Personalized longitudinal motion planning based on a combination of reinforcement learning and imitation learning

With advancements in autonomous driving technology, to minimize the decision-making disparities between human drivers and intelligent vehicles, the need for anthropomorphism and personalization in intelligent vehicles has become increasingly pressing. In planning longitudinal motion of intelligent vehicles, it is essential to consider multiple performance metrics as well as the driver's acceptance of the vehicle's driving style. This paper introduces a longitudinal motion planning policy that synergistically combines reinforcement learning with imitation learning. The primary framework is built on reinforcement learning, creating a foundational policy for longitudinal motion planning. Within this reinforcement learning context, this study incorporates a classic trajectory prediction method to construct an environment with prediction and deduction model (EPD). Generative Adversarial Imitation Learning (GAIL), a well-established imitation learning technique, is employed to assimilate human driver demonstration data into the reinforcement learning framework. The Deep Deterministic Policy Gradient (DDPG) algorithm, integrated with the EPD and GAIL models, is used to formulate a comprehensive personalized longitudinal motion planning policy. This policy is rigorously trained and tested on a natural driving dataset. The findings confirm that the proposed policy can adapt to the driving style of each target driver, achieving personalized driving while simultaneously meeting stringent performance indices in longitudinal motion planning compared to human drivers.