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
287 stars 87 forks source link

learning-to-drive-in-5-minutes on real RC car [question] #39

Closed MariaPiaGelos closed 2 years ago

MariaPiaGelos commented 2 years ago

Hi !

I want to apply your repo on a real RC car. Do you have any suggestions to do this? I have to shrink the VAE model ?

Thanks!

araffin commented 2 years ago

Hello, normally there is not much to change, you mostly need to write the adapter to connect to the real car. For reference, I used an auto-encoder with latent space of size 32 in the video here: https://www.youtube.com/watch?v=f_FmDFrYkPM

You also probably need to use a smoother exploration (link is also in the description of the video). For a more recent version of the auto-encoder code (in pytorch): https://github.com/araffin/aae-train-donkeycar

I would also recommend you to switch to Stable-Baselines3 (gSDE is included): https://github.com/DLR-RM/stable-baselines3