In cooperative intelligent transportation systems (C-ITSs), vehicular positioning performance can be augmented by network-shared data and cooperative positioning (CP) algorithms. This article proposes an adaptive Gaussian mixed model (AGMM) based message passing (MP) method for robust vehicular CP, named AGMM Gaussian belief propagation (AGMM-GBP). In realistic engineering applications, positioning sensors are often corrupted by non-Gaussian noises with uncertain features, which brings challenges to the many existing CP methods. To address this problem, we model the observation noises via a set of Gaussian mixtures and apply it to the MP framework, which can fully approximate the non-Gaussian distributed noises in the MP. Moreover, to adaptively estimate the GMM parameters, a variational inference method is adopted for online learning of the noise distributions. In the experimental dataset based on the wheel odometer and monocular camera, AGMM-GBP outperforms many other existing CP methods at all levels, such as Gaussian belief propagation (GBP) and factor graph (FG) optimization. The overall framework of AGMM-GBP can be extended to other MP-based estimators, and it has potential applications in other scenarios with unknown non-Gaussian noises.