imagine AI creating it's own movesets and combos. This is so cool! Thanks for sharing!
Please give us academics among your viewers a deep dive into the tye technical details. This is mind blowing cool and would love to know how you solved many of the problems
BIG respect for you making the AI open source (in the future). Thought the whole video till that point about how cool it would be if it was open source
Amazing work, your AI has come a long way and I'm excited to see how far you can push it, it is certainly exciting to see what different AI's are capable of and the future of AI's in general.
I really appreciate that you know what you're talking about compared to some of the new "ai creators" who are glorified prompt writers... that aren't even really good at prompts
Since you want to open this to the community, maybe it is worth to look into distributed machine learning: The learning process can be run on multiple machines in parallel and at any time the partial solutions can be sent to some master NN that combines all of them.
Very pleasant arrangement on the video. Nice presentation, segmentation, explanation. Well done! I've been following this for quite some time now. Let's see how far you can take this! PS: I totally dig the music!
Very interesting video, and looks like you had a lot of fun developing it! I loved the editing, it was very clean. I'm not sure it's fair to say the same AI beat humans 10-1, though - the way you're training it, it's effectively 11 different AIs competing, whereas the best human players will be similarly skilled across many tracks. If your 'general' racing AI manages to start taking records, though, then it's time to be scared...
@@-_-LouLou1234-_- That was listed as a next step. Not something that has been done. I think Linesight just wants Rastats to respond to the challenge being accepted, but IDK.
Your videos and those of other greats like Code Bullet have made the concept more clear to me, to the point that I now want to train an AI to make construction drawings of existing architectural models.
It’s crazy to think that if this game came it was released for the first time today, it would’ve been possible to learn all of these techniques from the AI in a matter of days, as opposed to the years that it took for them to be discovered by the community.
While I am rooting for you, I secretly hope humans still come out ahead. Their determination and skill truly show how some people are on a different level from the rest
What may reduce the computational effort even further (especially on the zero-shot AI) could be to scrape replays from leaderboards and store the state and input of each frame. Then assign a score according to each finishing time and train the AI on that dataset. As you wouldn't have to run the game loop, you could probably process a LOT more data simultanuously and more importantly converge faster to an almost record breaking skill level. To break the records you would then obv. have to let it play on the actual game again.
starting with the record route will allow for finding incremental improvements, but it would make it very unlikely to actually change the route. That world record route might be a dead end, and following it might be locking you out of improvements.
@@zaphod77 maybe, but the same could be said for any solution the AI finds (but yeah, it wouldn't make for a great video). But if you scraped and trained on hundreds of different replays on different maps it could probably generalize some trackmania strats and actually create a decent zero-shot AI.
Trackmania and computer science are evolving... First there were normal runs, then FWO joined the game and cuts got mainstream, then more and more tricks were found, kacky joined the game... Then everything evolved and now we have TAS, which is not that new, but it really got mainstream in TM too, and now I see AI evolving. I just love it.
i'm impressed by your skills, not the AI. as long as you have to go in and fix the AI to adapt to new challenges its not what was promised to humanity. youve just made a fancy TAS
You know what would be a cool project. You could set up some code to scrape all the world records for maps and used those records to generate your race lines, and deploy your code to some sort of compute cluster. Then you could have it automatically grind every WR
Around 18:50 where the AI is having trouble finishing the trail. I think you could change add something. 1. With gamma you could have the red line slowly disappear and the later reward system. 2. The line would start off not disappearing but once it started getting good times it could change 3. The longer it followed the red line the more it could get, but if it stayed the reward would disappear.
Impressive dedication and commitment to training the model on such a difficult task. I am impressed with how far you were able to improve it and all that was learned along the way. Well done!
Congrats on the awesome work! I follow your work as a fan of racing games and gamedev. So proud of the results, and to hear you are wanting to open source the code to get help from the community is just even more respectable!! THANKS!
i get in the end its a victory for ai on the first win. but seems unfair sense humans already beat the ai twice. but its a great video none the less. really fun seeing the process
I could not possibly be more impressed that your AI managed to learn wallbanging. This is just insane, learning a trick that requires extreme (by AI standards) investment with a high failure chance dependent on so many variables... That's amazing, it really is.
I haven’t played trackmania for a decade but I still like to watch the videos of it from time to time when RUclips graces me with a suggestion. I’ve been picking up AI recently for basic tasks but I think this puts it into perspective, the quote that in 24 hours the AI can learn something that humans took a decade to learn is crazy. The power of AI is going to shape the future and we should all be learning how to use it.
