Uncertainty-Aware Hybrid Estimation Architecture for Generalized Vehicle Lateral State Monitoring

Recent advances have leveraged a hybrid approach integrating data and physical laws for state estimation. However, the existing cascaded hybrid architectures, which connect physical and neural network components through inputs and outputs, fail to fully utilize the capabilities of data and physics, leading to limited accuracy and generalization performance. To address this problem, an uncertainty-aware hybrid estimation architecture consisting of a neural network, a physical lateral dynamics model, and a Monte Carlo sampling-based state estimator is proposed. Using this architecture, data and physical laws are integrated at the system modeling level, and the model uncertainty is dynamically quantified to fuse measurements for high-performance state estimation. First, a reconstructed hybrid lateral dynamics model is put forward, and the process noise is introduced for uncertainty description. Then, the aleatoric and epistemic uncertainty are characterized, in which the aleatoric uncertainty is learned via the neural network, while the epistemic uncertainty is quantified using the dropout Monte Carlo method. Finally, a Monte Carlo sampling-based state estimator is designed to use real-time measurements for state estimation considering different degrees of uncertainty. Simulation and experimental results showed that the proposed estimator can achieve high accuracy and strong generalization lateral state estimation, especially under uncovered working conditions.