Energy-Saving Full-Body Control of Hybrid Vehicles through Deep Reinforcement Learning

Deep reinforcement learning (DRL) algorithms have become an essential approach for enabling autonomous evolution in artificial intelligence (AI) models. As the representative form, autonomous vehicles equipped with new energy powertrains are expected to take into account the external dynamic environment, human-centered driving demand, and the internal power supply. This paper aims to integrate fundamental tasks of energy-saving and autonomous driving, realizing end-to-end self-learning-based improvements. Firstly, a Python-Simulink-Unreal Engine/Carla-based co-simulation framework is proposed after customizing 3D driving scenarios and enhancing vehicle modeling. Then, focusing on the vehicle-level embodied intelligence, a DRL-based energy-saving full-body control system is proposed, in which DRL agents are trained to collaboratively learn adaptive cruise control (ACC) and lane-keeping assist (LKA) at the vehicle driving layer, as well as energy management strategy (EMS) and transmission shift strategy (TSS) at the powertrain level. Moreover, to achieve the integration of energy-saving driving and end-to-end autonomous driving, Bird's Eye View (BEV) perception is incorporated into the state space. This not only constructs a dual-modal state space but also facilitates extensions to other end-to-end tasks. Finally, a hardware-in-the-loop (HIL) platform is established. The results show that the DRL-based energy-saving control system maintains a speed of 60 km/h while staying within the lane 3.82 m wide on a 7.041 km circular road. The complete drive lasts 557 seconds and achieves a fuel economy of 7.748 L/100 km.