Optimal vehicle dynamics and powertrain control of carbon-free autonomous vehicles: Large language model assisted heterogeneous-agent learning

Abstract

With the growing trend toward automation and decarbonization in heavy-duty transportation, ammonia–hydrogen hybrid electric vehicles (AHHEVs) equipped with autonomous driving capabilities are expected to play a significant role in long-haul freight applications. Although substantial progress has been made in the development of hardware propulsion systems, current AHHEVs generally lack advanced integrated energy-saving systems capable of coupled vehicle dynamics and powertrain control. To address this gap, this paper proposes a heterogeneous-agent reinforcement learning (HARL) framework for coupled optimization of velocity and energy management, where the large language models (LLM) serve as a high-level reasoning and coordination prior. Specifically, LLM-generated expert knowledge is leveraged to guide heterogeneous-agent policy initialization, constrain exploration within physically consistent and energy-efficient regions. Comprehensive tests demonstrate that this LLM-enhanced framework effectively reduces the sample complexity of reinforcement learning and accelerates convergence in dynamic driving scenarios, and the proposed method achieves more stable performance in terms of control accuracy, real-time response, and up to 2.6 % energy efficiency improvements.