This was an awesome demonstration of ML - well done! You made really excellent use of Trackmania's ghost car system for showing the human cars versus AI. Keep it up! :)
The AI is cool, but showing the learning process with this visualization method is crazy cool. The info just comes across so naturally. And it goes hand in hand with the racing genre. This sort of visualization style could work for other games, like platformers. But I don't think it'd be nearly as effective for something like an FPS game, for multiple reasons
This was fascinating. Lots of admirable skills on display. I wish they would use similar techniques to solve efficiency issues in many industries. Great job.
So here I am, watching a video about a game I did not play and have no intentions to. But this is such a well explained and put together video. Great stuff!
Great video! I would love to watch more of each individual track breakdown, i really enjoyed that part and it's really cool to see the AI battle the human driver!
Childhood dram of mine was to figure out a system of partial differential equations that given the vector data of the track boundaries solves for all the sets of best trajectories. You training AI in this Track Mania is incredible!!!
You did not trained AI.
AI does not exist, stop using words you do not understand.
People will start to think YOU do not have "I".
Not the pin of shame lmao
The pin of shame always gives me a chuckle. You love to see it!
realistically we're talking semantics
I mean they are right, AI doesn't exist.
@@AnymMusic please, its 2024, say sepersontics
Putting Wirtual in the video with a swedish car is wild
Shots fired fr
😂😂😂😂😂 yes
I believe he'd lost a bet and had to use a swedish car for about month iirc
the disrespect 😂
Absolutely devious
I love the visualizations with multiple copies of a run. It's really awesome to see the differences in human vs AI racing lines.
yeah that was such a good way of visualizing WR times.
It was like a 3D racing line… never thought of something like that before
Agree, very interesting way of visualizing it. Would be cool if racing games had that.
Absolutely i was Really confused for a second then i saw the ai gain time and was like oooooohhhh
i never played this game but god all mighty the story telling skills of this community has me hooked as a spectator.
and i dont even like race games.
Legit rollercoaster level drama, cheating scandals, amazing achievements, and now the most impressive AI I've seen play a game. I agree one of the best parts is how well the community tells stories and gets us engaged when we don't even play the game. Probably my favorite reaction content too.
@@Brad-dx9fd if you want to see an even more impressive 'AI' Gaming bot - you should check out Seer from Rocket League. It even plays live against pro players and wins.
My thoughts exactly !
Tell me about it
@@greenzct9970 *takes a deep breath to start a lecture about pacing, story telling and clear and understandable imagery *
Something tells me Wirtual is not going to try and beat it
He's welcome to try, but the challenge seems even harder than Deep Dip 2 😁
@@linesight-rl would be great to see if the AI could figure out Deep Dip 1.... maybe having height be a reward?
@@linesight-rl okay now i'm curious whether an ai can climb deep dip
(it'll be harder to train it on the new trackmania though, wouldnt it?)
@@rasol136you would have to take the floor of a height modulo something (assuming they are equally spaced) as one big jump could ruin that model. Or maybe it just uses that as a shortcut 🤷
The AI does not work for tm 2020 tho@@rasol136
i love how humans still beat it on a01, truly shows how insane the a01 wr is!
for now…
😢we’re really approaching the end times istg.
@@boom-jr8vi yeah in a couple months its gonna compete with tas runs so a01 wr will be the warm up lol
Remember that this Ai has a brain of a fruit fly. Nothing compared to the AIs currently out there
@@tom_skip3523 i mean a brain of a fruit fly is pretty impressive since it can use it perfectly. Yeah there might be better ais but not in trackmania (atleast i think there aren't)
s4d is not a mechanic neural networks just learn to use. Pretty sure the model is not adapted to work with so many degrees of freedom.
Hey, quick thought on optimizing pixel rendering for AI vision. You mentioned that the AI has a relatively low resolution of images that it looks at to make decisions. What you could do is check the activation functions for each visual input and reduce the resolution on groups that have low activations and increase resolution on groups that have high activation. This is similar to selective focus in our natural vision where we may generally process the whole picture as we walk into a room, but may not notice that the bear on the table is holding a bottle of tequila until someone says "Hey, notice something strange about that bear?" Then, without moving closer to the bear, we look much more closely at the specific details to determine what we are looking at.
