Development and Implementation of Safe, Human-like Autonomous Vehicle Control on Horizontal Curves

Recent advancements in deep learning and hard-ware technologies have enabled the application of novel control techniques for autonomous vehicles to handle challenging road conditions such as horizontal curves. This study presents a novel approach that blends human driver data with the Pure Pursuit algorithm, utilizing the law of propagation of uncertainty. Trained on the combined data, two deep learning architectures - Long Short-Term Memory (LSTM) networks and Fully Connected Deep Neural Networks (FCDNN) - are adopted: one predicts future speed profiles for calculating braking deceleration. In contrast, the other predicts steering rates for tracking a reference path, considering lateral acceleration and jerk constraints to ensure safety and comfort following the American Association of State Highway and Transportation Officials (AASHTO) and International Organization for Standardization (ISO) 15622 standards while emulating an expert human driving behavior. The proposed work has been implemented and tested using a Chevy Bolt Electric Utility Vehicle (EUV) during a U-turn, with test results demonstrating successful constraint satisfaction and acceptable cross-track error in reference path tracking.