Learn Driving Policy from Pixels
Imitation learning driving policies from image observations
About
We formulate an autonomous racing game with discrete action space as a classification problem and use the imitation learning technique to solve it. The robot has raw images as observations. We design a feature extractor that can extract edge information from the pixels. After feature extraction, we perform principal component analysis (PCA) for dimensionality reduction. As a result, the extracted feature vector is in a much lower dimensional space, is robust against the background colour, and can generalize to unseen datasets.
The project was along with the course RO47002 Machine Learning for Robotics (instructor: Prof. Jens Kober) at the Delft University of Technology. The final grade for this course was 9.1/10 (top in the class, no concrete statistics).