Here's a computational example:
Imagine the following 3x3 image (as text)
A B C
D E F
G H I
each frame will activate each pixel at a different amount, say 0 to 1. Here's an example activation array for the above:
0 1 0.5
0 0 1
0.2 0 1
Now, for the instances that have a low value, say less than 0.4, we will keep the resolution the same. But for higher activations, we may need to see more clearly so we have each pixel divide into another 3x3 that gives better resolution. In this case; B, C, F, I all appear to have high activation, meaning high importance, so we increase resolution to increase accuracy:
C now looks like this:
C.1 C.2 C.3
C.4 C.5 C.6
C.7 C.8 C.9
Same for all other pixel activations that are this high.
You can further subdivide as low as you needed to get the most detailed activations to get a pixel perfect view of the tight turn or barrier, without having to strain the model with extremely high activation rendering on every frame.
I don't know the exact architecture your AI is using, but this selective attention strategy works with most vision models.
Let me know what you think!
I am not very good at this kind of stuff, so don't take my word for it, but I think that might overload the ai.
Yea I feel like this would be great, especially since the screen more or less can look similar a lot of the time. Like the car is pretty much always in the same spot, taking up the same amount of space, which probably isn’t all the useful for the ai.
This sounds like a cool technique. Do you have any papers that apply this work? I've definitely seen the application of an attention mechanism, but never one that re-captures frames based on the attention weighting of the region
@@laurasisson1611 These people did something similar, and it seems like its working
The name of the paper: 3D CNNs with Adaptive Temporal Feature Resolutions
Sounds like you found some nice potential for improving the AI!
But just to understand the algorithm properly: the activation array would change dynamically during each step, which is the equivalent of „shifting focus“. But from a pc memory perspective (which he apparently already struggles with), would you have to save a full-scale image of each frame of the game to then be able to feed a small focus-part to the network? I think this might blow his current pc in terms of memory, even if the network itself only needs to be marginally larger (to now additionally cope with the focus area).
Please let me know what you think!
Needless to say the ai work is amazing, but I think this video is brilliant in many other ways. I think you’ve made the cleanest and easiest to understand visualisation of the learning, with the waves of different coloured cars, and the train of WR cars to beat. Fantastic job on both the technical stuff and the video production!!
I second this! I'm a machine learning researcher and the amount of effort you've put into not only getting good results but also saving the data and displaying the data of the runs in an easy-to-understand way is extremely commendable! Well done!
Can't agree more
The WR train was a great addition
I'd love to see even more details of the training, and full footage for each track. Hope we get to see a TAS challenge too, I love '92 Batman!
With comprehensible training videos like this, think of what a boon this will be to train human players to lose better to AIs!
Bruh... If I was picky I'd say that this is not exactly the same learning method, but I'll mark my words and eat my hat if you drive Deep Fear respawnless with AI v4.
You're probably safe: the AI has trouble with fullspeed maps as evidenced on A01.
But we'll definitely try!
@@linesight-rlhow come it struggles with full speed? As a human that naturally prefers the full speed elements of trackmania I feel like surely it should be easier because there’s less complex inputs? Just curious
@@tylercorrin1615 My guess based on the information given in this video, FS might be harder to train because of the low resolution screenshots and low fps the AI has available. If you look at 5:50, you can see how hard it is to see far ahead, which is critical in high speeds, because what is far ahead will be much closer much earlier in high speeds compared to low speeds.
@@vitriha3787 Maybe you could increase the resolution in the middle of the image
@@vitriha3787 I'm also going to bet that it's really hard for it to do accurate speed slides when the skidmarks are like 2 pixels wide.
One of the best videos I saw in a while
Thanks for the kind comment 🙂
i like how you even included the sections where humanity beats the ai
section
It was only 1 map lol
The AI just needed a bit more training on that map
I feel like it won't find speedslides without some human input, and I doubt it could win simply by training more.
@@lukeskyguy2238 Did it not find the speedslide by itself in the video? 😆
It would be really interesting instead of 0-shot learning to see how few-shot learning would work. Let the general AI fine tune on the new map for a few minutes to see how much it improves. Maybe we can see the AI play Random Map Challenge at some point?
having a 24/7 stream of ai on random map challenge would be so hype
@@StriderGW2 i would genuinely watch that every single second i could
i second this
AI cannot play different tracks on same neural network 😢
@@cyanhacker It can, but as you see the tchniques here applied are generally guiding the AI a lot and the neural network is small.
