Energy-Efficient Roundabout Crossing Strategy for Autonomous Electric Vehicles Based on Reinforcement Learning With Expert Demonstration

Learning-based methods, such as reinforcement learning (RL), have shown advantages over rule-based methods in energy-efficient strategies for autonomous electric vehicles (AEVs), showing the potential to achieve energy-efficient driving. However, existing energy-efficient strategies often optimize the energy allocation of the underlying controller of AEVs or adopt simple longitudinal speed planning at the overall driving behavior level to reduce energy demand, which is insufficient to achieve optimal energy efficiency. In view of this, this paper proposes an energy-efficient strategy that combines reinforcement learning and expert demonstration imitation learning, and adopts both lateral and longitudinal energy-efficient behaviors to maximize the energy efficiency of AEVs, and applies it to complex roundabout driving scenarios. The strategy incorporates adaptive imitation learning of expert demonstrations into RL training, so that AEVs have the energy-efficiency awareness of human experts to output energy-efficient driving actions. In addition, a rule-based safety-energy filter is designed downstream of the decision-making module to eliminate bad actions to improve learning efficiency and energy efficiency. Results show that compared with advanced baselines, the proposed strategy improves energy efficiency by an average of 11.51% while ensuring both driving safety and traffic efficiency.