Large language models (LLMs) have shown remarkable capabilities for perceiving
driving environments and making interpretable, logical decisions for autonomous
driving. However, their potential for more comprehensive driving strategies,
especially concerning energy efficiency, remains underexplored. Most existing
studies primarily focus on driving safety, which may inadvertently increase
energy consumption. To address this issue, this study explores the use of LLMs
as high-level controllers to jointly optimize driving safety and energy
efficiency. A textual prompt is designed for the LLM, incorporating few-shot
examples that describe scenarios, states, and actions. The LLM processes the
scenario and state prompts describing the surrounding traffic environment. It
generates a high-level control signal, which is then translated into low-level
vehicle motion commands in a high-fidelity traffic simulator with realistic
physics, vehicle dynamics, road slopes, and network topology. Experiments in
campus-scale digital twin car-following scenarios demonstrate that the proposed
LLM-based framework achieves an average reduction of 4.16% in energy consumption
compared to the reinforcement learning paradigm, while maintaining driving
safety and providing interpretable high-level decision-making. This study
highlights the potential of LLMs for longitudinal eco-driving applications under
the evaluated simulation settings, extending previous LLM-based autonomous
driving research that primarily focused on safety to also consider energy
efficiency.