With techniques that give the AI abilities to look ahead, it could very well surpass and even plan its own route faster than humans. It woudn'T need to have human reqard shaping, insteadyou could employ PPO for optimizing the reward function along a general goal.
so props to the author but a fruitfly brain is notgoing to cut it. I would love to see a breakdown of the Neual network used currently and maybe with some added hardware getting that up to speed already beats the challenge
Amazing video and congratulations on creating such an amazing piece of of work(?). Really cool!
This is a (very time intensive) hobby 🙂
Honestly, a map like Deep Fear seems perfect for an AI. The only scary thing about it is its length.
For a human, the problem with Deep Fear is remembering the track, focus, precise input and reaction time. The AI has no issues with any of these things.
And I think the original commenter must have a misunderstanding about how it works, since they specify "respawnless," because there's no reason a fully trained version would need to respawn, it would just do the right inputs the first time.
The AI may have a problem with its context window if the track is very long and has many unique sections.
@@SamuelBoshier Its possible some sections are easier to do with the standing respawn, every checkpoint on the map is respawnable and has boosters, doing it segmented like that eliminates some variation. The AI could potentially learn to do it like that instead.
@@SamuelBoshier You sure it isn't you that has a misunderstanding of how the map works?
@@TheSuperappelflap Regarding length I'm referring to the massive amount of computation it requires. Nothing else really.
Trackmania AI youtube videos are the reason why I am so invested into learning Neural networking now. This amazes me. You're amazing. I'm already subscribed, but if I could, I'd subscribe twice.
You can...
Sub #2 = Patreon
But there's MORE!...
Sub #3 = GoFundMe - Contribute towards PC upgrades.
Sub #4 = Secret OnlyFans 🤫- Can't imagine of what 'Content' could possibly be posted here that's Trackmania themed? AI can. 😉
/s
Well. Its more like a genetic algorithm approach than a neural network application
@@pahom2 :-B well akshually
@@pahom2You do realise that what u said just makes you look silly, right? If you’d the slightest idea how this AI works, you’d know that it is definitely a neural network. The genetic algorithm bit is (one part of) how it’s trained, not how it runs. Saying it’s one thing or the other just doesn’t make any sense; it’s not even wrong, it’s just daft.
I love this video. Your editing and style is up there with old wirtual videos. So engaging and intuitive. Well done
Glad you enjoyed it!
I can't get over how good this video production is mate. Well done!
If this doesn't go viral, I'm going to be so sad. The AI just gets so interesting every time. Can't wait for the time that FWO itself will have trouble beating it.
They'll recruit the AI
I really doubt they'll recruit AI. Jealous as they are there's no way that they'd welcome someone better than them at the game.
I mean, how many years has it been since TAS entered the WR scene and they don't even consider recruiting him to the team? He's been crushing every map he plays and they keep denying him membership
@@daniellima4391had me in the first half ngl 😂
@@daniellima4391 Fr, my boy TAS was robbed.
He does other games too, and it's disgusting how communities ignore him
I mean they might like give a version of this an honorary induction to FWO. Perhaps the first version capable of trying its own short cuts
The field of reinforcement learning is so varied and has so much potential that has yet to be seen. Thank you for showing what a mid-size gaming rig can do and cutting down a little of the hype of scale we're seeing with AI right now. The next big breakethrough is in RL!
Honestly this is ASTOUNDING! This AI is in a league of its own. The wheel clip for the trial map was simply incredible. Well done!
Man, I only watched 10m and I am already amazing of the quality of the video and the quality of the AI. Holy shit, this is the best AI I saw not supported by money. So inspirational!
Man, incredible. GJ
I don't usually comment on videos, but the production and video editing quality on this one made me do so. Keep it up !! :)
Much appreciated!
As a developer, this video has me worried about my job security. You probably think I'm referring to AI taking my job, but no. I'm talking about people like Agade and pb4 massively raising the programming bar! I feel pretty bad putting my dopey little pet projects on a resume after watching stuff like this!
dont worry agi is coming in 6 years and we will all be retired to a life of leisure by our ai overlords
@@archsys307 On what basis do you say AGI will be here in 6 years? We haven't found anything that even remotely resembles AGI.
