Really impressive results. I wonder if through in a wide array of different tracks, will it generalize and become a full player? Could it learn to use items and play against cpu as well?
My next video will be looking to see answer many of these questions! Using items is a tough one though, as it requires an extra action for the agent, and is also so dependent on what item the AI has, compared to the other actions which generalise across all situations.
@@pango8143 A fascinating question to be sure. I suspect that the COMs nintendo came up with don't actually drive like the AI in this video does. What I suspect they did is have the COMs get pulled along on a semi preset rail path with a little random variation thrown in to make it look like they are driving. I suspect item behavior is handled by a fairly standard finite state machine that chooses what to do based on current race conditions such as the place the COM is in, what item they currently have, and some other environment factors such as how many other players are around them. It's very likely just very clever smoke and mirrors that looks like driving to the player but isn't actually driving.
The first of those two corners you mentioned in around the middle of the video has been the toughest for me. The steep bank angle combined with the sharpness of the turn means that if you accidentally hit the barrier, you can easily bounce and be launched cleanly off the track.
I know it would be a nightmare to code, but it would be really cool to see the AI assign a difficulty map to the course. So we can see what it "thinks" is the most difficult for it to learn
That would be interesting, but I'm not even sure how to go about doing it! I feel like currently the best metric for that is just how long it takes to learn the track, since it definitely takes a very different amount of time to learn different tracks. In my newest versions of this project, it can take under 20 hours to learn a track like Luigi Circuit, but potentially hundreds of hours to learn rainbow road, even with multiple starting locations
@@aitango My best idea for it would probably be having it color map it over the course where the color blocks would be put between the lap counting checkpoints. SummoningSalt's videos have a good example of what I'm talking about with the lap counting checkpoints
Congrats on embracing and producing what I consider a new art form. Videos such as this merge AI learning and Presentation into a unique art. Your skill, knowledge, and of course TALENT are amazing. Thank you!
This was so cool! I think this was my favorite ai of all your mario kart ai's, because of how it learned to drive, not just learn the track exactly. So in some places it would drive differently if it messed up to stay on the track. Pretty cool stuff.
@@pokkohopper not at this particular section. It probably recognises what that ramp looks like and learns how to perfectly execute the low jump and since its not human precision isnt an issue. A good ai should learn different ideas for different situations
@@aitango just in case u forgor u know u can speed up dolphin by 2x right needa use that overload cpu thing that's how malleo drove backwards for 10000 laps on gv2
Yoo I just saw ur other video and subbed yesterday lol Damn this would've been especially hard since this map is like the definition of a noisy input. Also was that a low trick at 4:35 lol
Thanks for the sub! Yeah very noisy, the floor is a million different colors and there's random bits of rock floating everywhere so make things worse. Somehow yes haha
I loooooove rainbow road in mkwii and I really think it's the peak of all rainbow roads. So many memories of being a smaller kid, playing rosalina, and doing an all out bagging strategy while my brother would actually play the track right
It was my favourite track when I was younger too! I remember playing it with my parents and forcing them to play on this track despite them hating it haha
You know, I wonder how much the training would go in these types of scenarios if the ai had 2 frames to look at instead of what I presume is just one. After all, it's kind of difficult to guess the speed and direction you're going from a single frame.
What values does the AI have access to in each pixel? If it's just given the raw RGB, that might be hurting its performance. Giving it a brightness variable as well would probably help! On the topic of how it sees, though I'm not sure if this would count as cheating to you, it might be worth running an edge detect filter and also giving it that data in some way.
The AI has access to a slightly different version of the raw pixels, such that the resolution is far lower and instead of RGB the image is greyscale. This AI currently using 3 convolutional layers, which actually allows the AI to learn its own edge detection
the rainbow road fast staff ghost was by far the hardest for me, however, i managed to beat it with the flame runner equivalent kart (cant remember name but funkys is green with flames) hardest part imo is moonjump/that jump cut after the figure 8 section
Currently the bottleneck is the frame capturing / image processing. This is pretty unique as I had to find a way to allow my code to interact with dolphin
@@aitango Do you have intentions to share more of the technical details of this project regarding interfacing with dolphin etc? (Perhaps open sourcing at some point?) Would be very interesting to read.
You can see it drives off a corners because it thinks there's a sexy boost panel on the walls, but that's only true some of the time. It's also funny trying to see it go for a path that it doesn't have access to because it would be a faster path if it could.
I've typically stuck to Wii games since I have a setup which can run pretty much any Wii game. Maybe at some point I'll have a look at games on other platforms, but it would probably take me weeks or months to get initially setup!
i have no idea about how exactly this stuff works but would it be possible to tell the AI to "stop learning by trying" and to "play it safe" at some point to make him complete a lap without crashing? cause as i understand it the AI trys to constantly find ways to go faster and therefore sometimes try things that dont work even though it already learned inputs that do work for a certain part.
It might not have been done in this video, but in the new one with CPUs, it's explained that different things have different weight values for reward or punishment. At one point during that one, the AI started finishing the laps the long way instead of only going for the shortcut before figuring out the shortcut completely.
2:56 It really looks like your way of training the AI as shown in the previous video is inherently flawed; using merely graphics and nothing else, as well as overlaying lines to "reward" the AI on the map, the moment the track overlaps with itself, the AI gets confused, as seen here. Really hope this is solved in future tests.
