Imitation Learning

Our first attempts to train a self-driving agent to predict the best course of actions were based on imitation learning. Imitation learning is a supervised machine learning technique in which a labeled training set is generated by a human controller giving inputs to the system. To create our dataset, we drove the agent around the AirSim City environment, recording images and an action log. A CNN model is built for the agent to predict the actions by taking the snapshots of images as input in the environment.