@@Justin73791 well it was a day like any other when i discovered AGI would arrive in 2030. i woke up at 5, went for a run, cold shower, alright team lifts. i stuck around to run some drills with coach after and then headed back for food. then i read 90 pages of the fifth book on gaius julius caesar i pirated this year and had half a pound of beef with half a stick of butter. alright then i went to my graduate topics in poptropical geometry class and dozed off, i already know the material anyways. walking back to my penthouse i stopped at a busy crosswalk and remembered that yielding is the motion the tao, so i yielded to my intrusive thoughts and walked across anyways (like that one scene in gattaca.) well i downed my 2 PM half gallon of milk and then checked if kai trump was 18 yet. damn not yet. hell of a swing though
in conclusion AGI is coming in 2030
@@archsys307but it's always 6 years in the future. :) in 6 years it will still be 6 years in the future. same with fusion always being 30 years in the future
@@Decodeish1 go look at current graphs of various AI progress metrics over the past decade
hardware is the growth determining factor right now… it could actually happen much sooner isnt that crazy
in any case go look at some predictions from smart people in the field
i saw a list, bunch of openai guys, other founders, ai researchers, etc all rather knowledgeable and intelligent people, 2/3 of them put AGI pre 2032 (in fact a ton in the 20s) and 90% of them by 2040
AGI as in rivaling an average, 50th percentile, skilled worker. nothing crazy, but it’s gonna be escape velocity
extrapolating from progress patterns of fusion or say SD cars doesnt hold up to the fundamentals graphs and the resounding, damn near unanimous predictions from experts
I feel like such AI can be both a good thing and bad thing for the world recording competition. It showed how some minor improvements can be made in places human players might not think of, but also it might led most people to just blindly copy the AI's approach, like how it went in Go. One of the biggest charm of watching world records run is the creative ways players figured out to gain even the thinest margin.
But anyway, this video is such a joy to watch, absolutely great work!
Honestly, I am interested in how AI can be used in TASing. TASing requires so many human inputs and knowledge of what saves time or at least thought possible so it can be attempted via TAS.
The thought of AI being a tool that can be trained to help test theoretically possible shortcuts or how fast something can truly be makes me wonder about how far it can be pushed.
Yes AI needs to be trained, and so still requires at least a standard level of the aforementioned knowledge, but still I think it could be a massive tool for TAS's. Especially of AI can be developed for other games where TASing is heavily reliant on human play + savestates due to a lack of other available compatible tools.
@Ezz_Fr That’s definitely true, but there’s a catch: AI can try more times in days than human can in years. When you know that someone plays the game so much more times than you do, I think it’s hard to think that you yourself can even find a better way with the your ‘super inefficient’ approach. After all, it takes too much time and luck for a human player to find, test, and execute an approach to perfection.
In the world of Chess AI has overtaken human for a long time now. AI vs AI is what generates progress and humans are studying AI tactics to learn new strategies. We are a long way from that in physics based games. Chess is just a discrete decision matrix. The video game genre that could probably be optimized the fastest would be 2D fighting. It's discrete.
It's also going to be incredibly cathartic to beat the AI
@@christopherlperezcruz1507 Any real-time player vs player mutiplayer game would get easily dominated by an AI simply because of inhuman consistency and reactions. Per example, all an AI needs to dominate in CS:GO is to be able to identify the enemy and click on their head as fast as possible since the AI will be able to kill the human before the human brain even has time to send the order to shoot back + some basic patroling to find the enemies and pathfinding to get to the bombsites when the bomb is planted. The only thing the humans can do to fight back at that point is try to coordinate with granades and stuff, but the AI could also easily learn to run away from explosive/fire granades and look away from flashbangs at the exact frame where they explode.
The only videogame genres where the humans could come close are either very complex turn based games where there is a lot of time to think about what to do and too many options for an AI to calculate everything even for a single turn, or time attack games like trackmania where the human can just keep farming the same map thousands of times to get close to the AI until they reach the human limit.
The AI absolutely fiending for the slightest instant gratification is so funny. You made an addict 😂
The day trackmania AIs can scout and lab tracks on their own, we're all cooked.
I mean technically it can, it discovered the barrier jump on its own. the issue is that the scoutting and labbing that it does is very very inefficient as it doesn't actually know what its looking for. a human could say "i think this might be faster" and try something new, but a neural net simply doesn't have the capability to do something like that, it can only make small adjustments.
Ngl didn't expect you here lol
i mean it would be awesome to use it for mappers to be able to automatically validate and set author times on the maps
@@eightheve Yah it can, you could just tell it to go on the path you're trying to use (red line linesight talked about) and see if it works
Deep Dip is Doomed
Me: would be interesti...
