AWS DeepRacer : A Practical Guide to Reducing The Sim2Real Gap — Part 2 || Training Guide

To train the AWS DeepRacer to navigate safely around a track without crashing, it is crucial to select an appropriate action space, reward function, and training paradigm. Start by using a discrete action space with limited steering angles and throttle values. Design the reward function to encourage the car to stay on the track, slow down for turns, and avoid veering off-track. Consider rewarding proximity to the track's center line and penalizing all-wheel excursions. To prevent unintended zigzagging, incorporate a penalty for extreme steering angles. Iteratively train the model by cloning and improving the best performing version, gradually reducing the learning rate to fine-tune its performance. Switch between clockwise and counterclockwise track orientations to minimize overfitting. Aim for a consistent reward graph, even if 100% completion is not always achieved. By following these strategies, you can train a DeepRacer model that can reliably navigate a track without crashing.