물리 정보 신경망을 통한 차량 동역학 모델 파라미터 추정

초록

As autonomous driving technology advances, validation in virtual environments is becoming increasingly crucial, leading to ongoing efforts to reduce discrepancies between simulated and real-world vehicle dynamic behavior. Dynamic consistency is essential for evaluating the reliability and safety of autonomous algorithms. Although data-driven learning methods have been actively explored, simulation models cannot be fully replaced by learning approaches alone due to the importance of interaction with ground surfaces. Therefore, this study proposes a physics-based parameter estimation method for improving vehicle dynamics accuracy in virtual simulations for autonomous vehicles. The proposed methodology utilizes Physics-Informed Neural Networks (PINN) to effectively estimate parameters of both longitudinal and lateral vehicle dynamics models. This approach adheres to the physical laws governing vehicle dynamics and ensures robust interactions with the physics engines used in virtual environments, leveraging data-driven learning to enhance dynamic consistency. The proposed method is expected to serve as a core technology for verifying autonomous vehicle systems in virtual environments, contributing significantly to improving the safety and reliability of autonomous driving systems.

키워드