Sep-19-2020, 03:45 PM
Hi,
Over the past couple months (actually year), I was working on a car game for reinforcement learning, called CaRL. Since this project is in a very early release stage, I would be thankful for stars on my GitHub page:
https://github.com/MatthiasSchinzel/CaRL
With this game I want to bring reinforcement learning algorithms closer to real applications - making it easier to train your RL algorithm for a self-driving car, be it an RC car following a road or maybe one day a real car.
So, if you check out the GitHub, you will see it is actually just a black and white road. The reason for the high abstraction level is, to make it kind of “domain agnostic”. It is of course not really domain agnostic, but all you need is something to extract the ego road lane. This can be done with for example a UNET of you can find the lane markings on the left and right and construct the ego lane from that. I even made a proof of concept of the above Idea: I let my trained SAC steer a car in GTA 5: https://github.com/MatthiasSchinzel/Soft...laying-GTA
Another reason for this high abstraction level: it makes it really easy to make your own tracks. Use paint, download real roads from google maps, you name it! And as we all know: the more data the better.
One last thing about the abstraction: In my opinion, creating a 3D game, which is looking as realistic as possible is far more difficult, than annotating roads from images and train a segmentation algorithm on it.
Since RL is so far mostly a R&D topic, I would be happy if researches would use this! If you are a student or someone working on RL, please also give me feedback!
Thank you!
Over the past couple months (actually year), I was working on a car game for reinforcement learning, called CaRL. Since this project is in a very early release stage, I would be thankful for stars on my GitHub page:
https://github.com/MatthiasSchinzel/CaRL
With this game I want to bring reinforcement learning algorithms closer to real applications - making it easier to train your RL algorithm for a self-driving car, be it an RC car following a road or maybe one day a real car.
So, if you check out the GitHub, you will see it is actually just a black and white road. The reason for the high abstraction level is, to make it kind of “domain agnostic”. It is of course not really domain agnostic, but all you need is something to extract the ego road lane. This can be done with for example a UNET of you can find the lane markings on the left and right and construct the ego lane from that. I even made a proof of concept of the above Idea: I let my trained SAC steer a car in GTA 5: https://github.com/MatthiasSchinzel/Soft...laying-GTA
Another reason for this high abstraction level: it makes it really easy to make your own tracks. Use paint, download real roads from google maps, you name it! And as we all know: the more data the better.
One last thing about the abstraction: In my opinion, creating a 3D game, which is looking as realistic as possible is far more difficult, than annotating roads from images and train a segmentation algorithm on it.
Since RL is so far mostly a R&D topic, I would be happy if researches would use this! If you are a student or someone working on RL, please also give me feedback!
Thank you!