This dissertation explores visual perception within the game of autonomous racing.Pocket Racer, is a scaled, high performance vehicle designed to learn to imitate the behavior of the world champion driver with vision based internal models, similar to that of the dragonfly when exhibiting predictive prey interception behaviors. This algorithm, Race-Net, predicts and targets future vehicle trajectory states with a combination of offline deep supervised learning and predictive forward inverse modelling. Through Race-Net, we attempt to demonstrate super human autonomous driving capabilities on our Pocket Racer platform.
Pocket Car, our educational outreach platform, is designed for STEM education. Versatile while being low-cost and accessible, Pocket Car is designed with a wide assortment of sensors. Pocket Car is capable of learning to mimic driving behaviors and also compute state estimation for control via Adaptive Monte Carlo Localization (AMCL) using cones as landmarks. Alternatively, it is equipped with a low cost 2D LIDAR for Simultaneous Localization and Mapping (SLAM), applicable for competitions involving school hallways. We demonstrate the educational merits of Pocket Car through our successful collaboration with the IEEE undergraduate club at UCLA.