An Approach for Reliable End-to-End Autonomous Driving Based on the Simplex Architecture

Over the past decade, autonomous driving has been a subject of continued interest for research. In general, conventional approaches for autonomous driving consists of roughly two parts: Perception and motion planning. Recently, an alternative approach based on the deep neural network has been developed, called the end-to-end autonomous driving, that maps raw sensor data directly to driving command without requiring a separate perception process. However, the performance of the end-to-end driving highly depends on the quantity and quality of the datasets used in the learning process and can become unreliable if untrained situation is encountered. To overcome this fundamental drawback of the end-to-end approach, we adopt the simplex architecture for autonomous driving as a mean that combines the end-to-end approach together with the conventional approach to improve the overall driving reliability. The improved driving reliability of the proposed autonomous driving framework is shown through experimentation on a testbed system built on this work.