A detailed knowledge of a vehicle's characteristics makes it possible to monitor its dynamic behavior, energy consumption, and wear. In this preliminary study, we address the problem of learning a robust and consistent mass estimator from data provided by embedded sensors, which are subject to uncertainties. Consistency refers to the ability to comply with physical laws-Newton's second law in the case of mass estimation, robustness to the capacity to infer from uncertain or scarce data. This preliminary work aims at defining the problem and providing some guidelines with respect to constructing a robust and consistent mass estimator from uncertain data. Simple experiments on a Renault vehicle confirm the feasibility and the interest of learning a consistent vehicle model so as to increase the estimation accuracy of vehicle consumption.