search

araffin / learning-to-drive-in-5-minutes

Implementation of reinforcement learning approach to make a car learn to drive smoothly in minutes
https://towardsdatascience.com/learning-to-drive-smoothly-in-minutes-450a7cdb35f4
MIT License
284 stars 88 forks source link

[Question] What features are missing to make it possible to train the car in level 1 without using teleop? #19

Closed FrancescoTerrosi closed 4 years ago

FrancescoTerrosi commented 4 years ago

As the title says:

I would like to expand this project in order to make it possible for the car to learn by itself in every environment, could you please tell me why this is not permitted right now? thanks

araffin commented 4 years ago

Hello,

This comes from the simulator: it does not provide a distance to the center of the road, so you cannot automatically stop the episode and must use teleop.

See here in the code (cross track error)

FrancescoTerrosi commented 4 years ago

Yes I read that comment and it confuses me because in the repo's README you're saying that cte works only on level 0 but I understand that comment as if cte is available in all the 3 tracks. Also I'm not getting any error so is it actually implemented for now? Last question: does the simulator provide the distance for track 0 and not for the others? Thank you :)

araffin commented 4 years ago

Also I'm not getting any error so is it actually implemented for now? Last question: does the simulator provide the distance for track 0 and not for the others?

It is not, it just returns zero all the time for other levels. But this is not related to this repo.

FrancescoTerrosi commented 4 years ago

Okay thank you, I thought you were just missing some preprocessing or stuff like that.

Cheers