Safe Reinforcement Learning-Based Eco-Driving Control of Autonomous HEV in Off-Road Scenarios

With the advancement of intelligent vehicle technologies, eco-driving control integrating velocity planning and energy management offers a solution to address charge and range anxiety in autonomous HEVs. However, due to complex road conditions, off-road eco-driving control has received limited scholarly attention. This paper proposes a novel integrated control system using a safe reinforcement learning (SRL) framework for eco-driving in off-road scenarios. First, we develop a method to identify safe velocity boundaries by quantifying constraints from complex road conditions based on vehicle dynamics stability criteria, deriving an analytical representation of the dynamic safe velocity boundary. Additionally, an SRL algorithm based on Lagrangian-SAC and an action-shielding mechanism is proposed, leveraging the advantage of model predictive control (MPC) in handling explicit constraints to guide safe actions while filtering unsafe ones during RL agent exploration. Finally, simulation and real-vehicle experiment results show that the proposed method effectively balances fuel economy and travel time while ensuring vehicle driving safety by strict safe constraints, achieving zero constraint violations in challenging off-road scenarios. Compared with the benchmark method, the fuel economy of the proposed method is improved by up to 6.12%.