Diffusion-driven hybrid unknown input observer for vehicle dynamics estimation

Abstract

Vehicle sideslip angle (or lateral speed) and steering angle are essential variables for autonomous vehicle control and active safety systems. Existing estimation methods often depend on deterministic models or neural networks for mapping, which restricts their ability to capture vehicle dynamics distributions in complex driving conditions. This study introduces a diffusion-driven hybrid estimation framework to achieve real-time estimation of lateral speed and steering angle. The nonlinear vehicle dynamics model is represented as a linear parameter-varying (LPV) model. A score-based diffusion model is embedded into an LPV unknown input observer (UIO) to capture the multimodal distribution of dynamics modeling uncertainty under different state-space model (SSM) parameters without retraining. The model uncertainty prediction sequences from the diffusion model are asynchronously generated to ensure real-time performance. Furthermore, we construct an end-to-end-based hybrid observer incorporating a fully connected neural network as an expert model to identify model uncertainty for diffusion model training. Based on Lyapunov stability theory, the ℓ∞ gain performance can be guaranteed to minimize the impacts of uncertainty approximation errors and improve the estimation quality. Experimental results obtained from a real-world test track demonstrate the consistent effectiveness of the proposed framework across various driving scenarios and different SSM parameters, especially in extreme driving scenarios outside the training data distribution.