Linesight: in the future I could try TAS records
I could listen to you point out details of the AI learning tracks (as done in the first half) for hours. Great stuff.
The editing is amazing this time around! Crazy improvements
This was awesome, not just the AI, but also really well put together and presented. Having recently started video editing myself, this is so much above what I currently could reasonably do
Looking forward to your next video :)
Dude, you're killing it with this style! I love the balance between technical and simple. The editing makes it very engaging, but honestly I would watch it unedited. You're doing something never-before seen, and doing it very well. I love your voice, too. Great video!
Best trackmania video I've seen in a long time
Amazing job, huge effort and it's so nice to see how far your ai is coming. It's like watching your child learning to ride a bike or something
Maybe not something zero shot but more like how we currently use llms
Maybe have a model that is trained on a ton of different tracks(base model) and then fine tune it for the one it's currently attempting. This could lead to a large reduction in training time. Possibly even something use something like a Lora adapter so that only a few parameters have to be trained, and the rest can be frozen. This would mean the hardware requirements could be reduced while actually having a larger base model(no gradients and other training parameters in VRAM).
Just freeze everything except the last couple layers and then train on the track for a few iterations. This can be done easily in keras or pytorch or whatever hes using.
I was thinking similar thoughts how very inefficient all the Trackmania AI is when they're training from scratch every time they do a new track.
@@mithril_leaf well, that is how humans do it as well. People spend hours learning a track trying different lines and grinding. We just start out with more generalized knowledge and then take longer to optimize.
A human trying a track ten times will get a better time than an AI without prior knowledge but given enough attempts the AI will do better.
@@TheSuperappelflapthe problem is humans have general knowledge about how to play trackmania. Based on how he set up his AI, the AI does not have any transferable knowledge between courses. He is training it to follow the red line he already defined ahead of time for that specific track.
This approach is more akin to brute forcing a single map to find the optimal inputs, rather than creating an AI general knowledge for trackmania that can play on any track.
@@brunch. Right, this is not a generalized system. But it is a first step in that direction. The next step would be to take this system and use it as a component in a larger system that uses computer vision and some general knowledge about the limited amount of different blocks in the game to find the red line by itself.
This is one of the most interesting, well presented and entertaining videos I've watched on RUclips in a long time, and I've never played the game! I have however kept up with a few videos coming out about the game over the last year. What an exciting prospect; bravo!
I am speechless
@@ExhaustedPenguin I had to deeply think about, it was really hard but I managed to write it
hi speechless nice to meet u
Damn you for beating Rollin's 56.86 before we released the video!! Also nice run by the way :)
I have no idea about the viability of a D06 run with 2 laps intended way + 1 lap cut, what do you think about it?
@@linesight-rl ahah sorry about that :D
In my opinion it would give a huge advantage to the AI to do that 3rd lap jump, it clearly doesn t need for the 2nd one but at least, since you have to release otherwise, the last jump is very doable for the AI and would add an extra 1 or 2 seconds to the ai
@@loupphok what do you mean by "doable by AI"? robot always do perfect inputs, yes, but with current approach discovery is questionable. getting 'some' jump is not a problem for AI; problem that human can understand 'oh I see i can make it eventually' while AI punished over and over at mistake and not progressing on discovery. So... in theory AI can discover jump but as DL engineer I feel it will be very unlikely. Especially other 'brother competitors' of current model will finish the map. Discovery of jump only can found by AI if there are no competitor model that progress further IMO. (like with gamma anywhere greater than zero ofc)
I have been watching this journey since you started posting these. I remember the first try, the model couldn't even finish the map. It's amazing to see the progress in the model and also how you have adapted to training the model. This is really freaking awesome.
Well done, that's insane progress. Can't wait to see how good it will get at reading new maps
Damn, you are the first creator I see that manage to build an AI that can beat human across multiple map, this is so cool
Great vid! Your presentation has improved a lot and the idea of multiple cars of the same run is a BANGER idea! Can't wait for you to put AI to even more maps.
Amazing job, I loved the editing :D 14:10, and was surprised at the significant gain on D06 with just pure driving skill.
