Automotive industry increases the usage of control units to improve the vehicle’s overall dynamic behaviour and tries to reduce the number and the cost of tests needed during the design phase, as the number of vehicle variants is increasing. Therefore, the development of virtual channels is proposed, aiming at replacing expensive and cumbersome physical sensors. Machine learning (ML) algorithms can be used to create these virtual channels due to their ability to capture complex dynamics quickly. On the other hand, a significant open challenge is to ensure the robustness of ML models against changes in the system under test, as training data (Source domain) and test data (Target domain) often differ in feature space distribution. This issue arises when models trained on one vehicle are deployed on another with different dynamic characteristics. The problem is exacerbated with unlabeled Target domains, as the Unsupervised Domain Adaptation (UDA) methods, belonging to Transfer Learning (TL), can only address marginal distribution shift, while there is also conditional distribution shift due to the different dynamic characteristics of vehicles. This paper introduces a robust ML-based virtual channel for estimating vehicle wheel center loads (WCLs). This study consists of selecting informative input signals, introducing UDA incapabilities due to conditional distribution shifts, and proposing a Domain-Aware Generalization (DAG) model to enhance robustness by extracting domain-specific and domain-invariant features. The methodology is validated with simulated and experimental datasets from different car designs, demonstrating its effectiveness.