This project provides helping scripts to automate the training of DeepRacer models using the official AWS DeepRacer console for re-inforcement learning.
One big problem with Deeracer is overfitting to the virtual environment. This can be prevented by randomizing the world (textures and tracks). While manipulating the environment in a local training setup is rather easy, doing this in the official console is only possible by modifying the official deepracer simulation application. A set of scripts for doing this is provided in this repository.
Here are some impressions:
- Start a new DeepRacer job
- Get the Robomaker ARN from the Robomaker Console
- execute ./increase_speed.py <robomaker arn> <time in minutes> <percentage increase> (e.g. ./increase_speed.py "arn:aws:robomaker:us-east-1:000000000:simulation-job/sim-6z3jfvryz3dh" 120 1.10)