Bridging Regimes: A State-Dependent Blending Methodology for Parsimonious and Robust Heavy Vehicle Dynamics Modeling

Data-driven gray-box models for vehicle control often fail to generalize across distinct physical regimes. This study tackles the critical, yet often-overlooked, challenge of robustly blending model parameters between these regimes. The vehicle’s “expert poles” are defined using physically distinct maneuvers (steady state vs. transient). A three-way benchmark is used to prove that the blending method is more critical than the concept itself. Three architectures are compared: (1) a baseline single-parameter “Static Model”, (2) a common literature “Heuristic Model” that blends using lateral acceleration (ay), and (3) the proposed “Dynamic Model” using a systematically optimized “Angle-Only” architecture. The findings demonstrate significant differences: The common-sense “Heuristic Model” exhibits severe degradation in stability metrics, lowering overall ay accuracy by 23.7% and r (yaw rate) accuracy by 110.2% compared to the baseline. In contrast, the “Angle-Only” model is the only architecture that successfully improves the primary ay objective (by 17.7%). The Dynamic Model’s 35.0% r-metric degradation is demonstrated to be the minimal, quantified engineering trade-off required for achieving robust adaptation—a task completely failed by the Heuristic Model. This study provides a validated, data-driven path for developing control-oriented models, proving that a simple, systematically optimized input is methodologically superior to the state-dependent heuristic.