Abstract
Uncertainty in vehicle parameters poses a persistent challenge for longitudinal vehicle-following control and limits the reliability of model-based controllers. This paper proposes a learning-based longitudinal control framework that predicts throttle commands in an end-to-end manner using a composite 1D CNN–LSTM–DNN network. The 1D CNN captures local temporal patterns from multi-signal car-following inputs, the LSTM models longer-term dependencies, and the DNN maps the learned representations to throttle commands, reducing reliance on explicit vehicle dynamics modeling. Experiments on real-world vehicle data collected from on-road driving demonstrate robust performance under parameter uncertainty, achieving an RMSE of 0.11 and an R² of 0.93.