Imitation Learning Decision with Driving Style Tuning for Personalized Autonomous Driving

Existing autonomous driving systems mainly focus on the safety, comfort, and effectiveness of driving, but might ignore the personalized driving style of the vehicle. One model can hardly perform well in different driving contexts while training several models with different driving styles from scratch is not cost-effective. To address this problem, propose a personaliZed dEcision framework based on imitAtion Learning, named ZEAL, to enable autonomous vehicle driving in distinct driving styles. First, ZEAL uses a Base model trained with expert data from a reinforcement learning (RL) trainer, which is then fine-tuned to create models emphasizing either efficient or cautious driving style. We then leverage the Transformer blocks to weigh environmental inputs, making the autonomous vehicle enhance the attention to information that matters in a certain driving style. To rapidly adapt to various driving styles, we use small amounts of data to tune models based on a Base model while remaining safe, comfortable, and efficient. Extensive experiments conducted on a widely-used simulator demonstrate the effectiveness of ZEAL.