r7vme / learning-to-drive-in-a-day Goto Github PK

Hi @r7vme,
I have a quick question about controller.py. Can you please say what is the goal of this script? I am a bit confuse about the overall approach. I understand that the Gym environment has been modified such that the input is the latent representation (coming from VAE) of each image (and not the raw image), but the rest of you approach is still unclear. Pluse, It seems you save the raw images in a buffer, but why, when you have a latent representation. Would you please clarify a bit?

Hi,
Great work.
A quick question- How did you train VAE? I mean which data-set has been used to train VAE?

Hello,

Thanks for this nice project (and kudos for using stable-baselines ;) ).
I tried to run your code but I have some issues:

With docker:

./run-in-docker.sh 
starting DonkeyGym env
Missing DONKEY_SIM_PORT environment var. Using defaults
donkey subprocess started
binding to ('0.0.0.0', 9090)
waiting to load..
No protocol specified
No protocol specified
Error initializing Gtk+
No protocol specified
waiting to load..

I have intel graphics but the problem is still here even when removing device=/dev/dri:/dev/dri

In a conda env:
Everything starts properly except that I got this error message in the terminal:

Expecting property name enclosed in double quotes: line 1 column 67 (char 66) failed to read json  {"msg_type":"telemetry","steering_angle":0,"throttle":0 ...

Trying the scripts from the sdsandbox repo, I ran into the same issue (see https://github.com/tawnkramer/sdsandbox/issues/11), so I assume the problem comes from my Unity version that may be too recent.
What version are you using?

System Info:

• python 3.6
• Unity-2018.2.7f1
• requirements installed via pip -r requirements.txt

Right now policy learning process is vague and does not improve over time (i.e. good policy can be learned in 3 episodes, but after 10 episodes policy can degrade completely).

So after spending time adjusting optimal buffer size and optimization steps, i see that it's pretty random. I assume that prioritized experience buffer for DDPG will help imrove situation.

In short, it will make sure

• to sample unseen observations (by using infinite priority)
• to sample "valuable" observations (by computing priority based on TD error)