Trackmania AI using upcoming road information

[LaurensNeinders]

Hi,
This is a project where an AI was trained on information about the track ahead.
The first link has the files needed to test the AI yourself.
The second is a paper that gives more information about the thought behind this AI.
Trackmap tmrl files.zip
Paper.pdf

And here a video of the AI driving the tmrl-test track
https://youtu.be/SSabAy9nDeU

And finally here is the repository:
https://github.com/LaurensNeinders/tmrl-trackmap

[yannbouteiller]

Hi!

I have read your paper, congrats, this is cool work!

I have a question: in Fig18 (Appendix) you show how your additional input works at a conceptual level, but in practice how do you process this? From this figure it looks like a 2D patch that you would feed to a CNN sort of thing? Thanks!

Do you wish to upload your video on YouTube for the leaderboard, since the WT link will expire I suppose? (I can do it if you want)

Do you think you could package your branch as a standalone library such that it works out-of-the-box? There is no setup.py atm (I can also probably do it if you want)

(Also if you are going to present your paper, I spotted a couple typos in the Appendix)

[yannbouteiller]

So, theoretically you would need to clone the repo, install it locally in developer mode on your machine by executing pip install -e . where the setup.py file is, and then you could directly modify your local copy of the repo.

[NDR008]

I do indeed flatten the coordinates. however I don't use the LIDAR observation as input anymore, I do use some other variables like: speed, gear, rpm, acceleration, steering_angle, slipping_tires.

I cloned the repo, but then wasn't sure how to properly use it 😅 . So I ended up making changes in site-packages and then copying them to the repo.

Damn... that is something I was planning to look into in my work... been painfully trying to reverse engineer GT1 to find most of the variables you mentioned

[LaurensNeinders]

So, theoretically you would need to clone the repo, install it locally in developer mode on your machine by executing pip install -e . where the setup.py file is, and then you could directly modify your local copy of the repo.

Ah, that sounds a lot better way of working on such a project.
For now, do I just need to make a resources zip file (possibly make a release) and edit the link to it in setup.py?

[yannbouteiller]

So, theoretically you would need to clone the repo, install it locally in developer mode on your machine by executing pip install -e . where the setup.py file is, and then you could directly modify your local copy of the repo.

Ah, that sounds a lot better way of working on such a project. For now, do I just need to make a resources zip file (possibly make a release) and edit the link to it in setup.py?

You can simply make a resources.zip file, put it in the same directory as your setup.py and open it in setup.py instead of downloading the one that is in the tmrl repo.

[LaurensNeinders]

I pushed these changes to the repository. I made a resources.zip and changed setup.py

[NDR008]

I definitely want to try this out on my Gran Turismo work.

…

On Wed, 22 Feb 2023, 17:34 Yann Bouteiller, ***@***.***> wrote: So you get a set of 2 * 15 = 30 car-centric 2D coordinates, right? This, I understand from figure 18. Then, do you flatten these coordinates into a vector of 2 * 15 * 2 = 60 numbers, append this vector to the LIDAR observation, and feed this to the default MultiLayer Perceptron used by tmrl for LIDARs? If you cloned the tmrl repo, your local version of the repo should already be almost packaged as a standalone library. That would just be a matter of modding setup.py such that it uses your files instead of the content of resources.zip Unless instead of cloning the repo locally you installed it from pypi and modded it directly in your site-packages folder, which is not how pip is supposed to work 😅 — Reply to this email directly, view it on GitHub <#37 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AIKWRWURJLFJS26E7PAIRUTWYY5YTANCNFSM6AAAAAAVDHX2M4> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[yannbouteiller]

Your submission has been evaluated on 10 runs, its mean performance was 38.737 with a standard deviation of 0.398 :)

Here are the replays

It might be slightly under-evaluated because I got about 2-3 of these yet unexplained timestep timeouts of ~0.25s per run...

Updating the leaderboard shortly

[yannbouteiller]

Done! Congrats, you did set the level of this "off"" competition quite high :)

[NumseBacon]

I kinda feel like the training time should be stated? Or is there a reason for this? @yannbouteiller

[LaurensNeinders]

Indeed quite a lot, I'll let you know when I have retrained the model.

[NumseBacon]

awesome, thanks

[NumseBacon]

perhaps making a wandb and saving it could also help with viewing statistics?

[yannbouteiller]

tmrl already logs to wandb in a public project by default

[NumseBacon]

tmrl already logs to wandb in a public project by default

Damn one has been running for 14 months and one called test1 has 2,5 months