Deep Learning-Based Hybrid Dynamic Modeling and Improved Handling Stability Assessment for Autonomous Vehicles at Driving Limits

This paper addresses the critical challenges of dynamic modeling and quantitative assessment of handling stability for autonomous vehicles (AVs). To meet the high safety standards in AVs, especially under conditions at driving limits, it is essential to employ precise dynamic models that match real-world system responses as closely as possible and provide reliable, quantitative evaluations of handling safety levels. We propose a deep learning-based hybrid dynamic modeling approach that integrates a designed recurrent neural network with vehicle physical models to enhance the traceability and interpretability of prevailing data-driven models. Utilizing recorded vehicle data, we first train the designed neural network to learn vehicle dynamic responses, thereby enhancing modeling accuracy. Then, the knowledge-based physical model and trained network model interact with each other and are mixed by exploiting the unscented Kalman filter. Furthermore, a comprehensive metric is developed to quantify the level of vehicle handling stability. We propose a multivariate dynamic representation to facilitate the analysis of complex vehicle dynamic behaviors. The mathematical definition of the handling safety metric is provided. Finally, hardware-in-the-loop (HIL) experiments are carried out to validate the proposed methods. Experimental results verify the effectiveness and better performance of the proposed hybrid dynamic model and handling safety metric. This work holds promising implications for accelerating the release and deployment of AVs.