In kidney transplantation, dynamic predictions of graft survival may be obtained from joint modelling of longitudinal and survival data for which a common assumption is that random-effects and error terms in the longitudinal sub-model are Gaussian. However, this assumption may be too restrictive, e.g. in the presence of outliers, and more flexible distributions would be required. In this study, we relax the Gaussian assumption by defining a robust joint modelling framework witht-distributed random-effects and error terms to obtain dynamic predictions of graft survival for kidney transplant patients. We take a Bayesian paradigm for inference and dynamic predictions and sample from the joint posterior densities. While previous research reported improved performances of robust joint models compared to the Gaussian version in terms of parameter estimation, dynamic prediction accuracy obtained from such approach has not been yet evaluated. Our results based on a training sample from the French DIVAT kidney transplantation cohort illustrate that estimates for the slope parameters in the longitudinal and survival sub-models are sensitive to the distributional assumptions. From both an internal validation sample from the DIVAT cohort and an external validation sample from the Lille (France) and Leuven (Belgium) transplantation centers, calibration and discrimination performances appeared to be better under the robust joint models compared to the Gaussian version, illustrating the need to accommodate outliers in the dynamic prediction context. Simulation results support the findings of the validation studies.