Learning a CNN-based End-to-End Controller for a Formula SAE Racecar

Convolutional neural networks (CNNs) have had much success in the past decade on a many complex vision tasks, successfully passing human benchmarks on object recognition, image segmentation, and video classification [1, 2, 3]. Much of the success of CNNs can be attributed to their ability to automatically learn feature maps and scale with large datasets [5]. More recently, deep CNNs have been applied to learn control algorithms associated with steering a regular street-legal sedan [6]. In this report, we aim to adapt the work of Bojarski et al. for the specific purpose of learning a complete drive controller (steering, brake, and throttle) for a Formula SAE racecar navigating a pre-defined cone-delineated track setup.