Fantastic video and astonishing AI! I didn't think it would get this good this quickly. Having one system capable of learning the different tracks is impressive. I'd love to see a single general model playing next. Im looking forward to the code release 👌
One of the main reasons I love this project so much is maybe one of the reasons you chose Trackmania. The deterministic nature of the mechanics is perfect for reinforcement learning and the progress is just so cool to watch. The game is already set up for viewing that progress perfectly. Watching AI behave in such a fluid way, navigating obstacles faster and more precise than human drivers is inspiring. The game is so competitive, so AI getting the W is just more impressive. This really showcases what we will see in the future with games. Imagine if you just created an environment and allowed the AI to figure out the best way to survive, navigate, handle different combat scenarios, etc. Imagine AI creating it's own movesets and combos. This is so cool! Thanks for sharing!
Please give us academics among your viewers a deep dive into the tye technical details. This is mind blowing cool and would love to know how you solved many of the problems
BIG respect for you making the AI open source (in the future).
Thought the whole video till that point about how cool it would be if it was open source
It's now fully open-source :)
github.com/Linesight-RL/linesight
@@linesight-rl lets goo! Great job dude :)
Amazing work, your AI has come a long way and I'm excited to see how far you can push it, it is certainly exciting to see what different AI's are capable of and the future of AI's in general.
I really appreciate that you know what you're talking about compared to some of the new "ai creators" who are glorified prompt writers... that aren't even really good at prompts
Started watching your videos for the AI, but god damn your editing is also super good!
Your editing skills are top-notch. Loved the transitions!
Since you want to open this to the community, maybe it is worth to look into distributed machine learning: The learning process can be run on multiple machines in parallel and at any time the partial solutions can be sent to some master NN that combines all of them.
Very pleasant arrangement on the video. Nice presentation, segmentation, explanation. Well done! I've been following this for quite some time now. Let's see how far you can take this!
PS: I totally dig the music!
Very interesting video, and looks like you had a lot of fun developing it! I loved the editing, it was very clean. I'm not sure it's fair to say the same AI beat humans 10-1, though - the way you're training it, it's effectively 11 different AIs competing, whereas the best human players will be similarly skilled across many tracks. If your 'general' racing AI manages to start taking records, though, then it's time to be scared...
That's a fair point you're bringing up. It's something I'd like to work on in the future.
the world records are by different people also. and also people actually have to train maps too when they start hunting world records on them
Visualization is so top-notch. A lot of good world class engineers fail to present their work; but you did amazing. P.S. DL programmer myself.
I'm waiting for Rastats' comment 😏
A hat needs to be eaten
But did you show the deep fear completion, for real I haven't seen it
@@-_-LouLou1234-_-
That was listed as a next step. Not something that has been done.
I think Linesight just wants Rastats to respond to the challenge being accepted, but IDK.
go forth, linesight. make a man eat a hat
I responded in another comment unreal improvement on the AI !
This video and the whole concept is so incredible, instant subscribe. I've been mindblown
Now for the ultimate test: unleash it on Deep Dip!
it currently cant run in TM2020 since it uses TMinterface which is a tool only available in TMNF.
@@poruzu That's a shame, I'm just picturing a cascade of AI cars falling everywhere
This was a very nice and interactive watch ! Thanks for sharing and please keep progressing the A.I. :)
Your videos and those of other greats like Code Bullet have made the concept more clear to me, to the point that I now want to train an AI to make construction drawings of existing architectural models.
3:33 this is an amazing way to visualize that
It’s crazy to think that if this game came it was released for the first time today, it would’ve been possible to learn all of these techniques from the AI in a matter of days, as opposed to the years that it took for them to be discovered by the community.
dropping this well deserved comment for you guys, excellent work, thank you for sharing it with us all
what a great video. excited for the next installment
While I am rooting for you, I secretly hope humans still come out ahead. Their determination and skill truly show how some people are on a different level from the rest
phenomenal work! wow! I'm really excited to see how the general model will be able to handle brand new tracks once it is trained further!
What may reduce the computational effort even further (especially on the zero-shot AI) could be to scrape replays from leaderboards and store the state and input of each frame. Then assign a score according to each finishing time and train the AI on that dataset. As you wouldn't have to run the game loop, you could probably process a LOT more data simultanuously and more importantly converge faster to an almost record breaking skill level. To break the records you would then obv. have to let it play on the actual game again.
starting with the record route will allow for finding incremental improvements, but it would make it very unlikely to actually change the route. That world record route might be a dead end, and following it might be locking you out of improvements.
@@zaphod77 maybe, but the same could be said for any solution the AI finds (but yeah, it wouldn't make for a great video). But if you scraped and trained on hundreds of different replays on different maps it could probably generalize some trackmania strats and actually create a decent zero-shot AI.
