duckietown / gym-duckietown

Self-driving car simulator for the Duckietown universe
http://duckietown.org
Other
45 stars 16 forks source link

Domain Randomization API #133

Closed bhairavmehta95 closed 5 years ago

bhairavmehta95 commented 5 years ago

Switching to config files for domain randomization.

Some of the graphics stuff will need some reorganizing, but this is a pretty basic addition - looking for feedback. If we like this, we can merge and I can do the texture-based randomization the same way.

@liampaull

liampaull commented 5 years ago

@bhairavmehta95 have a look at the results of the circle ci test? is something missing? Otherwise I'm basically fine to merge but we should write down somewhere, either in here or in the docs, what are the variables that you can randomize and what the numbers mean. The API looks good though

bhairavmehta95 commented 5 years ago

Yeah so the CircleCI tests fail because of a variable reference:

https://github.com/duckietown/gym-duckietown/blob/master/learning/reinforcement/pytorch/train_reinforcement.py#L64-L84

It says that the variable is unreferenced, but it gets set before it actually would ever step into that if statement. (aka it's fine)

And yes - this PR was more about getting comments on the API - I can write up a little README to put inside the randomization/ folder.

Unfortunately though, while we can qualitatively describe what each variable does, certain things are just uninterpretable. i.e the global light position will always be a 3D -point in space, and there's not an intuitive range that says "this range of 3D coordinates makes sense while this other range doesn't."

I'll write something up though, atleast as a start.

liampaull commented 5 years ago

ok sounds great