Generative AI-based Adaptive Cruise Control Performance Evaluation in Car-Following Scenarios

Abstract

This study introduces a novel generative AI-based adaptive cruise control (GenAI ACC) model for autonomous vehicles, utilizing a time generative adversarial networks (TimeGAN) framework to simulate car-following behavior. The model was developed and evaluated using OpenACC and Vanderbilt ACC datasets, comparing its performance against traditional PATH ACC and cooperative adaptive cruise control (CACC) systems across safety, flow stability, comfort, and emissions metrics. The GenAI ACC demonstrated significant improvements in several key areas, including safety, flow stability, comfort, and emissions. With regard to safety, it achieved reduced conflict durations and a 40% higher minimum time to collision in high oscillation scenarios compared with PATH ACC. The model maintained neutral flow stability (relative mean absolute deviation ≈ 1) across all scenarios, effectively minimizing speed disturbances. Concerning comfort, GenAI ACC limited jerk variations to ±2 m/s³, substantially outperforming PATH ACC’s 12 m/s³ spikes in high oscillation scenarios. Most notably, it achieved 5–31% lower cumulative NO_x emissions across all scenarios compared with both PATH ACC and PATH CACC. Although GenAI ACC has not yet matched the stability and comfort levels of PATH CACC, it offers a promising solution for autonomous vehicles by relying solely on onboard sensors, reducing vulnerability to cyberattacks and connectivity issues. This study highlights the potential of generative AI in bridging the gap between automated and connected automated systems, paving the way for safer, more efficient, and environmentally sustainable transportation.