Off-Road Navigation With End-to-end Imitation Learning for Continuously Parameterized Control

Control of autonomous ground vehicles by deep neural networks has been an area of interest for over 30 years. However, the majority of this research has been directed to-wards urban navigation and traffic control. We instead attempt to develop an end-to-end controller for off-road navigation using a convolutional deep neural network. In pursuit of this, we make use of state-of-the-art machine learning algorithms to train a control policy using data collected from a limited number of human demonstrations. We evaluate a variety of network structures, primarily focusing on methods of goal parameterization within a continuous state space. We demonstrate the effectiveness and limitations of these methods in the real world using a 1/6 scale truck outfitted with an RGB-D camera and GPS sensor.