Trackmania and computer science are evolving... First there were normal runs, then FWO joined the game and cuts got mainstream, then more and more tricks were found, kacky joined the game...
Then everything evolved and now we have TAS, which is not that new, but it really got mainstream in TM too, and now I see AI evolving. I just love it.
i'm impressed by your skills, not the AI. as long as you have to go in and fix the AI to adapt to new challenges its not what was promised to humanity. youve just made a fancy TAS
3:25 this is a great graphic
Oh no, ai is even taking over tm now..
AI brilliance aside, this is a really well put together video. Subscribed!
OMG am an IWO memeber too
You need to beat a cut WR with the intended way, membership is probably even harder than FWO 😅
You know what would be a cool project. You could set up some code to scrape all the world records for maps and used those records to generate your race lines, and deploy your code to some sort of compute cluster. Then you could have it automatically grind every WR
Found the Big Tech company alt account.
How are you so good at Trackmania, AI, AND video editiing????? WHAT A GUY!
15:39 BRAZIL!
10:43 if we didn't had those 10 years, would you still be able to train it to do those things?
yep, he didnt train it to do anything but get to the end as fast as possible
"It can't see the small gap" Well yeah, you gave it a miniscule resolution to work with
Around 18:50 where the AI is having trouble finishing the trail. I think you could change add something.
1. With gamma you could have the red line slowly disappear and the later reward system.
2. The line would start off not disappearing but once it started getting good times it could change
3. The longer it followed the red line the more it could get, but if it stayed the reward would disappear.
Impressive dedication and commitment to training the model on such a difficult task. I am impressed with how far you were able to improve it and all that was learned along the way. Well done!
dude
you put so much work into making this a good watch and it shows
Congrats on the awesome work! I follow your work as a fan of racing games and gamedev. So proud of the results, and to hear you are wanting to open source the code to get help from the community is just even more respectable!! THANKS!
its open sourced by now
This was incredibly done! You have incredible talent and ingenuity, this was entertaining to watch.
i get in the end its a victory for ai on the first win. but seems unfair sense humans already beat the ai twice. but its a great video none the less. really fun seeing the process
Been following your journey, it's absolutely mesmerizing work, very impressive and inspiring :)
Really excited for your progress. Keep it up mate.
This is amazing man, I really can’t wait to see how your AI will be the next evolution in tricks for Trackmania
the visualization with the numbers was really helpful for understanding
I could not possibly be more impressed that your AI managed to learn wallbanging. This is just insane, learning a trick that requires extreme (by AI standards) investment with a high failure chance dependent on so many variables... That's amazing, it really is.
WOW!!! Very interested to see this and your upcoming projects!! And great video production too
I haven’t played trackmania for a decade but I still like to watch the videos of it from time to time when RUclips graces me with a suggestion.
I’ve been picking up AI recently for basic tasks but I think this puts it into perspective, the quote that in 24 hours the AI can learn something that humans took a decade to learn is crazy. The power of AI is going to shape the future and we should all be learning how to use it.
Congratulations, when setting gamma to 0.95 you've created AiDHD
This was an awesome demonstration of ML - well done! You made really excellent use of Trackmania's ghost car system for showing the human cars versus AI. Keep it up! :)
Amazing story telling! I can't wait to see this his next project!
The AI is cool, but showing the learning process with this visualization method is crazy cool. The info just comes across so naturally. And it goes hand in hand with the racing genre. This sort of visualization style could work for other games, like platformers. But I don't think it'd be nearly as effective for something like an FPS game, for multiple reasons
Impressive, hats off to all the brilliant work and this super elaborated video
This is an amazing video! Great story telling and pacing. This should have way more views.
I have nothing to say But this was Awesome so here's my comment for comment engagement. i hope you can go super duper viral.
Thanks 🙂
@@linesight-rl I'm watching again
This was fascinating. Lots of admirable skills on display. I wish they would use similar techniques to solve efficiency issues in many industries. Great job.
So here I am, watching a video about a game I did not play and have no intentions to. But this is such a well explained and put together video. Great stuff!
The video is made so good i love it good work man❤
Great video!
I would love to watch more of each individual track breakdown, i really enjoyed that part and it's really cool to see the AI battle the human driver!
Childhood dram of mine was to figure out a system of partial differential equations that given the vector data of the track boundaries solves for all the sets of best trajectories. You training AI in this Track Mania is incredible!!!