A simple resnet-18 model was trained in pytorch over ~20 GB data to learn and map the actions of the agent. I used a stratified k fold strategy and ran 5 folds on the data to help with class balancing and making predictions more robust. Images were flipped from Left to Right to double the image dataset.